Performance-Driven Generative Design in Buildings: A Systematic Review

Abstract

Buildings are under increasing pressure to address decarbonization and climate adaptation, which is pushing design practice from post hoc performance checks to performance-driven generative design (PDGD). This review maps the current state of PDGD in buildings and proposes an engineering-oriented framework that links research methods to deployable workflows. Using a PRISMA-based systematic search, we identify 153 core studies and code them along five dimensions: design objects and scales, objectives and metrics, algorithms and tools, workflows, and data and validation. The corpus shows a strong focus on facades, envelopes, and single-building massing, dominated by energy, daylight and thermal comfort objectives, and a widespread reliance on parametric platforms connected to performance simulation software with multi-objective optimization. From this evidence we extract three typical workflow routes: parametric evolutionary multi-objective optimization, surrogate or Bayesian optimization, and data- or model-driven generation. Persistent weaknesses include fragmented metric conventions, limited cross-case or field validation, and risks to reproducibility. In response, we propose a harmonized objective–metric system, an evidence pyramid for PDGD, and a reproducibility checklist with practical guidance, which together aim to make PDGD workflows more comparable, auditable, and transferable for design practice.

1. Introduction

1.1. Background, Challenges, and Highlights

Driven jointly by the “dual-carbon” transition and climate adaptation, the paradigm of building energy saving is shifting from equipment-efficiency–led approaches to performance-driven design (PDD). In parallel, generative design (GD) has become tightly coupled with PDD, giving rise to a new paradigm of generative design driven by building performance. In this paradigm, quantifiable performance indicators, such as energy use, daylighting, and thermal comfort levels are set as explicit objectives, and a closed loop of parametric modeling—simulation-based evaluation—with intelligent optimization/generation is used to automatically explore high-performance massing, envelope, and roof [1,2,3].

A large body of research has shown that early-stage design decisions (e.g., massing, orientation, window-to-wall ratio (WWR), shading, envelope assemblies) have a pronounced influence on building energy use and comfort and are often subject to trade-offs with daylighting/ventilation targets. Consequently, multi-objective optimization (MOO) and data-driven predictive assessment need to be introduced during scheme generation, so as to avoid “after-the-fact fixes” at later stages [4,5].

Meanwhile, steady advances in the building performance simulation (BPS) toolchain (e.g., EnergyPlus, Radiance, OpenStudio, IES Virtual Environment) and in intelligent algorithms (e.g., Non-dominated Sorting Genetic Algorithm II (NSGA-II), Particle Swarm Optimization (PSO), response surfaces, Bayesian optimization, surrogate/metamodels, ML (machine learning)/deep learning (DL)) have made the generate–evaluate–optimize loop far more efficient. Designers can now batch-generate alternatives on parametric platforms, evaluate them with high-fidelity simulation or surrogate models, and converge toward Pareto-optimal solution sets using multi-objective evolutionary or hybrid intelligent frameworks [6,7,8,9]. These developments have enabled cross-scale generative optimization, from components/subsystems (shading, facades, windows, roofs) to single buildings and beyond to groups and districts [10,11,12].

Despite this rapid growth, the PDGD literature remains fragmented across objects and scales, objective sets, and algorithmic paradigms, making it difficult to reuse results in practice. Designers and engineers still lack a consolidated view of what has been tested, under which conditions, and with what level of evidence. The specific pain points are as follows:

• Inconsistent metric systems and evaluation criteria across studies—e.g., Energy Use Intensity (EUI)/kWh, spatial Daylight Autonomy (sDA)/Useful Daylight Illuminance (UDI)/Annual Sunlight Exposure (ASE), Predicted Mean Vote (PMV)/Predicted Percentage of Dissatisfied (PPD) in various combinations—hamper lateral comparison and the accumulation of evidence [13,14].
• Coupled objectives with limited sensitivity/interpretability: insufficient emphasis on design-variable-centered explanations undermines decision support [15].
• High computational cost and weak reproducibility: the community is moving toward surrogate/hybrid-intelligence frameworks, yet their generalization boundaries and data dependencies remain unclear [16,17].
• Toolchain bias and limited interoperability: many PDGD workflows are tightly coupled to specific parametric and simulation platforms, which restricts the types of problems that can be addressed, makes replication on alternative toolchains difficult, and hampers integration with BIM and urban scale models [18].

Accordingly, the main highlights of the paper are as follows:

• We provide a clear definition and scope of PDGD at the early design stage, distinguishing it from conventional performance-driven design and from operation-side optimization.
• We systematically map 153 recent studies onto a five-dimensional space—objects/scales, objectives/metrics, algorithms/tools, workflows, and data/validation—so that typical PDGD pipelines and their evidence base become visible and comparable.
• We identify three dominant workflow archetypes for PDGD and discuss where each is most suitable in terms of the performance–cost–risk trade-off.
• We synthesize a working objective–metric system and an “evidence pyramid” and translate them into a reproducibility checklist and engineering oriented guidance that can support more consistent reporting and practical deployment of PDGD in real projects.

