Surface Rationality and Deep Mimicry: Regional Selection of Energy Priorities Under Smart Specialization 2021–2027

Abstract

Evidence-based prioritization is essential for effective specialization strategies (RIS3). However, there is a scarcity of evidence on whether regions are leveraging their own strengths or mimicking other policies. This study examines 236 EU regions, 178,314 publications, 116,336 projects and 470 RIS3 energy priorities (2021–2027) across 112 energy-related topics. We measure capability potential in two dimensions: proven areas of activity (inside strengths) and related areas with similar technologies (adjacent frontiers). Selective behavior is described using exploitation and exploration indicators, a stretch indicator and portfolio–priority adjustment indicators. Our findings reveal that surface rationality masks deep mimicry. Capabilities drive the direction of selection but not its scale. Regions select less than 10% of available strengths or adjacent areas. Instead, 40.3% of priorities are in ambitious areas, such as hydrogen and offshore wind energy, that exceed the potential opportunities. The portfolio–priority alignment is minimal at 0.10, and the wishful gaps are close to the maximum at 1.85 out of 2.0. Conversely, regions leave 93% of their potential untapped. RIS3 energy priorities indicate a greater desire to follow priorities than an ability to act. We suggest that policymakers should conduct capability audits to confirm absorptive capacity before setting priorities, establish benchmarks in strategy monitoring to measure the exploitation and exploration of regional assets and provide financial incentives that reward choices based on capabilities rather than historical alignment. Future research should examine whether capability-based priorities outperform choices that merely mimic others.

1. Introduction

The transition to environmentally friendly sustainable energy systems that ensure security presents significant obstacles to regional economic development [1,2]. The European Union must accelerate its efforts to decarbonize and modernize its energy infrastructure [3], which requires regional innovation systems to develop new opportunities and transform the current knowledge and technology framework [4,5]. The process of structural transformation, which Boschma [6] describes as ‘regional branching,’ requires economies to leverage their existing capabilities to develop new technologies that are closely aligned with their expertise. Public regional innovation funds must continue to support existing business strengths while building new industries and conducting innovative research and development, as these programs must deliver tangible results in their current state [7]. The current challenge is becoming more complex as different European regions demonstrate varying levels of innovation potential, institutional strength and industrial development patterns [8,9], which Rodríguez-Pose [10] describes as persistent geographical innovation gaps across Europe.

The EU Cohesion Policy currently uses research and innovation strategies for smart specialization (RIS3) as the main framework for helping regions discover their distinctive strengths that will drive sustainable economic development [11,12]. The entrepreneurial discovery process [13,14] forms the basis of RIS3, enabling regions to identify their strongest areas that can lead to global market leadership. The effectiveness of RIS3 implementation remains controversial despite its spread to all regions of Europe [11,12,15]. Researchers studying regional innovation policy must determine whether local strategies based on regional innovation policy can be considered a ‘smart’ approach that achieves success through real local opportunities and forward-looking methods, or whether they simply replicate existing systems through institutional mimicry [16,17]. Research by Di Cataldo et al. [18] shows that European RIS3s exhibit three main problems: multiple and unrelated objectives, an inability to adapt to regional opportunities and the practice of mimicking the investment choices of neighboring regions instead of identifying new business opportunities. This pattern is consistent with institutional theory on isomorphism [19], which shows that organizations implement identical practices because they face competitive forces, normative expectations and coercive requirements. Research by Pylak et al. [20] explains that the selection of RIS3 priority areas depends on mimicry behavior, as organizations choose to follow their peers and role models rather than using indicators that reflect their actual strengths.

The energy transition necessitates immediate solutions to address these specific questions [3,7,21]. The requirement to reduce carbon dioxide emissions has led to a socio-technical transformation, as described by Geels [22,23]. This involves a complete overhaul of the production and consumption system and the innovation framework, extending beyond standard technological advances. The RIS3 2021–2027 programming period requires EU regions to establish energy-related specializations, including renewable energy and hydrogen technologies, smart grid systems and energy storage solutions [21,24]. Regions face an uncertain future, as it is unclear whether these areas have operational potential or whether their stated goals are theoretical. Regions that set energy priorities must choose whether to remain in their current comfort zone or explore new adjacent opportunities. The relationship between business opportunities and selected priorities indicates whether entrepreneurs are using conventional methods to find new business opportunities or maintaining current operating procedures. What do the selection patterns reveal about the discrepancy between the normative vision of smart specialization and its practical implementation?

This study analyzes how EU regions use their RIS3 energy priorities through evidence-based development of their own capabilities or follow political trends without evidence. The study focuses on three main objectives: (1) identifying regional energy resources by assessing their core competencies and adjacent areas; (2) analyzing selection patterns through new assessment methods that measure exploitation and exploration as well as stretching activities; and (3) evaluating different theories explaining how organizations select their priorities through rational decision-making, exploratory approaches or mimetic behavior.

1.1. Conceptual Foundations: Portfolios, Capabilities and Energy Domains

Our research begins by defining the key concepts that form the basis of our analytical framework. The regional innovation portfolio comprises all research and innovation activities within a specific geographical area and is evidenced by scientific publications, patents and funded projects [6,25]. This portfolio represents the region’s entire knowledge base and technological capabilities for economic development. The portfolio concept employs evolutionary economic geography to study regional economies that develop over time, with their current capabilities determining future development paths [6,26].

The region’s demonstrated competencies in specific technological or thematic areas are revealed through its current activities rather than its declared goals [27,28]. In line with the literature on absorptive capacity [28], we distinguish between two types of competence resources. Inside strengths (IS) indicate areas where a region outperforms standard benchmarks, demonstrating its established expertise and capacity for knowledge development. A region that has not yet established its strengths can enter adjacent frontiers (AF) because it has capabilities that facilitate its entry into these areas. Boschma [6] uses the term ‘relatedness,’ and other researchers [25,26,29,30,31,32] measure this using indicators of technological proximity. Organizations need to understand their IS and AF because their current capabilities (IS) help them achieve maximum efficiency, but AF show them potential new business opportunities arising from related fields that they can understand.

The classification of energy-related innovation activities presents particular challenges in identifying these activities. Current taxonomies, such as NACE industry codes and IPC patent classes, as well as priority categories defined in RIS3 documentation, serve administrative and legal purposes but do not reflect the mental and technological aspects of innovative work [33,34]. Existing classification systems combine diverse activities into single categories, including photovoltaics wind energy biomass and geothermal energy, despite their differing knowledge requirements. The RIS3 energy priorities appearing in policy documents lack a structured framework, as different regions utilize varying terminology to describe their priorities at different levels of detail and with distinct thematic approaches. The analysis of selection behaviors necessitates a classification system that demonstrates the actual knowledge relationships between different energy fields as scientists, engineers and innovators utilize these links to connect their knowledge. The method chosen for semantic topic modeling (Section 2.1.2) enables the system to discover topic clusters through natural text analysis of publications and projects, thereby yielding knowledge structures derived from the data rather than utilizing predefined administrative categories.

1.2. The Two-Dimensional Challenge: Endowment and Selection

The literature on smart specialization focuses mainly on what regions choose as priorities, paying less attention to what regions could choose based on their existing capabilities [20,35]. Assessing strategic actions in different areas requires information about the options available to each region. Drawing on evolutionary economic geography [6,25,26], we argue that regional innovation policy operates in two basic dimensions. The potential dimension shows what a region actually has, through its IS that exceed EU benchmarks, and AF that are natural entry points due to their technological links. Second, the dimension of selection behavior shows what strategic choices a region makes, using its existing IS, checking AF or stretching for distant fields that are unrelated to its current capabilities.

The current organizational conflict between these elements aligns with March’s fundamental distinction [36], which distinguishes between the exploitation of existing organizational capabilities and the exploration for new innovative solutions. The implementation of regional policy introduces two important elements that influence exploratory activities through their effects on both geographical distribution and organizational constraints. The related literature on diversity [6,25,26] shows that organizations will expand their activities into technological AF because these fields enable better knowledge transfer and integration due to cognitive proximity. Organizations that conduct various related activities within their portfolios will achieve better results when entering new adjacent markets through unified operational management [25,26,30]. Organizations that enter new markets with characteristics different from their current activities will need to devote more time to learning, and their business success will become less likely [27,29,31,37].

The two-dimensional framework relates to current research discussions on the problem of ‘wishful thinking’ that occurs in RIS3 [18,20]. The selection of priorities by regions that do not correspond to their current capabilities [8,28] will lead to the inefficient use of policy resources and prevent the transformation of entrepreneurial discoveries into growth opportunities. Organizations that choose to focus on their current portfolio needs will become trapped in dependence on their existing development path, while losing important opportunities for organizational transformation [6,20,32]. An organization must find the right balance between leveraging its established IS and entering new AF that require the development of skills in existing competencies, but should not lead to the degradation of skills [18,26,32].

Research on the formulation of RIS3 has resulted in numerous studies [12,15,35], but researchers have not gathered sufficient empirical data to answer three fundamental questions. Currently, there is a lack of appropriate methods for identifying regional energy capabilities that separate a region’s IS from AF for future development and for tracking these capabilities over time. Current RIS3 assessment methods are based on general innovation indicators, which include patents and publications, as well as research and development intensity, but do not provide a detailed assessment of specific capabilities that would show whether regions are able to implement their chosen energy-related innovations [26,32]. A second important issue is that we do not understand how resources influence the way regions choose their priorities from among the available opportunities. Three possible regional behavior scenarios in terms of choice are the following: adapting opportunities, mimicking other regions and models, and setting unattainable goals based on wishful thinking [20]. Research on choice patterns has only addressed two aspects, namely thematic convergence and binary presence/absence of priorities, but has not measured the extent of mobilization of a region’s capacity base, the intensity of exploration and exploitation and the alignment of priorities with existing policy actions [18,32].

In this study, we address these knowledge gaps by examining the EU’s regional energy innovation capabilities and priority selection for 2014–2027. The study analyzes two questions, starting with descriptive landscape studies and then moving on to behavioral studies that test specific hypotheses.

1.3. Research Questions and Hypotheses

1.3.1. RQ1: What Energy Capability Endowments Do Regions Possess, and How Do These Evolve over Time?

Our first research question concerns the empirical picture of regional innovation capabilities in the field of energy across the EU. The research team analyzed 236 NUTS-2 regions in 112 energy-related areas to determine the distribution of their IS potential and the distribution of AF potential. The study identifies two types of regions based on their capability base: those that focus on their IS (high Portfolio Opportunity Index, POI) and those that focus on the untapped capabilities of their AF (low POI). The descriptive analysis serves three basic purposes, as it allows us to measure selection patterns (RQ2) while revealing which regions choose from available or limited skill sets. The study analyzes how different regions present their investment opportunities through their opportunity spaces, while also examining whether the EU energy innovation sector has a center-periphery structure [24] or exhibits different spatial distribution patterns. The study analyzes how portfolios naturally evolved over time, from 2014 to 2020, to create a baseline that can be used to assess future policy impacts.

Evolutionary economic geography identifies two opposing systems that work against each other. Path dependence combined with increasing returns causes regions to focus on their current IS, creating a self-reinforcing process that limits their overall scope of activity [26,32]. The logic of branching indicates that organizations with strong capability bases can use these platforms to enter new related markets, which would expand their business portfolio [6,27,31] in line with Jacobs’ externalities [38]. An analysis of trends prior to the introduction of coverage policy, which shows the percentage of potential energy fields (IS + AF) that regions actively pursue with above-median activity, helps us determine whether the capability landscape remains static or is naturally expanding, which affects how RIS3 interventions interact with regional systems.

1.3.2. RQ2: How Do Regions’ Capability Endowments Shape Their Priority Selection Behavior?

Research question RQ1 allowed us to create a map of possibilities, which we use to begin analyzing the logic behind the selection of strategic priorities for RIS3 2021–2027. The research team uses a number of complementary performance indicators to study selection patterns. The Exploitation rate (ER) shows what percentage of available IS receive priority selection, while the Exploration rate (ExR) indicates the percentage of AF areas that are being explored. The stretch ratio (SR) measures how often the selected priorities go beyond the boundaries of IS and AF, which is an ambitious goal or wishful thinking. The Selection Comfort-Zone Bias Index (SCZBI) shows the extent to which the selection focuses on IS rather than AF areas. The assessment of portfolio–priority alignment uses two methods to measure the similarity between pre-policy behavior patterns and selected priorities by calculating cosine similarity and L1 divergence.

The system for organizing regional strategic behaviors uses a 2 × 2 matrix that combines POI (capability endowment) with SCZBI (selective behaviors) to generate four strategic archetypes (Figure 1). Organizations belonging to the Strength Boosters archetype (high POI and high SCZBI) have many IS that they use to build their specialized capabilities. The Excelling Perfectionists (high POI, low SCZBI) archetype includes organizations that leverage their established strengths to build new capabilities because they have a solid foundation. The Explorers (low POI, low SCZBI) group operates with limited resources to enter untested markets through a risky business approach that requires external support. The Narrow Specialists (low POI, high SCZBI) group chooses to pursue unrelated or distant priorities because they lack core competencies, but they attempt to enter new markets thanks to their limited absorptive capacity to assimilate new information.

The research develops three opposing theories that explain different mechanisms through specific testable outcomes that help researchers distinguish between these theories.

H1 (Rational selection hypothesis):

An evidence-based entrepreneurship discovery system implemented by regions [13,14,35,39] will allow them to set priorities in line with their current capabilities. Research indicates that locations with many IS (high POI) need to increase their SCZBI and ER values above 15%, while locations with many AF (low POI index) need to decrease their SCZBI values and increase their ExR values above 15%. The research predicts that POI and SCZBI will show positive relationships in two distinct areas of the POI × SCZBI space, which correspond to comfort zone-oriented regions at the top and exploration-oriented regions at the bottom [20]. Empirically, we test whether POI predicts selection intensity (ER and ExR) after controlling for regional characteristics (GDP per capita, R&D capacity, legacy of energy priorities in 2014–2020).

H2 (explorative selection hypothesis):

Different regions decide to prioritize organizational learning because they want to build their organizational capabilities [19,36] through related diversity mechanisms [27,29,31]. Areas with a high POI that contain powerful bases will develop into new connections (Excelling Perfectionists), but areas with a low POI index will build their structure using numerous potential AF resources (Explorers). This hypothesis predicts high ExR (>15%) among low-SCZBI regions, regardless of POI, and a positive correlation between AF potential (number of adjacent frontier topics) and ExR. This analysis examines how organizations implement strategic diversification methods to maintain relatedness boundaries while stretching their business activities.

H3 (Mimicry and wishful thinking hypothesis):

This hypothesis shows that institutions follow institutional isomorphism [19] and policy mimicry [20,32] rather than using evidence when choosing options. The region selection process uses two methods, which involve following a reference group in terms of domain choices and mimicking successful innovators, such as the German hydrogen strategy and Danish offshore wind initiatives. This hypothesis predicts four main outcomes: (1) low-level portfolio–priority alignment (cosine similarity < 0.3), as pre-policy actions are independent of priority setting; (2) wishful gaps at high levels (L1 distance > 1.5, approaching the maximum of 2.0) due to different components; (3) high stretch ratios (SR > 30%), showing that areas beyond IS and AF are the subject of particular attention; and (4) substitution rather than addition patterns, as SR values above 30% will result in negative coefficients for ER and ExR in multivariate analyses (β < 0). The research results would show that regions use mimicry as a substitute for evidence-based selection methods rather than as an additional selection tool.

The three research hypotheses exist independently of each other, as ability determines which priority to choose between IS and AF, but mimicry determines which part of the ability base to activate. The evaluation of all three hypotheses shows which factor leads to a positive POI–SCZBI relationship through surface rationality or deep mimicry, which becomes evident through specific ER/ExR and SR values and substitution methods.

1.4. Integrated Theoretical Framework and Contributions

The research framework combines knowledge from evolutionary economic geography [6,22,25] with organizational learning theory [36] and related variety research [27,29,30], absorptive capacity [27,28], and institutional isomorphism [19] to examine regional innovation policy through its two operational elements, namely endowment and selective behavior. The tension between these two elements reveals whether smart specialization follows an evidence-based model of entrepreneurial discovery or evolves toward aspiration-based policy through imitative practices.

The research presents original findings that broaden our understanding of three independent academic disciplines. The research advances knowledge on smart specialization through the first large-scale study that analyzes how regions select their energy priorities, guided by indicators or by mimicking their peers and role models [20]. The research employs a two-dimensional framework that measures both regional capacity resources and patterns of their selection in terms of energy priorities. The developed indicators form a measurable system that enables a strategic assessment of behavior in regional innovation policy, while addressing existing problems related to mimicry approaches, unattainable goals, and unclear tasks [12,17,18,20,32,35]. The study documents three important elements: selected regions, the use of their potential through ER and ExR and the degree of alignment of the portfolio with selected themes and existing gaps.

Second, we contribute to evolutionary economic geography by empirically validating the concepts of internal assets and neighboring borders through detailed calculations of linkage density in 112 energy-related themes. Our finding that regions with high relatedness density (high AF potential) encounter fewer cognitive barriers to entry is consistent with branching theory [6,25,26,31,37], but our finding that these regions nevertheless exhibit low exploration rates (≈5%) reveals an important observation: adjacent benefits are necessary but not sufficient for diversification. The technological capabilities of regions enable them to enter new energy areas, but they consistently choose to focus on distant goals rather than related opportunities (high SR). The gap between what regions can achieve and what they actually choose to pursue indicates that policies aimed at stimulating new areas of activity through the development of related variety must establish performance-based systems that support regions in pursuing new opportunities rather than chasing unproven ones. The SR, which tracks priorities beyond IS and AF, creates a new way to measure “opportunity distance” that researchers can use to explore why distant diversification efforts fail compared to opportunity-based choices.

Third, our research supports the theory of learning and mimicry in policy, as mimicry extends beyond convergence on topic labels to encompass the selection process itself: systematically under-mobilizing capabilities (low ER/ExR) and over-allocating to wishes (high SR) with a substitution mechanism whereby stretch crowds out evidence-based choices. The RIS3 project necessitates a fundamental transformation, as it requires more than just diversification of strategies to prevent mimicry, which involves changing the way priorities are set by assessing absorption capacities [27,28]. The RIS3 development process will create a competitive environment between regions that will showcase contemporary sectors (Green Deal, hydrogen economy and offshore wind energy) without assessing their operational value, as there are no stabilizing mechanisms, including ER/ExR indicator controls and capacity justification requirements, and performance-based reward systems. Pylak et al. [20] examined in their research how users choose between following their peers and following systemic indicators.

The research presents three important contributions: (1) the introduction of the first complete assessment of EU NUTS-2 regional energy resources through a detailed analysis of relatedness density; (2) the creation of new performance indicators that help researchers evaluate the selection of RIS3 priorities through the analysis of ER, ExR, SR, alignment index and wishful gap analysis; and (3) showing that regions employ strategic stretching and policy mimicry tactics because they do not effectively exploit available opportunities, devoting excessive resources to achieving aspirational goals. The framework we have analyzed for energy priorities can be applied to other RIS3 areas, including health, digital transformation and circular economy.

The research design utilizes a standardized database containing 178,314 OpenAlex publications, 116,336 CORDIS and Kohesio innovation projects and 1159 RIS3 energy priorities from 2021 to 2027 across 238 NUTS-2 regions and 112 energy topics. These were extracted using BERTopic (version 0.17.4) with LLM-labeled semantic topic modeling. The existing data infrastructure enables three fundamental operations: (1) the generation of capability metrics from historical data (2018–2020) as reference points; (2) the construction of a sparse similarity matrix from soft topic assignments at the document level to represent the modular structure of energy topics; and (3) the determination of adjacent elements for each region–topic combination through relatedness density measurement; and (4) the assessment of research hypotheses via correlation analysis, multivariate regressions and subsample comparisons, which also reveal regional prioritization.

The remainder of the paper is organized as follows. Section 2 describes the data sources, semantic topic modeling approach, construction of the region–year–topic panel, operationalization of capability endowment and selection behavior metrics and analytical strategy for hypothesis testing. Section 3 presents information in a sequential order. Section 3.1 demonstrates how capability endowments develop (RQ1), and Section 3.2 describes selection patterns and the H1–H3 tests and identifies strategic archetypes (RQ2). Section 4 evaluates theoretical and policy implications through smart specialization and evolutionary economic geography frameworks to establish research connections while presenting the study’s limitations and potential directions for future research.

2. Materials and Methods

2.1. Data and Topic Space Construction

The following section explains how our analytical dataset was constructed through three distinct processes. Section 2.1.1 presents data sources: OpenAlex for scientific publications, CORDIS and Kohesio for innovation projects and Eye@RIS3 for regional priorities. We perform spatial harmonization of all recorded data to NUTS-2 2021 boundaries. The process of extracting 112 energy topics from BERTopic semantic modeling follows in Section 2.1.2, which uses LLM-based classification to match these topics with RIS3 priority descriptions. The regional panel construction in Section 2.1.3 unites all data sources into a single analytical framework for regions by year and topic.

2.1.1. Data Sources and Spatial Harmonization

RIS3 aims to align regional innovation activities with IS and AF, but assessing this alignment requires a harmonized empirical basis. To address RQs 1–2, we construct a multi-source regional panel of research outputs, innovation projects and RIS3 energy policy priorities covering the EU-27 between 2014 and 2025. The research method follows established principles from regional studies and innovation system research, which emphasize micro-level activity-based indicators to understand capacity development, diversification patterns and policy reactions [6,25,26,33,34].

Our choice of publications and innovation projects as the empirical basis for measuring capability endowments—rather than regional statistics or energy strategy documents—reflects three methodological considerations. The first issue with conventional regional statistics is that they measure total economic performance through energy production and consumption and sectoral employment data, but they do not track knowledge development activities, and they cannot identify specific energy domain capabilities like hydrogen storage, offshore wind and smart grids. Second, we use actual research and innovation activities from publications and projects to show revealed capabilities, while regional energy strategies only show what regions claim to do without proving their actual ability to absorb new knowledge. The research design of our study uses the comparison between actual capabilities shown in publications and projects and the declared priorities in RIS3 documents as its main analytical benefit. Third, the textual information found in publications and project descriptions enables semantic analysis, which reveals new knowledge patterns so we can detect specific thematic categories that standard administrative categories fail to identify.

The established data architecture serves as the base that supports our methods for semantic topic modeling and capability mapping. The empirical dataset draws on publicly accessible repositories: OpenAlex (an open-access bibliographic catalog of scholarly works, authors and institutions; https://openalex.org, accessed on 4 September 2025 [40]), CORDIS (the European Commission’s catalogue of EU Framework Programme research and innovation projects; https://cordis.europa.eu/, accessed on 4 September 2025), Kohesio (the Commission’s repository of Cohesion Policy projects and beneficiaries; https://kohesio.ec.europa.eu/, accessed on 4 September 2025), Eye@RIS3 (RIS3 priorities for 2014–2020; https://place-based-innovation.ec.europa.eu/projects-0/eyeris3-innovation-priorities-europe_en, accessed on 4 September 2025 [41]) and the S3 Community of Practice Observatory (RIS3 priorities for 2021–2027; https://ec.europa.eu/regional_policy/assets/s3-observatory/index_en.html, accessed on 4 September 2025). For thematic scope, we retain OpenAlex publications classified in energy with publication dates 2014–2025; CORDIS projects tagged Energy under Horizon 2020 (2014–2020) and Horizon Europe (2021–2027); Kohesio projects (2014–2020; 2021–2027) whose Thematic Objective Label corresponds to low-carbon economy, climate change adaptation and risk prevention, network infrastructure in transport and energy or environment protection and resource efficiency; and RIS3 priorities (2014–2020; 2021–2027) classified under the energy scientific category. All records are limited to EU Member States (EU-27) and harmonized to NUTS-2 (2021) as follows: publications by author affiliations; CORDIS/Kohesio by project beneficiaries; and RIS3 priorities by normalizing reported NUTS codes and cross-walking historical codes to the 2021 standard, including expansion of NUTS-0/1 entries to their constituent NUTS-2 units. We applied an area-weighted crosswalk method to transform historical NUTS code records without coordinates into NUTS-2 (2021) through polygon intersection analysis between legacy units and the 2021 layer. The analytical dataset excludes all records which cannot be assigned to NUTS-2 (2021) because their geodata is missing or their codes are invalid or non-existent. The data sources together with their respective coverage information appear in

Table 1. Data sources and coverage.

Data Source	Type	Content Description	Coverage Period	Records (N)	Spatial Unit	Key Variables
OpenAlex	Publications	Scholarly works with institutional affiliations	2014–2024	178,314	Institution → NUTS-2	Title, abstract, authors, affiliations
CORDIS	Projects	EU Framework Programme (H2020, Horizon Europe)	2014–2025	3817	Participant → NUTS-2	Title, objective, participants, funding
Kohesio	Projects	EU Cohesion Policy projects	2014–2025	112,519	Beneficiary → NUTS-2	Title, summary, beneficiaries, funding
Eye@RIS3	Priorities	RIS3 priority descriptions	2014–2020/2021–2027	470 (284 + 186)	NUTS-2	Priority text, thematic codes, region

Table notes: All records filtered to energy-related content using keyword and semantic classification. Spatial units harmonized to NUTS-2 2021 boundaries. Final analytical panel covers 236 EU regions across 112 energy topics.

.