Together, these contributions form a working framework that responds directly to the pain points outlined above. The five-dimensional mapping and workflow archetypes reduce fragmentation, the objective–metric system clarifies conventions and comparability, and the evidence pyramid with its reproducibility checklist offers practical guidance on validation and documentation for more robust PDGD workflows.

1.2. Scope and Definitions

We define PDGD as an integrated workflow applied at the design generation stage rather than in operational control, which targets energy, daylighting, thermal comfort, carbon, and economics. The workflow couples parametric geometry, performance simulation, and intelligent optimization/generation to automatically or semi-automatically produce and screen design schemes. Unlike traditional post hoc performance checking, this paradigm emphasizes objective functions that directly drive the generation and iteration of geometry and construction, typically under a multi-objective formulation that makes trade-offs explicit [5,19]. To reduce the computational burden of expensive simulations, surrogate/metamodels and Bayesian or ML predictors are widely employed to accelerate search [1,16,20]. Distinct from studies focused on equipment and operations, this review concentrates on design-side methods, objects, and metrics (e.g., facade/shading, massing, WWR, envelope, openings, and lightwells), across scales from components to single buildings and districts [10,21,22].

In practical terms, the scope of this review includes PDGD studies that

• Operate at the concept or schematic design stage and use building-side geometric and envelope parameters as primary design variables, including façade and shading systems, massing, window-to-wall ratio, envelope systems, openings, and lightwells.
• Target one or more of the five core performance objectives, namely energy, daylighting, thermal comfort, carbon, and economics, and rely on explicit and quantitative metrics.
• Implement a closed loop that couples parametric modeling, building performance simulation, and algorithmic optimization or data/model-driven generation.
• Cover spatial scales from components and facades to single buildings and, where available, districts.

1.3. Related Reviews and Distinctiveness

A substantial body of survey literature spans generative design, performance optimization, and AI in building design.

Table 1. Comparative summary of prior review.

Survey (Year)	Scope (Topics Covered)	Objects/Scale (Design Level)	Validation Focus (Gaps or Evaluation Addressed)
Ma et al. (2025) [2]	Building energy use factors; prediction and optimization	Building scale (energy consumption of buildings)	Identifies key influencing factors; highlights need for better prediction accuracy; no generative aspect
Yazdi Bahri et al. (2022) [13]	Parametric facades and thermal comfort	Facade systems (building component); residential focus	Highlights comfort improvements; lacks discussion of cross-building validation
Cataroğlu Coğul et al. (2025) [14]	Daylighting, solar energy and human factors	Building and neighborhood scales	Finds focus mostly on building-scale passive measures; calls for neighborhood-scale and experimental studies
Villano et al. (2024) [23]	ML/DL for energy sim., optimization, management	Building energy systems (HVAC, retrofits)	Compares ML vs. DL accuracy; notes ML for efficiency vs. DL for control; little on experimental validation
Al Mindeel et al. (2024) [24]	MOO of energy, comfort, IAQ	Whole building performance (mostly offices/homes)	Finds most studies tackle all 3 objectives but overlook occupant behavior; encourages AI to transform methods
Alexakis et al. (2025) [25]	GA-based MOO for building retrofitting	Existing building retrofits (energy, cost, comfort objectives)	Finds NSGA-II most used; highlights long compute times, lack of occupant preference modeling; recommends tool accessibility improvements
Li et al. (2025) [26]	Building performance MOO	Building design optimization (various types)	Stresses need for systematic up-to-date reviews; does not address generative design or workflow explicitly
Wang et al. (2023) [27]	Passive House design optimization	Single building (passive house retrofit)	Demonstrates 25% energy and 21% discomfort reduction via MOO; provides design guidelines, but scope is one case (not general framework)
Bienvenido-Huertas et al. (2023) [28]	Natural ventilation and mixed-mode in warm climates	Building performance (ventilation, comfort)	Reveals very few prior reviews; cluster analysis shows disjoint research; advocates generative design integration into NV studies
Xiang et al. (2025) [29]	Thermal comfort strategies (with ESG focus)	Educational buildings (classrooms, schools)	Links comfort to ESG metrics; likely notes need to balance comfort vs. energy; lacks generative workflow perspective
Bhote and Chauhan (2025) [30]	MOO for dynamic facades (energy efficiency)	Building facades (focus on hotels, etc.)	Identifies trade-offs (energy vs. daylight vs. comfort) needing holistic optimization; notes lack of real-time adaptability
Özlük et al. (2025) [31]	Adaptive facades optimization (AI and tools)	Facade systems (envelope adaptivity)	Notes many tools available; points out slow computation and integration challenges; no unified framework given
Lystbæk (2025) [32]	ML-driven architectural design processes	Building design process (concept to design development)	Proposes extended ML-ABD workflow; notes focus on performance and emerging generative autonomy, but challenges in adoption remain
BuHamdan et al. (2021) [33]	Generative systems in AEC (parametric and rule-based)	Building design (all stages); some construction aspects	Notes disconnect between innovations and practice; no formal validation framework
Khan et al. (2025) [34]	Generative AI for architectural design automation	Building design automation (ADA); layouts, forms	Eight ADA categories table; finds fragmented workflows, 2D bias, lack of structural logic; urges unified metrics and integration
Zhuang et al. (2025) [35]	ML in generative architectural design	Building design (forms, layouts, etc.)	Discusses opportunities and challenges (likely notes lack of standard evaluation and limited adoption in practice)
Zhang and Zhang (2025) [36]	Generative AI in design and planning	Buildings and urban planning (built environment)	Broad perspective; likely emphasizes potential for planning but notes absence of validation methods for AI-generated designs
Yan et al. (2025) [37]	Generative design for spatial layouts	Building space planning (floor plans)	Catalogs techniques; notes challenges in 3D context and functional validity (implied need for better integration)
Kookalani et al. (2024) [38]	From generative design to deep generative models	Building and structural design automation	Trajectory analysis; highlights integration of deep models with optimization as future trend; lacks detailed metric or process guidance
Wu et al. (2022) [39]	GAN applications in built environment	Multi-scale (urban data to building layouts)	Identifies 26 application domains; stresses need for datasets; no real-world testing yet
Abu-Shaikha (2025) [40]	ML in sustainable urban architecture	Urban scale and building design integration	Emphasizes need for real-world piloting and interdisciplinary data; obstacles like data quality noted
Paravantis et al. (2025) [41]	Statistical and ML approaches for energy efficiency	Building energy performance (schools, offices, etc.)	Extensive comparative tables of models; emphasizes model accuracy (MAPE, etc.) and hybrid outperformance; little on design integration

synthesizes prior reviews relative to this work, comparing their scope, scales, and validation emphases. Collectively, these surveys confirm sustained interest in performance-driven workflows, yet most concentrate on specific subdomains, scales, or techniques rather than offering an integrated, loop-centric perspective that tightly couples generation, evaluation, and optimization.

Energy- and optimization-focused syntheses consolidate drivers of energy demand and modeling practices for prediction and reduction [2], with broad mappings of ML and deep learning documenting growth in forecasting, control, and retrofit optimization [23]. Reviews of multi-objective optimization highlight tri-objective formulations (energy–comfort–Indoor Air Quality (IAQ)) in offices/residences and a have strong reliance on simulation-coupled evolutionary search, particularly NSGA-II [24], but long compute times have become a common limitation [25]. Bibliometric analyses trace collaborations and hotspots in performance optimization [26]. As a whole, these studies advance algorithmic clarity but tend to treat energy or comfort in isolation, rarely embedding generative form-finding or data-driven generation directly within optimization loops.

Single-objective reviews provide depth but rarely cross boundaries. Daylighting and solar design surveys catalog strategies and variables across building and neighborhood scales, yet most studies remain building-centric [27]; urban context and human factors are underrepresented, arguing for multi-scale approaches [14]. Bibliometric mapping of natural ventilation and mixed-mode cooling reveals few integrative reviews, methodological heterogeneity, and scarce inter-regional validation; recommendations include weaving BIM and generative design into airflow studies to expand solution spaces [28]. In thermal comfort, systematic work on educational buildings under ESG principles inventories passive/active measures and underscores the challenge of co-optimizing comfort with other goals [29].

Facade-focused surveys sit at the nexus of form and performance. Parametric studies show measurable comfort and energy gains from double-skin systems and adaptive shading while classifying geometry-performance relationships [13]. Global appraisals of facade MOO chart a shift from single-criterion tuning to balancing energy, daylight, and comfort, but also flag limited real-time adaptability and weak links to structural/operational constraints [30]. Systematic reviews of adaptive facades dissect AI algorithms, simulation engines, and software stacks, noting progress alongside computational cost and interoperability friction [31]. Despite valuable depth, these streams seldom articulate a generalizable pipeline that spans facade generation, building massing, and whole-building—or district-scale—behavior.