2.1.2. Extracting Topics and Aligning Them to Regional Priorities

To translate heterogeneous publications and projects into a comparable analytical space, we extract high-resolution semantic topics using BERTopic. This step is essential because traditional classifications (e.g., Web of Science categories, Horizon thematic tags) are too coarse to capture the fine-grained technological pathways along which regions diversify [25,26]. Topic extraction provides a data-driven representation of knowledge domains, while the subsequent LLM-based labeling and RIS3 matching enable a unified mapping between real-world energy activity and regionally declared policy priorities.

To build our aggregated corpus of publicly funded research and project results, we use titles from OpenAlex publications and CORDIS/Kohesio projects, supplemented with abstracts and keywords where available. We implement BERTopic because it is tailored to short texts and yields coherent topic groupings with sparse, interpretable descriptors. Concretely, documents are embedded with the Sentence-Transformer para-phrase-multilingual-MiniLM-L12-v2 to place all sources in a single semantic space, reduced via UMAP to preserve local neighborhoods while denoising, clustered with HDBSCAN to accommodate heterogeneous cluster shapes and explicitly handle noise and described with class-TF-IDF (c-TF-IDF) to produce compact, human-readable topic terms. A document’s topic is defined as the index of the global cluster to which HDBSCAN assigns it; documents not assigned by HDBSCAN are marked as outliers (no topic). Outliers are optionally reassigned to the nearest topic centroid only when the cosine similarity is greater than or equal to 0.20; otherwise, they remain without a topic label. In addition to the hard topic index, we retain diagnostic soft evidence: for each document, we compute cosine similarities to all global cluster centroids and store the top ten raw (unnormalized) values.

To extract topics closely related to energy and label them, we use a local LLM (llama4) model for the most frequently occurring terms in each cluster. The LLM model flags energy-related clusters and assigns them concise titles that are understandable to policymakers based on the most frequently occurring terms.

To place RIS3 2021–2027 priorities in the same space, we embed each priority topic with the same Sentence-Transformer used for publications/projects and compare it only to the previously selected energy topics. Using FAISS nearest-neighbor search (cosine on L2-normalized vectors), a priority topic is linked to its single closest energy topic only if the similarity exceeds a conservative threshold of 0.70, chosen through stratified hand audit to prioritize precision over recall and ensure that matched priorities genuinely correspond to the same semantic domain.

2.1.3. Regional Panel Construction

Having placed all documents and policy priorities into the same semantic topic space, we next construct a region–year–topic panel. This panel structure aligns with evolutionary economic geography frameworks in which regional capabilities emerge cumulatively, path-dependently, and often unevenly [6,22,25]. The weighting approach ensures that documents contribute proportionally to multiple affiliated regions and that soft semantic evidence is preserved—critical features for studying subtle shifts in capability formation.

We build a comparable panel of topic intensities for each region and year to (i) quantify how research and policy-funded activity is distributed across topics within regions, (ii) track shifts over time and (iii) use harmonized, additive measures in downstream analyses. We retain only documents whose topic is flagged as energy-relevant.

Let $\epsilon$ be the set of all energy-relevant topics. From the stored diagnostics, for every document we have the top ten unnormalized cosine similarities to topics (i.e., global cluster centroids). For document $d$ observed in year $t$, let $k^{*}$ denote its topic (assigned global cluster index) and let ${cos}_{d,k,t}$ denote the cosine similarity between $d$ and topic $k$. Define the top ten topics set for $\left(d,t\right)$ as $N_{d,t}$. We keep only energy topics that appear in the document’s top ten and lie sufficiently close to its topic $k^{*}$, using a data-driven threshold ${\theta }_{d,t}=max\left\{\right(1-\tau )\cdot {cos}_{d,k^{*},t},{cos}_{d,k^{*},t} - \mathsf{\delta } , 0\}$ with fixed $\tau \in \left[\mathrm{0,1}\right)$ and $\delta \ge 0$ set ex-ante. The retained set is

$$K_{d,t}=\{kϵ\epsilon \cap N_{d,t}:{cos}_{d,k,t}\ge {\theta }_{d,t}\}$$

We form raw weights ${\tilde{p}}_{d,k, t}={cos}_{d,k,t}$ for ${kϵK}_{d,t}$ and renormalise within $(d,t)$:

$$p_{d,k,t}={\displaystyle \frac{{\tilde{p}}_{d,k,t}}{\sum _{j\in K_{d,t}}{\tilde{p}}_{d,j,t}}}.$$

Hence, $\sum _k{p}_{d,k,t}=1.$

If $d$ lists multiple NUTS-2 affiliations in year $t$, we allocate the document’s unit mass across regions $r$ proportionally to their frequency in the affiliation list and normalize to one. Let $R_{d,t}$ be the set of NUTS-2 regions listed in $d$’s affiliations in year $t$. Let $c_{d,r,t}$ be the affiliation count of region $r$ in that list. We allocate the unit mass of $d$ across its affiliated regions proportionally to these counts:

$$w_{d,r,t}\hspace{0.25em} =\hspace{0.25em} {\displaystyle \frac{c_{d,r,t}}{\sum _{r\mathrm{’}\in R_{d,t}}{c}_{d,r\mathrm{’},t}}}\hspace{0.17em}.$$

Hence, $\sum _{r\in R_{d,t}}{w}_{d,r,t}=1.$

For document $d$ let $p_{d,k,t}=0$ for non-related topics $k$ and $w_{d,r,t}=0$ for $r\notin R_{d,t}$ in year $t$. Topic $k$ mass in region $r$ and year $t^{\mathrm{*}}$ is

$${mass}_{r,t^{\mathrm{*}},k} =\hspace{0.25em} \sum _{d:t=t^{\mathrm{*}}}{w}_{d,r,t}·p_{d,k,t}.$$

If in region $r$ in year $t^{*}$ there are no documents with topic $k$, we assume that ${mass}_{r,t^{*},k} =0.\hspace{0.25em}$

The region–year total mass is ${mass}_{r,t^{*}}^{total}={\sum }_k{mass}_{r,t^{*}, k}$. The topic share within a region–year is

$${share}_{r,t^{\mathrm{*}},k}\hspace{0.25em} =\hspace{0.25em} \left\{\begin{array}{c}{\displaystyle \frac{{mass}_{r,t^{\mathrm{*}},k}}{{mass}_{r,t^{\mathrm{*}}}^{total}}}\\ 0\end{array}\right.\begin{array}{cc},& {mass}_{r,t^{\mathrm{*}}}^{total}>0,\\ ,& \mathrm{otherwise}.\end{array}$$

2.2. Mapping Regional Capability Endowments and Evolution (RQ1)

This section puts capability endowments into practice to answer RQ1: what energy capability endowments do regions possess and how do these evolve over time? We develop six measures. Section 2.2.1 defines the pre-policy topic shares that set the baseline portfolio for each region. Section 2.2.2 builds a topic similarity matrix that shows how technology topics relate across energy domains. Section 2.2.3 computes relatedness density to assess how close each region is to topics it has not yet explored. Section 2.2.4 adds capability endowment tags that separate IS from AF. Section 2.2.5 shows the POI that summarizes the region’s strategic position. Section 2.2.6 examines coverage dynamics, i.e., whether regions expand into the capability domains that are available during the pre-policy period (2014–2020).

2.2.1. Pre-Policy Topic Shares

For each region $r$ and topic $k$ we define the ex-ante baseline share $v_{r,k}^{\mathrm{base}}$ as the average over only those baseline years with positive portfolio mass in that region:

$$v_{r,k}^{\mathrm{base}}={\displaystyle \frac{1}{\left|T_{0,r}\right|}}\sum _{t\in T_{0,r}}{share}_{r,t,k},$$

where $T_{0,r}=\{t\in \{2018,2019,2020\}:{mass}_{r,t}^{total}>0\}$ and $\left|T_{0,r}\right|$ denotes the number of years in set $T_{0,r}$.

This baseline anchors each region’s portfolio composition before the 2021–2027 policy window. By averaging only over years with observed activity, we avoid mechanically pushing shares toward zero in regions/years where the corpus simply contains no documents. Using a multi-year mean (2018–2020) dampens idiosyncratic noise and business-cycle fluctuations, yielding a stable pre-policy reference.

2.2.2. Topic Similarity Matrix

To capture technological proximity among energy topics, we construct a corpus-based similarity matrix using documents’ soft topic assignments. Let $\mathcal{D}=\{1,\dots ,D\}$ index all document–year pairs $\left(d,t\right)$ and let $\mathcal{K}=\{1,\dots ,K\}$ index topics. Define the document-cluster matrix $X\in {\mathbb{R}}^{D\times K}$ by $X_{ik}=x_{ik}:=p_{d,k,t}$ for $i\in \mathcal{D},k\in \mathcal{K}$, where $p_{d,k,t}$ are the within-document soft topic weights constructed in Section 2.1.3.

Let $C=X^{\mathrm{\top }}X\in {\mathbb{R}}^{K\times K}$, where

$$C_{kk\mathrm{’}}=\sum _{i=1}^D{x}_{ik}{x}_{ik\mathrm{’}}=\sum _{\left(d,t\right)\in \mathcal{D}}{p}_{d,k,t}{p}_{d,k\mathrm{’},t}$$

and $\parallel X_{\cdot k}\parallel =\sqrt{\sum _{i=1}^D{x}_{ik}^2}=\sqrt{\sum _{\left(d,t\right)\in \mathcal{D}}{p}_{d,k,t}^2}$, $k\in \mathcal{K}$.

We then define the cosine-similarity matrix $S\in {\mathbb{R}}^{K\times K}$ as

$$S_{kk\mathrm{’}}=\left\{\begin{array}{ll}{\displaystyle \frac{C_{kk\mathrm{’}}}{\parallel X_{\cdot k}\parallel \cdot \parallel X_{\cdot k\mathrm{’}}\parallel }},& \mathrm{if} k\ne k\mathrm{’} \mathrm{and} \parallel X_{\cdot k}\parallel \cdot \parallel X_{\cdot k\mathrm{’}}\parallel >0,\\ 0,& \mathrm{otherwise} (\mathrm{in} \mathrm{particular} \mathrm{if} k=k\mathrm{’}).\end{array}\right.$$

By construction, $S$ is symmetric, has a zero diagonal and its entries lie in $\left[0,1\right]$. This similarity matrix provides a corpus-driven topology of topics: two clusters are close when documents frequently assign them positive weights simultaneously. The use of cosine similarity makes the measure scale-free and resistant to uneven topic sizes [42,43]. To further reduce noise, we thin out $S$ by setting values below the 5th percentile to zero, eliminating micro-similarities from rare co-occurrences.

2.2.3. Relatedness Density

We next build the corpus-level topology to obtain a region-topic measure of AF potential. For region $r$ and topic $k$, relatedness density (RD) is the weighted average of similarities with other topics, where weights are that region’s ex-ante share in those topics [44,45]:

$$R{D}_{r,k}={\displaystyle \frac{\sum _{k\mathrm{’}\ne k}{S}_{kk\mathrm{’}}\cdot v_{r,k\mathrm{’}}^{\mathrm{base}}}{\sum _{k\mathrm{’}\ne k}{S}_{kk\mathrm{’}}}}.$$

In case the denominator is zero (i.e., no non-trivial neighbors under $S$), we set $R{D}_{r,k}=0$. Intuitively, $R{D}_{r,k}$ combines proximity $S$ with the region’s pre-policy capabilities $v^{\mathrm{base}}$ into a single ‘neighborhood readiness’ value. It quantifies the strength of a region’s portfolio in domains that are already close to $k$. A high $R{D}_{r,k}$ indicates concentration near $k$’s neighborhood and thus lower frictions for diversification toward $k$ [45].

2.2.4. Definitions of Capability Endowment Tags

We establish two fundamental capability endowment tags on the $\left(r,k\right)$ grid. These tags allow us to assess the regions’ capabilities prior to any policy intervention during 2021–2027.

A region $r$ possesses inside strength potential (IS_potential) in topic $k$ when its ex-ante baseline share $v_{r,k}^{\mathrm{base}}$ exceeds the EU-wide 60th percentile for that topic, i.e.,

$${\mathrm{IS}\_\mathrm{potential}}_{r,k}=\left\{\begin{array}{ll}1,& \mathrm{if} v_{r,k}^{\mathrm{base}}>Q_{0.60}\left(v_{•,k}^{\mathrm{base}}\right),\\ 0,& \mathrm{otherwise}.\end{array}\right.$$

This flags topics where the region already demonstrates a large presence compared to the EU average distribution, indicating existing capability and absorptive capacity [27,28].

A region possesses adjacent frontier potential (AF_potential) in topic $k$ when it lacks current IS but exhibits high relatedness to its existing portfolio, formally:

$${\mathrm{AF}\_\mathrm{potential}}_{r,k}=\left\{\begin{array}{ll}1,& \mathrm{if} R{D}_{r,k}>Q_{0.60}\left(R{D}_{\cdot ,k}\right) \mathrm{and} {\mathrm{IS}\_\mathrm{potential}}_{r,k}=0,\\ 0,& \mathrm{otherwise}.\end{array}\right.$$

Here, $Q_{0.60}\left(R{D}_{\cdot ,k}\right)$ is the EU-wide 60th percentile of $RD$ for topic $k$ (computed across regions). The condition ${\mathrm{IS}\_\mathrm{potential}}_{r,k}=0$ ensures that AF_potential targets nearby opportunities rather than current strengths. This captures the branching logic: regions can more easily diversify into domains that are cognitively proximate to their existing portfolio [6,26,45].