Relevant studies have also traced the rise in AI in architecture, engineering, construction (AEC) [32], especially deep generative methods. Early systematizations map parametric modeling and shape grammars and identify a persistent gap between computational innovation and practice [33]. Recent syntheses of diffusion, Generative Adversarial Network (GAN), and transformer models describe rapid advances in automating layouts and floorplans while emphasizing fragmented workflows, limited engineering awareness inside generative steps, and the absence of standardized evaluation criteria for architectural quality [34]. Without cross-disciplinary integration, generative AI risks accelerating drafting rather than transforming methodology. Complementary reviews show ML supporting generative form creation via surrogates and reinforcement learning [35], extend the canvas to urban planning while stressing real-world constraints [36], and classify spatial-layout generators from evolutionary solvers to GANs, noting frequent needs for post-processing to meet context and user requirements [37]. Trajectory studies link rule-based generative design to deep generative models and promote hybrid pipelines where networks propose options later optimized for performance—still rare in fully integrated form [38]. Cross-stage scans covering massing, facade detailing, and space planning reveal patchwork adoption and the lack of cohesive, multi-stage pipelines [26].

The tension between creative output and performance fidelity recurs. Reviews of GAN applications enumerate dozens of built-environment use cases but cite limited domain-specific datasets, weak generalization, and difficulty routing physics-based feedback into generative loops [39]. Large-scale mappings of ML-driven design foreground a turn toward building performance and autonomous generation while arguing that genuine intelligence requires interoperability with established design stages and deliverables [32]; obstacles like data quality have become the main limitations [40]. Across AI-oriented surveys, three bottlenecks dominate: (1) poor interoperability between design and simulation environments, (2) scalability gaps when moving from components/rooms to complex buildings or districts, and (3) uncertain generalization beyond training regimes. These help explain why many workflows still treat generation and evaluation as loosely coupled, sequential steps.

Synthesis across the three strands—performance optimization, single-objective depth, and AI-enabled generation—shows technical maturity in parts but incompleteness in the whole [41]. Prior reviews typically emphasize either the engines that evaluate, the algorithms that optimize, or the models that generate; few adopt a closed-loop perspective that treats generate–evaluate–optimize as one integrated cycle across scales. Even where MOO is prevalent, links to generative geometry are ad hoc; where deep generative models flourish, performance metrics are often appended after the fact and lack standardization. Metric and validation practices mirror these divides. Objectives and metrics tend to track domain boundaries—daylight-centric, airflow-centric, or envelope-centric—undermining cross-study comparability. Validation discussion is common but narrow: simulation accuracy and algorithm benchmarks dominate, whereas structured, hierarchical evidence that triangulates simulation, lab-scale prototyping, and in-use monitoring is uncommon. Reproducibility practices vary widely; reporting checklists tailored to performance-driven generative experiments are scarce. Scale integration is also weak. While some work acknowledges building-to-urban or design-to-operation links, no widely recognized workflow unites facade rules, whole-building behavior, and district context in a single pipeline. Occupant-centric objectives appear inconsistently despite their centrality to comfort, usability, and acceptance.

In summary, the existing survey literature is rich but remains fragmented. Reviews on energy and optimization mainly clarify algorithms and drivers of demand. Daylighting, façade, and ventilation surveys deepen understanding within single performance domains. AI and generative design surveys focus on methods and automation, while only briefly touching on performance metrics and validation. None of these works provide a critical and integrated review of performance-driven generative design that links performance objectives, generative workflows, and validation practices across design objects and scales, and they do not capture the most recent wave of hybrid workflows and evidence-based practices that has emerged since 2020. This lack of an up-to-date, loop-centered synthesis leaves a gap between method innovation and engineering-ready application, which is the gap that the present review aims to address.

Building on this diagnosis, the present review adopts a loop-centered and engineering-oriented stance. Instead of only cataloging existing studies, we propose a coherent framework that links performance objectives, generative workflows, and validation practices and that is grounded in evidence and reproducibility. Together, these elements express our view of how performance-driven generative design can move from fragmented experimentation toward reliable and engineering-ready workflows.

2. Theoretical Background

This section provides a concise theoretical background that supports the review in the later sections. It clarifies how we understand performance-driven design and generative design and it lays out the main concepts that underlie the workflows discussed later. This section is intended to give readers with different technical backgrounds a common language about design variables, performance metrics, optimization, and generative mechanisms. With this shared basis, the following sections can focus on patterns, evidence, and engineering implications instead of repeating basic definitions.

2.1. Core Ideas of PDD

The essence of PDD lies in driving the generation and iteration of geometry through quantifiable performance indicators (such as energy consumption/carbon, daylight/glare, thermal comfort, ventilation/indoor air quality, etc.) rather than relying on passive validation after the design is completed. The theoretical core of PDD can be summarized in three aspects:

• From performance-response mapping to multi-objective trade-offs.

PDD treats design as a high-dimensional mapping $f:x\to y$, where $x$ are design variables (massing, orientation, WWR, shading parameters, etc.) and $y$ are performance metrics (EUI, sDA/UDI, PMV/PPD, carbon, etc.). Because this mapping is typically nonlinear, multi-modal, and strongly coupled, PDD uses the Pareto frontier to characterize the boundary of “no simultaneous improvement” among objectives, and—supported by interpretability and sensitivity analyses—reveals dominant relationships among key variables and constraints [3,42]. At the morphology–climate coupling level, studies show that climate-driven form-finding can markedly reshape the trade-off structure among energy, daylight, and thermal comfort, offering an evidence-based path for making climate a first-class design variable [43].