2.2.5. Portfolio Opportunity Index (POI)

To summarize each region’s capability and resource profile, we define the Portfolio Opportunity Index:

$$PO{I}_r={\displaystyle \frac{\sum _k{\mathrm{IS}\_\mathrm{potential}}_{r,k}}{\sum _k{\mathrm{IS}\_\mathrm{potential}}_{r,k}+\sum _k{\mathrm{AF}\_\mathrm{potential}}_{r,k}}}.$$

$PO{I}_r\in \left[0,1\right]$ measures whether a region’s capability base leans toward the existing strengths (high POI, many IS_potential domains) or toward adjacent opportunities (low POI, many AF_potential domains relative to IS_potential). Regions with high POI possess deep portfolios in proven domains but may face saturation; regions with low POI have shallower current strengths but broader AF. This index is not defined for a region with neither IS_potential nor AF_potential topics; in practice, we set $PO{I}_r=0.5$ (neutral) for such cases, which are few in our energy-related sample.

2.2.6. Coverage Dynamics: Temporal Evolution of Capability Portfolios

To assess whether regions naturally expand into their available capability domains before policy intervention (organic portfolio broadening vs. path-dependent concentration), we construct a coverage measure tracking the fraction of potential energy topics that regions actively pursue over time.

For each region $r$ and year $t\in \{\mathrm{2014,2015},\dots ,2020\}$, define:

$${\mathrm{Coverage}}_{r,t}={\displaystyle \frac{\#\{k:{mass}_{r,t,k}>0\}}{\#\{k:{\mathrm{IS}\_\mathrm{potential}}_{r,k}=1 \mathrm{or} {\mathrm{AF}\_\mathrm{potential}}_{r,k}=1\}}}.$$

This measures what fraction of the region’s capability space (IS_potential + AF_potential) exhibits positive observed activity in year $t$. An increasing trend in ${\mathrm{Coverage}}_{r,t}$ over 2014–2020 would indicate organic portfolio broadening; a flat or declining trend would suggest path-dependent concentration or lock-in.

We estimate a simple panel regression to test for systematic temporal trends:

$${\mathrm{Coverage}}_{r,t}=\alpha +\beta \cdot {\mathrm{Year}}_t+{\gamma }_r+{\epsilon }_{r,t},$$

where ${\gamma }_r$ are region fixed effects and ${\mathrm{Year}}_t$ is a linear time trend. A positive $\beta$ indicates regions are organically expanding coverage before policy intervention; a null $\beta$ suggests static portfolios.

2.3. Characterizing Priority Selection Behavior (RQ2)

This section characterizes priority selection behavior to address RQ2: how do regions’ capability endowments shape their priority selection under RIS3 2021–2027? Seven measures are created. Section 2.3.1 sets treatment assignment and priority flags that link RIS3 documents to the topic space. Section 2.3.2 adds priority positioning tags that group each priority selection as strength-based, adjacency-based or stretch. Section 2.3.3 demonstrates the Selection Comfort-Zone Bias Index (SCZBI) that reflects strategic orientation. Section 2.3.4 defines exploitation (ER), exploration (ExR) and stretch (SR) rates. Section 2.3.5 builds portfolio–priority match metrics including alignment and wishful gap. Section 2.3.6 adds the Opportunity Cost Index (OCI) to assess lost capability potential. Section 2.3.7 classifies regions into four archetypes using a 2-by-2 grid based on SCZBI and POI.

2.3.1. Treatment Assignment and Priority Flags

To integrate policy priorities into our analytical framework, we assign treatment status at the region-topic level. The RIS3 2021–2027 dataset does not report exact adoption dates for each priority. To avoid imputing uncertain timing and maintain comparability, we adopt a conservative rule: all priorities recorded in a region’s RIS3 2021–2027 documentation are treated as becoming active at the start of the programming period.

For each region-topic pair $\left(r,k\right)$, we define:

$${\mathrm{priority}}_{r,k}=\left\{\begin{array}{ll}1,& \mathrm{if} \mathrm{topic} k \mathrm{appears} \mathrm{in} \mathrm{region} r’\mathrm{s} \mathrm{RIS}3 2021–2027 \mathrm{priorities},\\ 0,& \mathrm{otherwise}.\end{array}\right.$$

Let $P_r=\{k:{\mathrm{priority}}_{r,k}=1\}$ denote the set of topics selected in region $r$’s priorities, with cardinality $\left|P_r\right|$. This indicator is the foundation for all subsequent selection behavior measures.

2.3.2. Priority Positioning Tags

We classify each selected priority according to its position in the region’s capability landscape:

• Priorities that fall within the region’s IS potential (Inside Strength Selected, IS_selected):

$${\mathrm{IS}\_\mathrm{selected}}_{r,k}={\mathrm{priority}}_{r,k}\times {\mathrm{IS}\_\mathrm{potential}}_{r,k}.$$

• Priorities that fall within the region’s AF potential (Adjacent Frontier Selected, AF_selected):

$${\mathrm{AF}\_\mathrm{selected}}_{r,k}={\mathrm{priority}}_{r,k}\times {\mathrm{AF}\_\mathrm{potential}}_{r,k}.$$

• Priorities that fall outside both IS_potential and AF_potential, i.e., wishful thinking or distant exploration (Stretch_selected):

$${\mathrm{Stretch}\_\mathrm{selected}}_{r,k}={\mathrm{priority}}_{r,k}\times \left(1-{\mathrm{IS}\_\mathrm{potential}}_{r,k}\right)\times \left(1-{\mathrm{AF}\_\mathrm{potential}}_{r,k}\right).$$

These tags split the priority set $P_r$ into three mutually exclusive categories based on pre-policy capability positioning.

2.3.3. Selection Comfort-Zone Bias Index (SCZBI)

To quantify whether regional strategies lean toward exploitation of existing IS or exploration of AF, we define the Selection Comfort-Zone Bias Index (SCZBI) as follows:

$${\mathrm{SCZBI}}_r={\displaystyle \frac{1}{\left|P_r\right|}}\sum _{k\in P_r}\mathbf{1}\{{\mathrm{IS}\_\mathrm{selected}}_{r,k}=1\}-{\displaystyle \frac{1}{\left|P_r\right|}}\sum _{k\in P_r}\mathbf{1}\{{\mathrm{AF}\_\mathrm{selected}}_{r,k}=1\},$$

where $\mathbf{1}\{\cdot \}$ is the indicator function (equals 1 if the condition holds, 0 otherwise).

${\mathrm{SCZBI}}_r\in \left[-1,1\right]$ measures the difference between the share of priorities that target IS and the share of priorities that target AF. Values near +1 indicate selecting proven strengths, while values near −1 indicate selecting adjacencies; values near 0 reflect a balanced mix. This operationalizes March’s [22] exploration–exploitation framework at the regional strategy level and is similar to revealed comparative advantage/specialization indexes [46].

2.3.4. Selection Rates: Exploitation, Exploration and Stretch

The assessment of selection efficiency, which indicates the number of available opportunities a region selects for its priorities, requires the measurement of three specific rates.

The fraction of available IS that are selected as priorities (Exploitation Rate, ER):

$$E{R}_r={\displaystyle \frac{\sum _k{\mathrm{IS}\_\mathrm{selected}}_{r,k}}{\sum _k{\mathrm{IS}\_\mathrm{potential}}_{r,k}}}.$$

A high ER rating indicates that the region operates at its peak capacity due to its current strengths. A low ER rating suggests that vital operational capabilities remain unutilized.

The fraction of available AF that are selected as priorities (Exploration Rate, ExR):

$$Ex{R}_r={\displaystyle \frac{\sum _k{\mathrm{AF}\_\mathrm{selected}}_{r,k}}{\sum _k{\mathrm{AF}\_\mathrm{potential}}_{r,k}}}.$$

High ExR indicates expansion activities, while low ExR suggests untapped diversification opportunities.

The fraction of priorities that fall outside both IS and AF potential (Stretch Rate, SR):

$$S{R}_r={\displaystyle \frac{\sum _k{\mathrm{Stretch}\_\mathrm{selected}}_{r,k}}{\left|P_r\right|}}.$$

A high SR indicates wishful thinking or high-risk distant exploration, while a low SR reflects priorities based on regions’ actual capabilities.

2.3.5. Portfolio–Priority Concordance: Alignment and Wishful Gap

To assess how closely regional priority vectors match their pre-policy activity portfolios, we compute two complementary measures as core indicators of strategic alignment.

For every region $r$, let $s_{r,2020}\in {\mathbb{R}}^K$ be the region’s vector of topic shares in the final pre-policy year 2020, with components $s_{r,2020,k}={share}_{r,2020,k}$ for $k=1,\dots ,K$. The region’s priority vector $p_r\in {\mathbb{R}}^K$ is defined as:

$$p_{r,k}=\left\{\begin{array}{ll}{\displaystyle \frac{1}{\left|P_r\right|}},& \mathrm{if} k\in P_r,\\ 0,& \mathrm{if} k\notin P_r,\end{array}\right.$$

so that $\sum _{k=1}^K{p}_{r,k}=1$. If a region has no recorded priorities ($\left|P_r\right|=\mathbf{0}$), we set $p_r=0$.

Alignment (cosine similarity):

$${\mathrm{align}}_r={\displaystyle \frac{\sum _{k=1}^K{s}_{r,2020,k}\cdot p_{r,k}}{\sqrt{\sum _{k=1}^K{s}_{r,2020,k}^2}\cdot \sqrt{\sum _{k=1}^K{p}_{r,k}^2}}},$$

with ${\mathrm{align}}_r=0$ if the denominator is zero. This indexes directional agreement between the portfolio and the priority vector on $\left[0,1\right]$, where 1 indicates perfect alignment.

Wishful gap ($L^1$ divergence) [47]:

$${\mathrm{gap}}_r=\sum _{k=1}^K\left|s_{r,2020,k}-p_{r,k}\right|.$$

This captures a significant mismatch on [0, 2]; smaller values indicate portfolios closer to stated priorities. High alignment and low gap suggest evidence-based, portfolio-grounded selection; low alignment and high gap indicate aspirational or mimicry-driven selection disconnected from capabilities.

2.3.6. Opportunity Cost Index (OCI)

To quantify the extent to which regions leave capability potential “on the table” (foregone opportunities), we define the Opportunity Cost Index (OCI) as follows:

$$OC{I}_r={\displaystyle \frac{\sum _k\left(1-{\mathrm{priority}}_{r,k}\right)\times \left({\mathrm{IS}\_\mathrm{potential}}_{r,k}+{\mathrm{AF}\_\mathrm{potential}}_{r,k}\right)\times {mass}_{r,2020,k}}{\sum _k\left({\mathrm{IS}\_\mathrm{potential}}_{r,k}+{\mathrm{AF}\_\mathrm{potential}}_{r,k}\right)\times {mass}_{r,2020,k}}}.$$

$OC{I}_r\in \left[0,1\right]$ measures the proportion of the region’s pre-policy capability mass (in IS + AF potential topics) that is not selected as a priority. A high OCI indicates essential untapped opportunities; a low OCI suggests comprehensive mobilization of available IS and AF.

2.3.7. Strategic Archetype Classification

We classify regions into four strategic archetypes using a 2 × 2 matrix crossing POI with SCZBI. We split both dimensions at the median:

• High POI: $PO{I}_r\ge \mathrm{median}\left(POI\right)$ (many inside strengths);

• Low POI: $PO{I}_r<\mathrm{median}\left(POI\right)$ (few inside strengths, many adjacencies);

• High SCZBI: ${\mathrm{SCZBI}}_r\ge \mathrm{median}\left(\mathrm{SCZBI}\right)$ (comfort-zone bias);
• Low SCZBI: ${\mathrm{SCZBI}}_r<\mathrm{median}\left(\mathrm{SCZBI}\right)$ (frontier exploration).

Therefore, we identified four archetypes:

• Strength Boosters (high POI, high SCZBI): Many existing IS and prioritize them. The process of deepening IS creates conditions that may lead to lock-in situations.
• Excelling Perfectionists (high POI, low SCZBI): Strong capability base but deliberately pursue AF. Optimal related variety strategy.
• Narrow Specialists (low POI, high SCZBI): Limited IS but select non-adjacent or overly ambitious IS.
• Explorers (low POI, low SCZBI): Limited existing IS but pursue AF. High-risk growth strategy requiring external support.

2.3.8. AI Tools

During the preparation of this manuscript, Claude 4.5 Opus was used for the purposes of English editing. The authors have reviewed and edited the output and take full responsibility for the content of this publication.