2.

The simulation–surrogate computational loop and early-design first.

When high-fidelity BPS (EnergyPlus, Radiance, etc.) is coupled with multi-objective optimization, computational cost grows rapidly with dimensionality and physics complexity. PDD therefore widely adopts metamodels/surrogates (response surfaces, Kriging/GP, ensemble learning) and Bayesian optimization to replace brute-force full simulation with a small-sample, high-efficiency loop of “simulation sampling–surrogate fitting–fast optimization/uncertainty quantification” [44,45,46,47]. This paradigm has delivered efficient and robust performance for envelope heat transfer, phase change materials, and pedestrian wind comfort, and naturally extends to inverse design (inferring geometry/parameters from target fields or metrics) [17,48,49].

3.

From design–operation to lifecycle consistency.

PDD emphasizes the value density of early decisions: the earlier key variables converge, the lower the downstream change cost and rework risk. Recent work integrates model predictive control (MPC)/operational strategy simulations with uncertainty evaluation in a unified framework to keep objectives consistent from concept through operation while continuously feeding operational data back to update surrogates and design assumptions [50,51,52]. This lays the theoretical foundation for reusable workflows and the evidence pyramid discussed later.

2.2. Technical Pathways of GD

GD is the “action mechanism” of PDD, responsible for systematically generating and filtering design alternatives under given objectives and constraints. Considering task physics, data availability, and computing budgets, three reusable pathways can dominate.

• Physics-simulation-driven

On a parametric platform, schemes are generated in batches, evaluated with high-fidelity BPS, and steered toward the Pareto frontier with NSGA-II, PSO, differential evolution, simulated annealing, etc. [53,54,55,56]. This route offers strong physical consistency and interpretability, suiting problems with clear constraints and low-to-mid dimensional variables (e.g., facade/shading optimization, massing–orientation co-optimization, etc.), and it couples well with engineering constraints and codes [57,58].

2.

Rule/structure-driven

When design knowledge can be formalized (shape grammars, topological rules, construction templates), combining grammars with DoE/hierarchical search contracts the prior space and enables systematic sampling. Interpretable structural priors ensure constructability and robustness [59]. When cost/schedule/carbon are explicit constraints, they can be integrated as objectives or constraints to form a “cost–performance” co-optimization framework [60,61,62]. This pathway excels in scenarios with strong engineering constraints and high implementability, and it naturally aligns with PDD’s multi-objective trade-offs.

3.

Data/model-driven

In high-dimensional or computationally expensive scenarios, GD tends to employ SBO as its “generative engine.” This approach allows for the construction of performance predictors with quantifiable uncertainty from small sample sets and facilitates rapid convergence to the Pareto front via acquisition functions such as Expected Improvement [45]. Furthermore, generative DL—including Generative Adversarial Networks (GANs), diffusion models, and conditional generation—along with RL, are enabling “direct solution generation” [63] and “policy learning” [58]. In information-dense tasks involving plans, morphology, and facades, these methods demonstrate near-real-time feedback and the potential for end-to-end workflows [64]. The synergy between these advanced techniques and traditional ML is poised to drive the evolution of the design process from “human–computer symbiosis” to “partial autonomy” [65].

3. Materials and Methods

A structured literature review was conducted using a rigorous, systematic approach [66]. The process comprised five sequential stages. (1) The research objectives were refined and specific research questions formulated. (2) Comprehensive Boolean search strings were employed across major databases to identify potentially relevant records. (3) These records were screened and selected using the PRISMA framework, with irrelevant and duplicate studies excluded. (4) Quantitative bibliometric analyses were carried out to map the research landscape, identify influential works, and examine collaboration networks and temporal trends. (5) The included articles underwent qualitative thematic analysis to synthesize evidence and derive key insights. Together, these steps ensured a rigorous, transparent, and reproducible overview of the investigated research domain [67]. Figure 1 presents this five-step workflow in the form of a flow chart.

3.1. Search of Publications

We used PRISMA as the overarching protocol for identifying and screening studies on PDGD [68]. To ensure consistency, topic-specific keywords were devised to minimize the risk of missing relevant literature. Because search engines differ slightly, Boolean operators were adapted for Scopus and WOS. The search strings combined the two domains—building performance and generative design—as summarized in

Table 2. Search keywords for PDGD.