3. Results

We present results sequentially following our two research questions. Section 3.1 maps regional capability endowments and their evolution (RQ1), establishing the empirical landscape. Section 3.2 characterizes priority selection behavior and strategic archetypes (RQ2). All analyses use data from 236 NUTS-2 regions, 112 energy topics, 178,314 publications and 116,336 projects. The data span 2014–2025 to capture both the pre-policy baseline period (2014–2020), which establishes regional capability endowments, and the RIS3 2021–2027 programming period, which is the focus of our policy analysis.

3.1. Regional Capability Endowments and Evolution (RQ1)

RQ1 asks what energy capability endowments regions possess and how these evolve over time. This section presents results corresponding to the methodological framework developed in Section 2.2. Section 3.1.1 reports the topic similarity and relatedness structure derived from the methodology in Section 2.2.2 and Section 2.2.3 (topic similarity matrix and relatedness density calculations). Section 3.1.2 presents the distribution of capability endowments—inside strengths, adjacent frontiers, and the Portfolio Opportunity Index—operationalized in Section 2.2.4 and Section 2.2.5. Section 3.1.3 analyze portfolio dynamics during the pre-policy period (2014–2020), applying the coverage metrics from Section 2.2.6.

3.1.1. Topic Similarity and Relatedness Structure

The filtered document-topic probability matrix contains 310,057 work-topic pairs with probability values above 0.70 from the original 1.84 million pairs. This matrix was used to construct an 112 × 112 topic similarity matrix for assessing technological proximity within the energy domain. The filtering process effectively minimizes noise by reducing the number of topic pairs with non-zero similarity to 29.1% after applying a 5th-percentile threshold at a similarity level of 0.0003. The energy innovation space exhibits a modular structure rather than uniform interconnectedness. The median non-zero similarity value is 0.007, while the distribution shows a right-skewed pattern extending to 0.99 for the most similar pairs including ‘solar photovoltaics’ and ‘renewable energy systems’. This sparse similarity structure is consistent with the technological relatedness literature [26,45] because energy innovation consists of separate yet partially connected areas that experience restricted knowledge transfer between them.

The RD measure, which weights topic similarities by regional portfolio composition, shows substantial variation (mean = 0.013, median = 0.009, range [0, 1]). This heterogeneity indicates that regions occupy diverse positions in the energy innovation space: some concentrate in tightly connected domains (high RD for many topics) while others pursue scattered unrelated activities (low RD across the board). This variation provides the foundation for distinguishing IS from AF.

3.1.2. Distribution of Capability Endowments

Table 2. POI distribution across EU regions.

Number of Regions	Mean	SD	Min	Q25	Median	Q75	Max	Gini	Mean Number of IS	Mean Number of AF
236	0.655	0.204	0.091	0.538	0.695	0.809	1	0.172	42	21.3

summarizes the distribution of capability endowments across the 236 EU regions in our sample. On average, regions possess strength potential in 42 energy topics and adjacent frontier potential in 21.3 topics, yielding a mean POI of 0.655. The median POI (0.695) is slightly higher, indicating a right-skewed distribution: most regions tilt toward exploitation of existing strengths rather than exploration of adjacent opportunities. However, substantial heterogeneity exists, with POI ranging from 0.091 (a region with few proven strengths but many adjacencies) to 1.0 (regions with proven strengths but no identified adjacent opportunities given their portfolio composition). The Gini coefficient of 0.172 suggests moderate inequality in capability profiles—less concentrated than income or patent distributions, but far from uniform.

Figure 2 visualizes this distribution. The histogram reveals a unimodal distribution centered slightly above the midpoint, with a long left tail representing regions that are “opportunity-rich” (many AF_potential, few IS_potential) and a compressed right tail of “strength-rich” regions (many IS_potential, few AF_potential). The median POI = 0.695 (dashed red line) divides the sample into relatively balanced halves, facilitating our subsequent classification of strategic archetypes in Section 3.2.

Table 3. Capability endowments by POI quartile.

POI Quartile	Number of Regions	Mean POI	SD POI	Mean Number of IS	Mean Number of AF	Mean Total
Q1 (Low POI)	59	0.368	0.139	24.8	37.9	62.7
Q2	60	0.620	0.044	38.2	23.4	61.6
Q3	58	0.762	0.031	48.2	15.0	63.2
Q4 (High POI)	59	0.871	0.048	57.2	8.6	65.8

disaggregates capability endowments by POI quartile, revealing stark differences in regional profiles. Quartile 1 (Low POI, mean = 0.368) comprises regions with an inverted portfolio structure: 24.8 strengths versus 37.9 adjacent frontiers. These are regions at early stages of energy innovation development or those with shallow but broad portfolios—typified by peripheral or post-transition economies building from limited bases. Quartile 4 (High POI, mean = 0.871) comprises mature energy innovation hubs with deep specialization: 57.2 strengths but only 8.6 adjacent frontiers. Notably, the total potential (IS + AF) remains relatively stable across quartiles (62.7 to 65.8 topics), indicating that POI reflects portfolio composition rather than size. This pattern is consistent with life-cycle theories of regional development [6]: regions do not simply accumulate more capabilities monotonically; they transition from broad, shallow adjacency portfolios to narrow, deep strength portfolios as they mature.

Figure 3 plots the trade-off between inside strengths and adjacent frontiers at the regional level, with region labels identifying notable cases. The dashed diagonal line (number of IS equal to the number of AF) marks theoretically balanced portfolios, though the observed data exhibit a clear negative slope, confirming the compositional nature of POI: regions with more inside strengths tend to have fewer adjacent frontiers, and vice versa. Color gradients transition from red (low POI, opportunity-rich) through yellow (moderate POI) to green (high POI, strength-rich), making regional positioning immediately interpretable.

Several patterns emerge. First, the upper-left quadrant contains opportunity-rich regions such as Friesland (NL12), Burgenland (AT11) and Severna Hrvatska (HR06), which possess 50–60+ adjacent frontier topics but relatively few (10–25) established strengths. These represent either peripheral regions building from limited bases or specialized regions with narrow but potentially expandable portfolios. Second, the lower-right quadrant contains strength-rich hubs such as Île-de-France (FR10), Mittelfranken (DE25) and various German Länder, which have accumulated 60–80+ strengths but face diminishing adjacent opportunities (fewer than 20 AF topics). These are mature innovation systems approaching portfolio saturation—they possess deep capabilities but limited obvious adjacencies given their existing composition.

Third, the central cluster around the diagonal exhibits moderate POI (yellow/light green), representing regions with balanced or transitional profiles. Notable examples include regions in Sweden (Småland med öarna, SE21), Italy (Abruzzo, ITF1; Sardegna, ITG2), Denmark (Sjælland, DK02) and Portugal (Algarve, PT15), which maintain roughly 30–45 topics in both IS and AF categories. These regions may have greater strategic flexibility: they can either consolidate existing strengths or pursue adjacencies without facing binding constraints.

Fourth, extreme outliers merit attention. At the far left, Ionia Nisia (EL62), Luxembourg in Belgium (BE34) and Zeeland (NL34) possess almost no inside strengths (0–2 topics) but moderate adjacent frontiers, indicating very early-stage or highly specialized portfolios. At the far right, no region exceeds approximately 80 inside strengths, suggesting an effective ceiling on capability accumulation even in the most advanced regions. Finally, the negative slope (implied correlation between number of IS and AF) is clearly visible and statistically confirmed (Pearson r = −0.18, p = 0.006), consistent with the compositional interpretation: POI reflects portfolio allocation rather than scale. Regions do not simply accumulate both IS and AF indefinitely; they face implicit trade-offs in directing limited research capacity, talent and funding.

Figure 4 maps POI across EU regions, revealing pronounced geographic clustering that partially contradicts simple core-periphery narratives. Low-POI regions (red/orange, opportunity-rich) concentrate in three distinct clusters. First, a Southeastern European belt encompasses Greece (especially Peloponnese and Central Greece), Bulgaria (most regions except for Southwestern, i.e., with Sofia province), Romania, Croatia and parts of Italy (Sardegna, Umbria, Marche, Abruzzo). Second, a Central European corridor includes Hungary, Eastern Austria, Western Slovakia, the Czech Republic, Eastern and Northeastern Poland and the Baltic states (Lithuania, Latvia). Third, peripheral pockets appear in the Netherlands, Belgium and Central Norway. These low-POI regions are not innovation deserts—Table 3 shows Q1 regions average 37.9 adjacent frontiers—but they lack established strengths, indicating portfolios dominated by potential rather than proven capabilities.

High-POI regions (green, strength-rich) exhibit clear geographic concentration. Germany displays nearly uniform high POI across all Länder, with particularly strong performance in Bavaria (Oberbayern, Mittelfranken), Baden-Württemberg, North Rhine-Westphalia and the former East German regions (Saxony, Thuringia). France shows high POI in Île-de-France, Auvergne-Rhône-Alpes and Bourgogne. Spain displays a mixed pattern with high POI in Madrid, Catalonia and Valenciana, but lower POI in La Rioja and Cantabria. Northern and Central Italy (Lombardy, Emilia-Romagna, Veneto, Lazio, Piedmont) exhibit high POI, contrasting sharply with the South. Scandinavia presents a split: northern Denmark and southern Sweden show high POI, while central Sweden and Finland exhibit more moderate values. Moderate-POI regions also include much of Western Poland (excluding Lubuskie), Estonia and part of Ireland. These represent either regions in transition from opportunity-rich to strength-rich profiles, or regions maintaining deliberately balanced portfolios.

The spatial pattern suggests that POI reflects industrial maturity and innovation system thickness [33,34] rather than simple GDP per capita. Notable deviations from wealth-based predictions include: (1) Eastern German regions (Saxony, Thuringia) exhibit high POI despite lower GDP than Western counterparts, likely reflecting inherited industrial capabilities from reunification and targeted Cohesion Policy investments; (2) Parts of the wealthy Netherlands and Belgium show surprisingly low POI, potentially due to narrow specialization in non-energy sectors (financial services, logistics); (3) Central European regions (Poland, the Czech Republic, Slovakia, Hungary) show within-country heterogeneity, with Western/capital regions exhibiting moderate-to-high POI while Eastern/peripheral regions remain opportunity-rich.

From a policy perspective, this geography has important implications. Low-POI regions in Southeast and Central Europe face a dual challenge: they possess adjacency potential but lack the absorptive capacity [28] to convert it into strengths without sustained support. High-POI regions in Germany, France and Northern Italy risk lock-in and portfolio saturation—they have deep capabilities but few obvious adjacencies. RIS3 interventions should therefore be differentiated: low-POI regions require capacity-building and foundational investments to activate adjacent frontiers, while high-POI regions need mechanisms to break path dependencies and re-open exploration [6].

3.1.3. Portfolio Dynamics in the Pre-Policy Period (2014–2020)

Figure 5 tracks portfolio coverage—the fraction of IS and AF potential topics with substantial activity—over the pre-policy period 2014–2020. Mean coverage hovers around 0.59 throughout, indicating that on average, regions actively pursue roughly 60% of their capability space (IS_potential + AF_potential) with above-median research intensity. The 95% confidence bands remain narrow and exhibit no clear temporal trend, consistent with the panel regression results reported below. This stability is notable: despite significant external shocks during this period (Paris Agreement 2015, Clean Energy Package 2016–2019, COVID-19 onset in 2020), regional energy innovation portfolios exhibit remarkable inertia.

Table 4. Coverage dynamics regression results.

Dependent Variable	Coverage (Fraction of IS + AF Potential with Substantial Activity)
Specification	Coverage~Year + Region FE, clustered SE (n = 235 regions)
Observations	1608 region–year pairs
Year coefficient (β)	0.00142 (SE: 0.00124)
t-statistic	1.143
p-value	0.254
Within R2	0.0012
Adjusted R2	0.248

formalizes the visual impression from Figure 5. The year coefficient β = 0.00142 is substantively small (implying a 1-percentage-point increase in coverage over 7 years) and statistically insignificant (p = 0.254). The within R² of 0.001 confirms that time trends explain virtually none of the variation in coverage after absorbing region fixed effects. We conclude that regional portfolios are static during 2014–2020: there is no evidence of organic broadening or concentration prior to the 2021–2027 policy intervention. This finding has an important implication.

From a substantive standpoint, portfolio inertia suggests that without deliberate policy intervention, regions are unlikely to spontaneously diversify into adjacent energy domains despite possessing the relatedness structure to do so. This locks in path-dependent specializations [6,35] and may perpetuate core-periphery divides. The strong stability of coverage (adjusted R² = 0.248 driven entirely by region fixed effects) implies that portfolio composition is deeply rooted in regional capabilities, institutions and industrial structures—factors that change slowly absent exogenous shocks. RIS3, by design, aims to be such a shock: a coordinated effort to redirect regional portfolios toward strategic priorities.

3.2. Priority Selection Behavior: Rational, Explorative or Mimicry? (RQ2)

RQ2 examines how regions select energy priorities under RIS3 2021–2027. The smart specialization literature emphasizes evidence-based entrepreneurial discovery [13,14,39], suggesting regions should align priorities with capability endowments—either by exploiting strengths (rational selection) or exploring adjacent frontiers (explorative selection). However, institutional isomorphism theory [19] and recent empirical work [20] suggest that policy mimicry may dominate, with regions “following peers” or “following role models” rather than “following indicators.”

This section presents results corresponding to the methodological framework developed in Section 2.3. Section 3.2.1 provides a descriptive overview of selection behavior using the metrics defined in Section 2.3.1, Section 2.3.2, Section 2.3.3, Section 2.3.4 and Section 2.3.5 (priority flags, positioning tags, SCZBI, selection rates and alignment measures). Section 3.2.2 presents the strategic archetype classification based on the 2 × 2 typology introduced in Section 2.3.7. Section 3.2.3 tests the three hypotheses on selection behavior (H1–H3 from Section 1.3.2), examining relationships between capability endowments and selection patterns. Section 3.2.4 explores spatial patterns and legacy effects, investigating whether pre-policy portfolio dynamics (Section 2.2.6) and geographic clustering of SCZBI (Section 2.3.3) shape selection behavior.

3.2.1. Descriptive Overview of Selection Behavior

Table 5. Selection behavior statistics.

Number of Regions	Metric	Mean	SD	Min	Q25	Median	Q75	Max
182	SCZBI	0.231	0.399	–1.000	0.000	0.232	0.500	1.000
182	ER	0.062	0.051	0.000	0.022	0.055	0.091	0.250
182	ExR	0.053	0.063	0.000	0.000	0.039	0.083	0.308
182	SR	0.403	0.293	0.000	0.186	0.375	0.571	1.000
182	alignment	0.102	0.103	0.000	0.007	0.078	0.160	0.365
182	gap	1.851	0.175	1.000	1.778	1.904	1.985	2.000
182	OCI	0.927	0.092	0.491	0.884	0.957	0.999	1.000

summarizes selection behavior across the 182 EU regions that declared energy priorities in RIS3 2021–2027. The mean Selection Comfort-Zone Bias Index (SCZBI) is 0.231 (median 0.232, range [−1, 1]), indicating a slight population-level tilt toward selecting inside strengths (IS) over adjacent frontiers (AF). However, substantial heterogeneity exists: 25% of regions exhibit SCZBI ≤ 0 (exploration bias), 50% fall between 0 and 0.5 (balanced or moderate comfort-zone bias), and 25% exceed 0.5 (strong comfort-zone bias). This variation suggests diverse strategic orientations—some regions consolidate strengths, others pursue frontiers—consistent with heterogeneous capability profiles documented in RQ1.

Yet beneath this surface diversity lies a striking commonality: universally low mobilization of capability potential. The Exploitation Rate (ER) averages only 6.2% (median 5.5%), meaning regions select as priorities fewer than one-tenth of their available strengths. The Exploration Rate (ExR) is similarly anemic at 5.3% (median 3.9%), indicating regions pursue a tiny fraction of their adjacent frontiers. In stark contrast, the Stretch Rate (SR) averages 40.3% (median 37.5%), with 75% of regions exceeding 18.6%. This means that on average, four in ten regional energy priorities fall outside both potential IS and potential AF—domains where regions possess neither demonstrated strength nor adjacency readiness.

Portfolio–priority concordance metrics reinforce this interpretation. Mean alignment is 0.10 (median 0.08), far below moderate concordance (0.5 threshold) and indicating minimal overlap between priorities and pre-policy activity portfolios. The wishful gap averages 1.8 (median 1.9), approaching the theoretical maximum of 2.0 (complete non-overlap). The Opportunity Cost Index (OCI) averages 0.93, revealing that regions leave 93% of their combined IS and AF potential unselected. Taken together, these descriptors suggest that priority selection is largely decoupled from capability endowments: regions neither systematically exploit strengths nor systematically explore adjacencies, but instead allocate the plurality of priorities to distant, aspirational domains.

3.2.2. Strategic Archetype Classification

We classify regions into four strategic archetypes (see Figure 1) based on the 2 × 2 matrix crossing POI (capability endowment) with SCZBI (selection behavior).

Table 6. Strategic archetype characteristics.

Archetype	Number of Regions	Mean POI	Mean SCZBI	Mean ER	Mean ExR	Mean SR	Mean Alignment	Mean Gap	Mean Number of Priorities
Strength Boosters	67	0.833	0.536	0.063	0.033	0.367	0.121	1.855	6.209
Excelling Perfectionists	25	0.787	0.060	0.036	0.083	0.637	0.060	1.904	6.680
Narrow Specialists	24	0.560	0.553	0.083	0.025	0.253	0.169	1.806	5.208
Explorers	66	0.480	−0.131	0.063	0.072	0.406	0.074	1.844	6.667

and Figure 6 present the results. Strength Boosters (36.8%, n = 67) exhibit high POI and high SCZBI, combining strength-rich portfolios with comfort-zone selection—typified by German Länder (Bayern, Baden-Württemberg), Northern and Western Italy (Piemonte, Lombardia, Veneto, Emilia-Romagna, Toscana, Lazio), Slovenia, Ireland, Western France, Central Spain, Northern Portugal. Excelling Perfectionists (13.7%, n = 25) combine high POI with low SCZBI, utilizing strong bases to pursue adjacencies—appearing in select German regions, especially in East Germany, some Central and Western Polish regions (Warsaw, Łódzkie, Wielkopolskie, Dolnośląskie), scattered French and Spanish regions and most of Finland. Narrow Specialists (13.2%, n = 24) pair low POI with high SCZBI, attempting ambitious leaps from weak foundations—forming a rather fragmented region in Central Europe (Austria, Croatia, Lubelskie in Poland, Nord-Vest in Romania), a few regions in Germany, Spain, Belgium, Denmark and Sweden. Explorers (36.3%, n = 66) display low POI and low SCZBI, pairing adjacency-rich portfolios with frontier-oriented selection—concentrated in Central and Eastern Europe (Poland, Romania, Bulgaria), Western regions (France, the Netherlands, Denmark, Ireland) and parts of Southern Europe (Greece, Southern and Eastern Italy, upper central Spain).

However, archetype labels mask a deeper uniformity: all four groups exhibit high Stretch Rates (Table 6, column 7). Strength Boosters dedicate 36.7% to stretch, while Excelling Perfectionists, despite low SCZBI implying AF focus, allocate 63.7% of priorities to stretch domains—the highest rate of any archetype. Explorers dedicate 40.6% to stretch, and even Narrow Specialists exhibit “only” 25.3% SR because they also select some IS (ER = 8.3%, highest among archetypes). The implication is that stretch behavior is not deviant but normative: it pervades all strategic orientations, suggesting systemic forces (policy trends, EU narratives, mimicry) shape selection beyond capability logic.

Figure 7 visualizes ER, ExR and SR by archetype with 95% confidence intervals. The dominant visual feature is the tall orange bars (SR), which overshadow the red (ER) and blue (ExR) bars in all groups. This graphically confirms that, regardless of POI or SCZBI, stretch dominates selection. Excelling Perfectionists exhibit the highest SR (64%) despite possessing the strongest capability bases (mean POI = 0.787, Table 6), while Narrow Specialists—paradoxically—show the lowest SR (25%) alongside the highest ER (8.3%). This inversion contradicts expectation: weak-capability regions should not be more grounded in their (limited) strengths than strong-capability regions. The pattern suggests that strong regions feel greater ‘license to dream,’ while weak regions select opportunistically from scarce activity.

3.2.3. Testing Hypotheses on Selection Behavior

We now test three competing hypotheses about the mechanisms driving selection: rational (H1), explorative (H2) or mimicry-driven (H3).

First, we test H1 (Rational selection hypothesis), i.e., “Strength-Based Strategies”. If regions engage in evidence-based entrepreneurial discovery [13,14,39], they should align priorities with capability endowments. Regions with many strengths (high POI) should exploit them (high SCZBI, Strength Boosters). Regions with few strengths but clear niches (low POI) should focus narrowly on those domains (high SCZBI, Narrow Specialists). Both strategies reflect capability-grounded rationality, yielding a positive or null correlation between POI and SCZBI among comfort-zone-biased regions (upper half of POI × SCZBI space) and high Exploitation Rates (ER > 15%) among high-POI regions. This hypothesis aligns with the “follow the indicators” logic [20].

To test this hypothesis, we analyze the correlation within the SCZBI > 0 subsample. We subset the sample to regions exhibiting comfort-zone bias (SCZBI > 0, n = 120) and test whether POI predicts SCZBI intensity. The Pearson correlation yields r = 0.374 (p < 0.0001), statistically significant and positive. This supports H1: among regions tilting toward IS selection, those with more strengths (high POI) do indeed exhibit stronger comfort-zone bias. However, the correlation is moderate (r < 0.5), and the explained variance (R² = 0.14) is modest, indicating that capability explains only 14% of selection intensity among rational selectors.

Then, we test the Exploitation Rate among high-POI regions. We examine the mean ER in the top POI quartile (n = 59, POI ≥ 0.82). The mean ER is 6.8% (95% CI: [5.4%, 8.2%]), far below the 15% threshold that would indicate systematic exploitation. Even among the most strength-rich regions, fewer than 7% of available strengths are prioritized. H1 is partially supported: capability does predict selection direction (positive r), but not selection magnitude (low ER). Regions exhibit rational orientation (choosing IS over AF when they have more IS) but do not mobilize their strengths at scale.

We now move on to testing hypothesis H2 (explorative selection hypothesis), i.e., “Growth-Oriented Strategies”. It states that regions, alternatively, may pursue exploration over exploitation [36] by targeting adjacent frontiers to build new capabilities through related variety [6]. High-POI regions may branch into adjacencies (low SCZBI, Excelling Perfectionists), while low-POI regions may build from abundant AF potential (low SCZBI, Explorers). Both strategies reflect organizational learning, yielding a positive or null correlation between POI and SCZBI among exploration-biased regions (lower half of POI × SCZBI space) and high exploration rates (ExR > 15%) among low-SCZBI regions.

To test this hypothesis, we conduct two tests. First, we analyze the correlation within the SCZBI < 0 subsample. We subset the sample to regions with exploration bias (SCZBI < 0, n = 62) and test whether POI predicts the intensity of AF selection. The Pearson correlation yields r = 0.184 (p = 0.152), positive but statistically insignificant. This provides weak support for H2: among exploratory regions, higher POI is associated with slightly lower exploration bias (less negative SCZBI), consistent with stronger regions having fewer adjacencies to pursue (per Table 3, Q4), but the relationship is not robust.

Then, we test the Exploration Rate among low-SCZBI regions. The mean ExR in the exploitation-biased subsample (SCZBI < 0) is 7.2% (95% CI: [5.1%, 9.3%]), again far below 15%. Even among regions explicitly tilting toward AF, fewer than one-tenth of adjacent frontiers are selected. H2 is weakly supported: exploration orientation exists (negative SCZBI), but like exploitation, it does not translate into high mobilization rates. Regions prefer AF over IS in their limited selections, but most potential AF remains untapped.

Both rational and explorative selection logics operate directionally—POI correlates with SCZBI as predicted (Figure 8 shows clear quadrant structure)—but neither drives intensive capability mobilization. ER and ExR remain universally low (~5–8%), far below thresholds indicating systematic evidence-based selection. This suggests that while capability exerts a weak influence on which type of priority to select (IS vs. AF), it does not determine how much of the capability base to activate. The question becomes: what does determine selection, if not capabilities?

Therefore, we move to the third hypothesis, H3 (mimicry and wishful thinking hypothesis), i.e., “Capability-Decoupled Selection”. Despite normative aspirations for evidence-based selection, institutional isomorphism [19] and policy mimicry [20] may dominate. Regions may “follow peers” or “follow role models” by selecting globally prominent domains (hydrogen economy, offshore wind, battery storage) regardless of capability fit. This produces low portfolio–priority alignment (<0.3), high wishful gaps (>1.5), high Stretch Rates (>30%), and critically, substitution rather than addition: high SR should crowd out ER and ExR (negative coefficients in regressions) as regions have a limited “priority budget” and allocate slots to wishes rather than capabilities.

We run two tests. First, we analyze alignment and gap thresholds. Mean alignment is 0.102 (95% CI: [0.087, 0.117]), far below 0.3 and decisively rejecting the hypothesis that priorities moderately align with portfolios (one-sample t-test vs. μ = 0.3: t = –26.03, p < 0.0001). Mean wishful gap is 1.851 (95% CI: [1.826, 1.877]), exceeding 1.5 and approaching the maximum of 2.0 (t-test vs. μ = 1.5: t = 27.13, p < 0.0001). Mean SR is 40.3% (95% CI: [36.0%, 44.6%]), exceeding 30% (t-test vs. μ = 0.3: t = 4.77, p < 0.0001). All three thresholds are decisively confirmed: priorities are minimally aligned with portfolios, maximally distant in composition and heavily tilted toward stretch domains.