Search Fields	Search String
PDD	“Building Performance” OR “Energy Use Intensity” OR “EUI” OR “Daylighting” OR “Natural Lighting” OR “Thermal Comfort” OR “Carbon Footprint” OR “Performance Simulation” OR “Energy Simulation” OR “Multi-objective Optimization” OR “Building Performance Simulation” OR “EnergyPlus” OR “Radiance”
GD	“Generative Design” OR “Algorithmic Design” OR “Parametric Design” OR “Evolutionary Algorithm*” OR “Genetic Algorithm” OR “NSGA-II” OR “Particle Swarm Optimization” OR “Generative Adversarial Network” OR “GAN” OR “Reinforcement Learning” OR “Deep Learning” OR “Form-finding” OR “Design Grammar”

.

3.2. Paper Filtering and Selection

Deduplication and initial screening. The results from the two databases were merged into an offline table, and deduplication was performed using titles, Digital Object Identifiers (DOIs), and author information, resulting in 1671 publications. We excluded non-English literature and review articles, leaving 1207; then we removed publications published before 2020, leaving 974. Next, through manual inspection of the relevance of titles, author keywords, and abstracts, studies unrelated to PDGD were excluded (such as those in the mechanical/medical/chemical fields), leaving 792.

Finally, a full-text evaluation was conducted based on the specific content of the studies. The included publications needed to research at least one of the following aspects to meet the requirements of GD: building envelopes, internal building spaces, or small-scale blocks, etc. At this stage, publications that focused primarily on using optimization algorithms to optimize energy consumption of building equipment, such as air conditioners, without addressing the aforementioned aspects were excluded, and 153 papers were ultimately included in the study, as shown in Figure 2.

3.3. Analysis of Author Keywords

The analysis of author keywords in the context of PDGD reveals key themes and trends that are pivotal to understanding the current state and future potential of this research area. A bibliometric and qualitative review of the selected 153 papers, as shown in Figure 3, demonstrates that the keywords most frequently associated with the subject include “multi-objective optimization,” “energy efficiency,” and “thermal comfort.” These keywords align with the core focus of the research, which involves optimizing building designs for energy efficiency, thermal comfort, and other performance criteria through computational and algorithmic methods.

The GD process, often coupled with parametric and algorithmic design methods, is central to addressing the multifaceted challenges of building performance. Key terms like “energy simulation,” “daylighting,” and “carbon footprint” highlight the emphasis on using simulation tools such as EnergyPlus, Radiance, and other software platforms that support the optimization of building envelopes, systems, and layouts. These tools are integral in achieving the goals of energy efficiency, sustainability, and occupant comfort, which are essential components of the broader objective of reducing buildings’ environmental impact.

The growth of this field, as evidenced by the increasing number of publications from 2020 onward, reflects a shift towards more integrated and efficient design processes, where performance-driven design is paired with advanced computational tools. This evolution is evident not only in single-building scale studies but also in larger-scale investigations that consider the impact of urban and district-level design interventions.

3.4. Bibliometric Analysis

To track field evolution and journal/topic distribution, we performed a bibliometric analysis on the 153 included papers. The output shows a steep rise after 2019, approaching 60 publications in 2025. Articles are distributed across 64 journals, with the top five being Journal of Building Engineering (n = 15), Energy and Buildings (n = 18), Buildings (n = 15), Sustainability (n = 11), and Building and Environment (n = 9). Visual summaries are provided in Figure 4 and Figure 5.

4. Results

To enhance reproducibility and strengthen evidence anchoring, we conducted cross-labeling on the PRISMA-screened corpus (n = 153), coding each study along five dimensions—objects/scales, objectives/metrics, algorithms/tools, workflows, and data/validation—with multi-label assignments permitted. The specific sources of these data can be referred to in the Supplementary Materials, where all publications are organized into four different tables. Supplementary Materials A provides an overview of the studies; specifically, for each paper, it includes an identifier (ID), title, year of publication, building usage type, case study type, case study location, and construction year. Supplementary Materials B summarizes the design object and scale of each study, categorizing them into the following groups: single-building massing/layout, envelope/facade, roofs/BIPV, room/floor plan, district/urban. Supplementary Materials C summarizes the objectives and metrics in each study, dividing them into the following groups: daylight, energy, thermal comfort, carbon, economics. Supplementary Materials D presents the workflows involved in each study, as well as the algorithms and tools used, to meet the requirements for reproducibility.

4.1. Overview of the Results

Supplementary Materials A contains a detailed table, which is included here to accommodate the table’s length and improve the document’s readability. This table presents key information for each study: an ID, title, publication year; building use type: Residential (R), Public (P) (including Education, Social, and Healthcare), Commercial (C) (such as Office and Retail), Industrial and Warehouse (I), or Other (O) (including Traffic and Assembly, Heritage, Traditional, or Listed Building) (H/T/LB); case study type: Real Building (RB), Archetype Building (AB), or Simplified Building Model (SBM); case study location; and construction year, as illustrated in Figure 6, which details the building use types covered in this review.