Figure 9 disaggregates alignment and gap by archetype, revealing that low alignment is universal, not archetype-specific. Even Narrow Specialists, who exhibit the highest median alignment (~0.19), remain far below 0.5. Excelling Perfectionists, despite possessing the strongest capability bases (mean POI = 0.787), show the lowest alignment (~0.04)—an inversion suggesting that strong regions feel greater freedom to pursue aspirational pivots. All archetypes exhibit wishful gaps near 2.0, confirming a systematic disconnect between priorities and portfolios.

Then, we run tests on the substitution mechanism. To test whether high SR adds to or substitutes for capability-based selection, we estimate regressions with ER and ExR as dependent variables, including SR as a predictor alongside POI, legacy and controls (

Table 7. Selection behavior predictors with SR.

Term	Model 1 ER	Model 2 ExR	Model 3 SR
(Intercept)	0.245 (0.128)	0.238 (0.178)	2.280 ** (0.851)
POI	0.012 (0.022)	0.018 (0.03)	0.146 (0.145)
SR	−0.062 *** (0.012)	−0.054 ** (0.016)	not applicable
Legacy	0.016 (0.013)	−0.004 (0.018)	0.233 ** (0.086)
Logarithm of GDP per capita (in purchasing power standard)	−0.019 (0.013)	−0.020 (0.018)	−0.203 * (0.085)
Logarithm of population density	−0.006 (0.004)	−0.004 (0.005)	0.043 (0.027)
GERD (as % of GDP)	−0.005 (0.003)	−0.01 * (0.005)	0.017 (0.022)
Share of fossil fuels (country level)	0.001 * (0.000)	0.001 ** (0.000)	−0.005 ** (0.002)

Note: *** p < 0.001, ** p < 0.01, * p < 0.05.

, Models 1–2).

In Model 1, SR exhibits a large, negative, highly significant coefficient (β = −0.072, p < 0.001). For every 10-percentage-point increase in SR, ER declines by 0.7 percentage points. Given a mean ER = 6.2%, this implies that regions at the 75th percentile of SR (57.1%, Table 5) have an ER approximately 3 percentage points lower than those at the 25th percentile (18.6%)—a 50% reduction. Critically, POI remains insignificant (β = −0.001, p = 0.959) even after controlling for SR, confirming that capability does not predict exploitation intensity once wishful thinking is accounted for.

In Model 2, SR again exhibits a negative, significant coefficient (β = −0.063, p < 0.001), indicating similar crowding-out of exploration. POI remains insignificant (β = 0.018, p = 0.516). The pattern is consistent: stretch substitutes for, rather than complements, capability-based selection.

Additional controls reveal intriguing patterns. GDP per capita negatively predicts both ER (β = −0.029, p = 0.018) and ExR (β = −0.024, p = 0.149), suggesting wealthier regions select less from their capabilities—consistent with the ‘license to dream’ interpretation. GERD intensity also negatively predicts ExR (β = −0.011, p = 0.020), implying that stronger R&D capacity paradoxically reduces capability mobilization. Legacy exhibits a marginally positive effect on ER (β = 0.023, p = 0.084), hinting at weak learning effects, but the coefficient is small and barely significant.

The regression evidence is unambiguous: regions face a “priority budget” constraint, and when they allocate slots to stretch domains (chasing hydrogen, offshore wind or other EU policy darlings), they do not also select IS or AF—they instead select wishes. This is not additive but substitutive, indicating that mimicry and wishful thinking crowd out evidence-based selection. The mechanism operates as follows: regions participate in entrepreneurial discovery processes, stakeholder consultations and strategy workshops, yet the final priority lists reflect aspirations (signaling to EU funders, emulating successful role models, following policy trends) more than grounded capability assessments. As documented in Pylak et al. [20], regions “follow peers” and “follow role models” rather than “follow indicators.”

H3 is strongly supported. Low alignment, high gaps, high SR and the substitution mechanism all confirm that mimicry/wishful thinking dominates selection in practice, even though rational and explorative logics exert directional influence (H1–H2). The positive POI–SCZBI correlation (r = 0.477, Figure 8) initially suggested rationality, but the low absolute ER/ExR rates and strong negative SR coefficients reveal this to be surface rationality masking deeper mimicry. Regions gesture toward their capabilities when choosing between IS and AF but allocate most priorities to neither.

3.2.4. Spatial Patterns and Legacy Effects

Figure 10 maps SCZBI across EU regions, exhibiting prominent spatial clustering that partially reflects, but also complicates, the capability gradient from RQ1. Comfort-zone bias (green, SCZBI > 0.5) concentrates in Ireland, Western France, Northern Italy, Northern Greece, Croatia, Austria, Denmark and part of Sweden. These are mature innovation systems with a high POI (Figure 4, RQ1), selecting strengths as predicted by H1. Exploration bias (red/orange, SCZBI < 0) appears in scattered regions: the Netherlands (Noord-Nederland), France (Corsica, Centre—Val de Loire, Hauts-de-France), Spain (La Rioja) and Sweden (Västerbottens län). Interestingly, the red belt of regions in Germany with low SCZBI scores also has low POI scores, suggesting that these regions are deliberately seeking to explore new areas, even if their internal strengths are not numerous.

Legacy effects are modest. Regions with 2014–2020 energy priorities (Legacy = 1, n = 165) exhibit slightly higher mean SCZBI (0.242 vs. 0.124), ER (6.3% vs. 4.7%) and SR (41.1% vs. 32.7%) than non-legacy regions (n = 17), but t-tests reveal no significant differences (all p > 0.10) except a marginal effect in the ER regression (Table 7, β = 0.023, p = 0.084). This suggests that prior RIS3 experience does not substantially alter selection behavior—both legacy and non-legacy regions engage in high-stretch, low-capability mobilization prioritization. The lack of a legacy premium in selection implies that regions do not “learn” to select more strategically over successive cycles, or alternatively, that the 2021–2027 period introduced new pressures (European Green Deal, hydrogen economy hype, offshore wind narratives) that reset selection dynamics regardless of history.

4. Discussion and Conclusions

This study set out to understand how EU regions select energy priorities under RIS3 2021–2027, asking whether selection aligns with capability endowments—exploiting strengths or exploring adjacent frontiers—or follows other logics such as policy mimicry and aspirational targeting. By constructing a comprehensive empirical landscape of regional energy innovation capabilities and analyzing priority selection behavior across 236 NUTS-2 regions, we uncover a striking pattern: surface rationality masks deep mimicry. While capability endowments exert directional influence on selection (regions with more strengths are more likely to select them), this influence is shallow. Regions mobilize fewer than 7% of their capabilities and instead allocate 40% of their priorities to distant, aspirational domains disconnected from their portfolios. We now place these findings in the broader context of the literature, discuss theoretical and policy implications, acknowledge limitations and identify directions for future research.

4.1. Key Findings and Contributions

Our analysis yields two sets of integrated findings. First, regarding regional capability endowments (RQ1), the energy innovation space exhibits a clear modular structure, with only 29% of topic pairs showing non-zero similarity, indicating that technological relatedness shapes feasible diversification pathways as predicted by evolutionary economic geography [6,25]. Substantial heterogeneity exists in regional capability profiles: POI ranges from 0.09 to 1.0, with low-POI regions characterized by many adjacencies but few strengths (opportunity-rich) and high-POI regions exhibiting the reverse (strength-rich). This distribution aligns with life-cycle theories of regional development [33,34], wherein regions transition from broad, shallow portfolios to narrow, deep specializations as they mature.

Critically, portfolios proved remarkably static during 2014–2020 (coverage trend β = 0.0014, p = 0.254), suggesting that organic diversification into adjacent energy domains is rare without deliberate policy intervention. This finding resonates with the path-dependence literature emphasizing regional lock-in [48] and the need for exogenous shocks—such as RIS3 policy interventions—to disrupt entrenched trajectories [6]. The absence of pre-trends also strengthens the parallel trends assumption for future causal analyses, as treated and control topics exhibit flat evolution before the 2021 policy window.

Second, regarding priority selection behavior (RQ2), our findings point to a clear conclusion: rational and exploratory logics operate directionally, but mimicry and wishful thinking dominate. Three hypotheses were tested. H1 (rational selection) posited that regions align priorities with capability endowments, with high-POI regions exploiting their strengths. We find partial support: POI and SCZBI correlate positively (r = 0.477, p < 0.0001), indicating capability influences selection direction. However, this correlation explains only 23% of variance (R² = 0.23), and Exploitation Rates remain dismally low (mean ER = 6.2%), far below the 15% threshold that would indicate systematic mobilization. H2 (explorative selection) proposed that low-POI regions or exploration-biased regions (low SCZBI) would pursue adjacent frontiers. Again, partial support emerges: regions with low SCZBI do exhibit slightly higher exploration rates (ExR = 7.2% vs. 4.7% among high-SCZBI regions), but absolute rates remain minimal, and most potential AF goes untapped.

H3 (mimicry and wishful thinking) receives decisive support across multiple tests. First, portfolio–priority alignment is minimal (mean = 0.10, far below the 0.5 threshold for moderate concordance), and wishful gaps approach the theoretical maximum (mean = 1.85 out of 2.0), confirming that priorities are systematically decoupled from pre-policy portfolios. Second, Stretch Rate dominates universally: approximately 40.3% of priorities fall outside both potential IS and potential AF, indicating widespread targeting of aspirational domains. Third, and most critically, high SR substitutes for rather than complements capability-based selection: multivariate regressions reveal that SR negatively predicts both ER (β = −0.072, p < 0.001) and ExR (β = −0.063, p < 0.001), demonstrating that when regions allocate “priority slots” to wishes, they do not also select IS or AF—they select wishes instead. This crowding-out mechanism confirms that regions face a “priority budget” constraint and that mimicry displaces evidence-based selection. Finally, the Opportunity Cost Index averages 0.927, meaning regions leave 93% of their combined IS and AF potential immobilized—a staggering degree of foregone capability activation.

4.2. Theoretical Implications: Reconceptualising Smart Specialization

Our findings contribute to three theoretical debates. First, we advance the smart specialization literature by providing the first large-scale empirical test of whether regions “follow the indicators” (capability-grounded selection) or “follow peers” and “follow role models” (mimicry-driven selection), as theorized by Pylak et al. [20]. While Foray et al. [14] and subsequent normative work [13,39] emphasize evidence-based entrepreneurial discovery wherein stakeholders identify regional strengths and adjacencies through data-driven processes, our evidence suggests this vision is not realized in practice for energy priorities in 2021–2027. Instead, regions engage in aspirational positioning: they select globally prominent domains most likely to signal ambition, attract EU co-funding under Horizon Europe and Cohesion Policy instruments and align with European Commission narratives (i.e., European Green Deal, Fit for 55 package, REPowerEU). This behavior aligns with institutional isomorphism theory [19], wherein organizations adopt similar practices not for efficiency but due to coercive (EU funding conditionality), normative (consultant “best practices”) and mimetic (emulating successful role models) pressures.

Importantly, our finding of surface rationality masking deep mimicry adds a new dimension to the debate. Regions are not entirely irrational as capability does predict selection direction (r = 0.48), and the four strategic archetypes exhibit distinct POI × SCZBI profiles. However, this rationality is shallow: it influences which type of priority to select (IS vs. AF) but does not determine whether to mobilize capabilities at scale. The substitution mechanism (SR crowds out ER/ExR) reveals that mimicry is not merely one force among many but the dominant logic, displacing capability-based selection. This challenges the entrepreneurial discovery paradigm’s assumption that regional stakeholders, when properly engaged, will converge on evidence-based priorities. Our results suggest that even when stakeholder processes occur, the final priority lists reflect political economy dynamics—lobbying by aspirational sectors (hydrogen associations, offshore wind consortia), consultant-driven template strategies and competitive signaling to EU evaluators—more than grounded capability assessments.

We should note an important distinction between our descriptive findings and their interpretation. The empirical patterns we document—low exploitation and exploration rates, high Stretch Rates and minimal portfolio–priority alignment—are robust descriptive results. However, the mechanisms driving these patterns cannot be directly observed. While we interpret these findings through the lens of institutional isomorphism and policy mimicry, alternative explanations merit consideration. Some degree of aspirational stretch may reflect rational strategic behavior: regions may deliberately position themselves for emerging industries (hydrogen, offshore wind) anticipating future capability development or respond strategically to EU funding incentives that reward alignment with Green Deal narratives. From this perspective, “following the policy narrative” could represent informed adaptation rather than unreflective imitation. Our data cannot definitively distinguish mimetic behavior from strategic signaling or rational anticipatory positioning. What the evidence does establish is that current RIS3 energy priorities are largely disconnected from existing regional capabilities—whether this disconnection reflects mimicry, strategic aspiration or funding-driven adaptation remains an important question for future research employing qualitative methods or natural experiments.

Second, we contribute to evolutionary economic geography by operationalizing and empirically validating the concepts of inside strengths and adjacent frontiers through relatedness density measures. Our finding that regions with high relatedness density to a topic (high RD) face lower barriers to entry aligns with the branching logic [6,26] and related variety mechanisms [29]. However, the disconnect between potential (high AF_potential) and selection (low ExR) reveals a critical gap: adjacency advantages are necessary but insufficient for diversification. Regions may possess the cognitive proximity and knowledge base to branch into adjacent energy domains, yet they systematically fail to prioritize them, instead chasing distant domains. This suggests that policy interventions aiming to catalyze branching—such as RIS3—must do more than identify adjacencies; they must create incentive structures that reward capability-grounded exploration over aspirational leaps. Our Stretch Rate (SR) metric, capturing priorities outside both IS and AF, provides a novel tool for future research to measure “capability distance” and test whether distant diversification attempts succeed or fail.

Third, we inform the policy learning and mimicry literature. Recent work documents that regions emulate peers’ RIS3 priorities [20], import consultant-designed strategies [49,50] and respond to EU Commission signals about “priority sectors” (such as hydrogen, batteries, etc.). Our contribution is to show that mimicry is not confined to what regions select (topic labels) but extends to how they select: by systematically under-mobilizing capabilities (low ER/ExR) and over-allocating to wishes (high SR). This implies that addressing mimicry requires more than diversifying regional strategies (reducing convergence on hydrogen/wind); it requires fundamentally altering the selection process to anchor priorities in absorptive capacity assessments [28]. Without such anchoring, even regions with diverse priority lists may simply be engaging in “diverse wishful thinking.”

4.3. Policy Implications: Rethinking Smart Specialization Design

If RIS3 aims to enable evidence-based entrepreneurial discovery that mobilizes regional strengths and builds on adjacencies [13,14,39], our findings suggest this vision faces severe implementation challenges. Instead of systematic capability activation (which would yield high ER and ExR), we observe aspirational signaling (high SR, low alignment, high OCI). Three policy implications follow. These recommendations are directed at multiple institutional levels: regional managing authorities and RIS3 coordinators responsible for strategy design; national ministries overseeing innovation and regional development policy; and EU institutions—particularly DG REGIO and the JRC Smart Specialization Platform—that set guidelines, evaluate strategies and allocate Cohesion Policy funding.

First, strengthen mechanisms anchoring priorities in capability assessments. Current RIS3 guidelines emphasize stakeholder consultation and entrepreneurial discovery but provide limited tools to ensure selections reflect absorptive capacity. Our Exploitation Rate (ER), Exploration Rate (ExR) and alignment metrics offer actionable indicators that regional authorities and EU evaluators could use to audit strategies. Regions with ER < 10%, ExR < 10% and alignment < 0.2 should be flagged for revision, as their priorities are demonstrably disconnected from portfolios. Requiring regions to justify each priority by citing pre-existing strengths (publications, patents, firms, projects) or demonstrating adjacency (relatedness density calculations) would impose discipline on selection. Alternatively, performance-based funding could reward regions that show high ER/ExR and low SR in implementation, creating incentives to mobilize capabilities rather than chase headlines.

Second, differentiate policy instruments by strategic archetype. Our typology—Strength Boosters, Excelling Perfectionists, Narrow Specialists and Explorers—reveals that regions occupy distinct positions in capability-selection space, yet EU and national policies often apply uniform instruments (competitive calls, tax credits, cluster programs). Excelling Perfectionists (high POI, low SCZBI), who utilize strong bases to pursue adjacencies, may benefit from challenge-driven innovation instruments (e.g., Horizon Europe Missions) that direct branching toward societal goals. Strength Boosters (high POI, high SCZBI), who deepen narrow specializations, risk lock-in and may require disruption incentives (e.g., funding for radical pivots, support for entrepreneurial entry in unrelated domains). Explorers (low POI, low SCZBI), who build from weak bases into adjacencies, face steep learning curves and may need capacity-building support (training, infrastructure, institutional strengthening) before competitive instruments are effective. Narrow Specialists (low POI, high SCZBI), who attempt distant leaps, should be counseled toward realism: redirecting priorities toward their limited IS or abundant AF rather than pursuing unattainable stretch goals. One-size-fits-all RIS3 implementation ignores these differences.

Third, address systemic mimicry through transparency and benchmarking. If regions “follow peers” due to information asymmetries—uncertain about what constitutes “good” priorities—then publicizing selection metrics (ER, ExR, SR, alignment) for all regions could reduce mimicry by making capability-grounded strategies visible. Creating a RIS3 dashboard showing which regions achieve high ER/ExR (evidence-based selectors) and which exhibit high SR (aspirational selectors), alongside post-policy outcome measures (to be developed in future research), would allow peer learning based on performance rather than headlines. Currently, regions emulate Germany’s hydrogen strategy or Denmark’s offshore wind strategy because these are high-profile, but they lack data on whether such strategies actually worked for those regions’ capabilities. Transparency might shift mimicry from “follow the loudest” to “follow the effective.”

However, we acknowledge a deeper challenge: mimicry may be rational for individual regions even if collectively suboptimal. If EU Horizon Europe and Cohesion Policy funding flows preferentially to regions that signal alignment with Commission priorities (Green Deal, hydrogen economy), then selecting stretch domains maximizes funding probability even if it minimizes capability fit. All regions declare ambitious plans for hydrogen/offshore wind energy, which undermines the credibility of individual declarations, but no region can afford to take a realistic approach and risk losing funding. To solve this problem, the criteria for allocating EU funds need to be reformed to reward alignment of capabilities rather than alignment of narratives—a politically difficult but necessary step.

4.4. Limitations and Boundary Conditions

Our study has several limitations that bound the scope of the conclusions. First, we focus on energy domains, which may exhibit unique dynamics given the European Green Deal’s prominence and the hydrogen economy hype during 2019–2021 (when RIS3 2021–2027 strategies were designed). Mimicry pressures may be weaker in less politicized domains (e.g., advanced materials, aggrotech). Future research should replicate our two-dimensional framework (POI and SCZBI) in other sectors to test generalizability.

Second, we measure priorities, not outcomes. Although we show that regions make choices based on aspirations, we do not examine whether such choices lead to innovative actions or whether priorities based on capabilities outperform those based on wishful thinking. This causal question—central to policy design—requires difference-in-differences analysis comparing treated (prioritized) and control (non-prioritized) topics over 2021–2025, which we reserve for future work. If high-SR priorities nonetheless succeed (e.g., because they attract disproportionate EU funding or catalyze entrepreneurial entry), then aspirational selection may be defensible as forward-looking foresight rather than wishful thinking. Conversely, if high-ER/low-SR priorities systematically outperform, this would vindicate the evidence-based paradigm.

Third, our capability measures (potential IS and AF) rely on pre-2021 publication and project activity, which may not capture tacit knowledge, firm capabilities or human capital stocks. A region with low publication-based POI might nonetheless possess industrial strengths (e.g., manufacturing expertise applicable to energy transitions) not visible in our data. Integrating patent data, firm surveys or labor market analytics would enrich capability measurement.

Fourth, we cannot directly observe mimicry mechanisms—whether regions copied peers, followed consultants or responded to EU signals. Survey or interview data on RIS3 strategy formulation processes would complement our revealed-preference analysis.

4.5. Future Research Directions

Our findings open multiple research avenues. First and foremost, causal evaluation of whether priority selection matters for outcomes is needed. Using the difference-in-differences framework, future work should answer the following questions: (1) Do prioritized topics exhibit faster growth in publication/project activity than non-prioritized topics? (2) Does this effect vary by ER, ExR or SR—i.e., do capability-grounded priorities (high ER, low SR) outperform aspirational priorities (low ER, high SR)? (3) Does strategic archetype moderate impacts—do Excelling Perfectionists’ priorities succeed more than Narrow Specialists’ priorities? Answering these questions would provide empirical evidence on the mimicry debate: if high-SR priorities consistently underperform, policymakers should discourage aspirational targeting; if they succeed, the current practice may be justified.

Second, our two-dimensional framework (capability endowment POI × selection behavior SCZBI) should be extended to other domains and policy cycles. Replicating the analysis for RIS3 2014–2020 (retrospectively) would test whether mimicry intensified or diminished over time. Comparing energy to health, digitalization or advanced manufacturing would reveal whether the Green Deal pressures uniquely drive stretch behavior or whether it is systemic across sectors. Cross-national studies could test whether RIS3 governance quality (stakeholder inclusiveness, data infrastructure, institutional capacity) predicts lower SR and higher ER/ExR, isolating best practices.

Third, the mechanisms underlying mimicry warrant deeper investigation. Why do regions with high POI—possessing abundant strengths—nonetheless allocate two-thirds of priorities to stretch domains (Excelling Perfectionists SR = 64%)? Is this strategic foresight (anticipating that current strengths will become obsolete, proactively pivoting) or institutional capture (lobbying by aspirational sectors overwhelming capability evidence)? Case studies comparing high-ER and high-SR regions within the same country (controlling for national institutions) could illuminate micro-level dynamics. Similarly, understanding why Narrow Specialists exhibit the lowest SR (25%) despite lacking capabilities—seemingly the most “grounded” archetype—requires qualitative work. Our interpretation is that weak regions select opportunistically from scarce activity, but alternative explanations (e.g., smaller strategy teams unable to articulate stretch ambitions) merit exploration.

Fourth, the role of legacy (prior RIS3 experience) deserves attention. We found no significant legacy effects on selection behavior, suggesting regions do not “learn” to select more strategically over successive cycles. However, legacy may matter for implementation rather than selection: experienced regions may translate priorities into activity more effectively even if their selection behavior is equally aspirational. Testing whether legacy moderates the priority-to-outcome relationship (DiD estimates of treatment effects) would clarify whether experience compensates for wishful thinking through superior governance.

Fifth, the normative question of whether mimicry is “bad” remains open. If all regions pursue hydrogen and offshore wind, collective overinvestment may occur, but if EU funding scales to match demand, individual regions may rationally compete for shares. Modeling the equilibrium of the priority-selection game—wherein regions signal alignment with EU trends, the EU allocates funding based on those signals, and post hoc some regions succeed while others fail—would formalize the strategic dynamics. Our descriptive evidence documents widespread stretch behavior; game-theoretic analysis could explain why it persists and what policy reforms would shift equilibria toward capability-grounded selection.

4.6. Concluding Remarks

This study provides the first comprehensive empirical portrait of how EU regions select energy priorities under RIS3 2021–2027, revealing a landscape where capability endowments shape strategic orientation but mimicry and aspirational targeting dominate selection outcomes. Regions leave 93% of their inside strengths and adjacent frontiers immobilized, instead allocating 40% of priorities to distant domains disconnected from portfolios, with stretch behavior substituting for rather than complementing evidence-based choices. While the entrepreneurial discovery paradigm envisions regional stakeholders converging on capability-grounded priorities through data-driven processes, the reality is that EU policy narratives (Green Deal, hydrogen economy), competitive funding pressures and institutional mimicry produce strategies that signal ambition more than they mobilize absorption capacity.

These findings challenge policymakers to rethink smart specialization design. Without mechanisms anchoring priorities in capability assessments—auditing ER/ExR rates, requiring justifications based on relatedness density, rewarding performance over promises—RIS3 risks devolving into aspirational signaling contests where regions compete to claim the most fashionable domains regardless of fit. Differentiated instruments by strategic archetype, transparency through public benchmarking and reformed EU funding criteria that reward capability alignment could mitigate mimicry pressures. Yet we also recognize that individual regions may rationally engage in mimicry if it maximizes funding access, pointing to a collective action problem that requires coordinated policy reform at the EU level.

For scholars, our two-dimensional framework—operationalizing capability endowments (POI) and selection behavior (SCZBI) to classify regions into strategic archetypes—offers a replicable template for evaluating RIS3. The Exploitation Rate, Exploration Rate, Stretch Rate and alignment metrics provide concrete indicators to audit whether strategies “follow indicators” or “follow peers,” moving the smart specialization literature from normative prescription toward empirical evaluation. Future research should test whether capability-grounded priorities outperform aspirational ones, providing the causal evidence needed to settle the debate between evidence-based and mimicry-driven regional innovation policy.

The energy transition represents one of the defining challenges of the 21st century, demanding coordinated action across EU regions to decarbonize economies while maintaining competitiveness. Smart specialization, in principle, could enable this by directing regions toward energy domains where they possess competitive advantages and realistic prospects for global leadership. Our findings suggest that realizing this potential requires confronting uncomfortable truths: current practice prioritizes signaling over substance, aspiration over absorption and headlines over capabilities. Bridging the gap between the normative vision of evidence-based entrepreneurial discovery and the empirical reality of mimicry-driven selection is the central challenge for the next generation of smart specializations policy.