The reviewed studies predominantly address commercial and residential buildings, with a marked focus on offices, retail spaces, and typical housing. This emphasis is logical given that these sectors constitute the largest share of building stock and energy use, making them high-impact targets for performance improvement. For example, office buildings are frequently chosen for optimization case studies [69], and residential dwellings like single-family homes or apartments are similarly common [70]. In contrast, specialized building types—heritage landmarks, industrial facilities, schools, etc.—appear only sporadically in the corpus. Research on heritage buildings is especially scarce; one notable example optimized skylight design in a historic palace, illustrating that such cases require delicate retrofitting strategies [71]. Overall, the paucity of studies on historic, industrial, or other niche buildings underscores a gap in generalizing PDGD insights to those contexts. Future work could broaden applicability by extending methods to these underrepresented building categories, which often have unique constraints and conservation considerations.

In terms of case study type, the reviewed literature demonstrates a balanced application of different case study models. Figure 7 presents the frequency of use for Real Buildings (RBs), Archetype Buildings (ABs), and Simplified Building Models (SBMs).

The analysis shows that ABs are the most frequently used case study type, appearing in 68 studies. The widespread use of archetypes, which are representative models of a specific building type, age, and location, highlights an emphasis on generating scalable and generalizable findings that can inform regional or national energy policies, reflecting its core value in guiding macro-level building-related strategies.

Following closely, RBs were utilized in 60 studies. This indicates a strong preference for research grounded in empirical data and real-world conditions, which is crucial for validating theoretical models and assessing the practical effectiveness of new technologies or strategies, as real-world cases can best bridge the gap between academic theories and on-site applications.

Lastly, SBM were employed in 25 studies. These models are essential for fundamental research, allowing for the isolation of specific physical phenomena or the execution of extensive parametric analyses that would be computationally prohibitive with more complex models—their application, though less frequent, plays an irreplaceable role in deepening the understanding of building physical mechanisms.

From an applied perspective, this profile of building types and case study models has several implications. First, the dominance of commercial and residential buildings is consistent with their share of floor area and energy use in many regions, but it also means that healthcare, educational, and industrial facilities remain underrepresented. Second, the frequent reliance on archetype models highlights the need for carefully documented assumptions on schedules, systems, and envelope assemblies, since these models are often reused across studies. Third, the substantial but still limited presence of real buildings suggests that validation under actual operation should be expanded so that simulated performance gains can be compared more systematically with measured outcomes. Finally, simplified benchmark models play a valuable role in isolating physical mechanisms, but they should be explicitly connected back to realistic typologies so that their insights are not confined to artificial settings.

4.2. Object and Scale

Based on the cross-annotation of 153 studies, the distribution of the design objects identified in this study across the five levels is significantly uneven. The specific data can be found in Supplementary Materials B’s Supplementary Data. Envelope/facade dominates (128), followed by single-building massing/Layout (89), roofs/BIPV (40), room/floor plan (39), and district/urban (20). Within each scale (multi-labeling allowed; hence subtype sums may exceed group totals), facade-oriented studies are led by window-to-wall ratio (WWR = 90), glazing/window-opening proportion (GWOP = 93), and shading geometry (SG = 68). At the single-building scale, the most frequent variables are orientation (Or = 58) and form/massing (FM = 52), complemented by openings organization (Op = 29), Volume Ratio/Compactness (V/C = 11), and courtyard/atrium features (C/A = 8). Roof strategies prioritize roof profile (RP = 30) and slope/tilt (S/T = 10), while BIPV (3) and skylights (Sk = 4) appear less often. At the plan scale, emphasis is on space organization (SO = 26) and ventilation paths/opening arrangement (VP = 14; OA = 11). At the district scale, attention concentrates on development intensity/density (Dn = 19) and urban form (UF = 17), with street-canyon morphology (SC = 8) as a secondary theme. This distribution corroborates our earlier finding that research efforts cluster around single-building massing/layout and facade systems, and it is consistent with the trends shown in Figure 8.

The distribution of design objects across scales shows a clear emphasis on envelope–geometry–opening variables, predominantly at the envelope/facade. Their prevalence is explained by two properties. First, these objects exhibit high interpretability and strong physical coupling: parameters such as WWR, glazing/opening proportion, and shading geometry directly govern solar heat gains, visible light transmittance, and the diurnal radiation profile, which in turn shape annual energy use, daylight thresholds, and thermal comfort through well-understood mechanisms. Second, they offer moderate dimensionality and well-bounded constraints (e.g., orientation, massing, roof slope), which yield stable sensitivity structures: a small set of high-elasticity levers (orientation, WWR, Shading Geometry) coexist with lower-elasticity controls (e.g., double-skin facades being less frequent in the corpus), together defining a tractable and trade-off-rich performance space.

The multi-label patterns further indicate pervasive interaction effects among subtypes. Typical triads, such as “WWR–shading geometry–operability openings”, set the attainable frontier under peak-load and glare constraints [72], while “massing–orientation–roof profile” couples wind-driven exchange with short-wave solar gains and shifts the inflection points of energy–comfort trade-offs across climates [73]. Hence, the cross-object interactions, rather than single parameters alone, often determine the reachable region of the Pareto set and explain the strong concentration around a few core subtypes in the evidence base.

Differences by scale clarify where data requirements and uncertainties originate:

• At the building and facade scales, boundary conditions (envelope properties, set-points, typical meteorological years) can be standardized, and measurement conventions (e.g., sDA/ASE, UDI, annual energy) are widely reused, supporting comparability and transferability.
• The relatively lower attention to Roofs/BIPV reflects their stronger context-dependence on shading, structural capacity, and maintenance regimes;
• Room/floor-plan problems introduce topological and occupancy-driven discrete variables, producing threshold-like, non-smooth responses;
• At the district scale, the small sample size mirrors a high reliance on exogenous drivers (micro-climate, ventilation corridors, street-canyon effects), making conclusions more sensitive to input uncertainty.

From our perspective, this uneven pattern suggests that PDGD has developed along the path of least resistance in terms of modeling effort and data, and that more attention is needed at the plan and district scales, where design decisions have strong social and climatic implications but are currently supported by a thinner evidence base.

4.2.1. Single-Building Massing/Layout

At the single-building massing/layout, emphasis is placed on orientation and form modifications, complemented by openings, volume ratio/compactness, and courtyard/atrium features. Their prominence arises from the direct modulation of diurnal solar exposure and incidence angles, which governs heat gains, glare risk, and the potential for wind- or buoyancy-driven ventilation; it also reflects well-bounded geometric freedoms that yield a stable structure of a few high-elasticity levers with compensatory controls. Evidence indicates that principal orientation and dominant massing dimensions largely shape the energy–daylight trade-off curve, while introducing atria/lightwells can simultaneously enhance daylight and temper thermal stress in low-latitude, high-irradiance contexts, provided that shading, smoke exhaust, and natural ventilation paths are made explicit [74,75,76,77].

With increased massing complexity, self-shading and windward/leeward shifts depress effective insolation and alter local convective coefficients; designers therefore compensate through openings organization and volume ratio/compactness to restore daylight uniformity and pressure pathways. Conversely, simplified or high–aspect-ratio forms rely more heavily on openings placement and ventilation paths to re-establish indoor airflow patterns [78,79,80]. Because boundary conditions at this scale (envelope properties, set-points, typical meteorological years) can be standardized, the associated measurement conventions (EUI, sDA/UDI/ASE, overheating hours) are widely reusable, supporting comparability across studies; nonetheless, results remain climate-sensitive—in high-latitude, cloudy regions, the marginal daylight benefits of orientation and WWR increase, whereas in high-irradiance/urban-heat-island settings, threshold effects of self-shading and sun-control intensify, enforcing a balance among compactness, ventilation potential, and peak-load constraints [20,81].

Overall, the prominence of massing/layout is rooted in a stable structure of few dominant variables with compensatory controls, which facilitates knowledge transfer and benchmarking across projects.

From a design strategy perspective, these findings indicate that early control of orientation, compactness, and courtyard or atrium configuration is one of the most powerful levers for PDGD at the single building scale. Massing-level decisions define the baseline for solar access, self-shading, and cross ventilation, and later envelope or system level optimizations can only refine what has already been set by the overall form. We therefore recommend that PDGD workflows explicitly stage massing optimization as a first phase and treat facade and system optimization as subsequent refinement, rather than trying to compensate for weak massing with detailed controls.

4.2.2. Envelope/Facade

The envelope/facade group shows the highest concentration of evidence, led by glass/walls optical properties, WWR, and shading geometry, followed by operable openings and airtightness/infiltration. These variables act directly on the incident spectrum–surface energy balance–indoor light distribution chain, affecting both annual energy fluxes and hourly glare/overheating thresholds.

Across studies, increasing glass/walls optical properties and WWR improves daylight availability but raises cooling demand and glare probability; effective shading geometry, through projection depth, porosity, and orientation-dependent masking, reshapes summer peaks and mitigates low-solar-altitude glare, thereby expanding the attainable performance frontier [82,83]. At the same time, operability and insulation govern nighttime/shoulder-season heat release and ventilation routes, setting the balance between peak-load reduction and annual payback [84]; in overheating-sensitive housing and educational buildings, in-use operation of openings and shading can markedly shift the distribution of overheating hours [85,86].

Multi-label co-occurrence indicates that glass/walls optical properties and WWR commonly pair with shading geometry, implying that interactions, rather than single parameters, set the ceiling: in high-irradiance climates, transmittance gains quickly meet cooling and glare constraints without effective shading [87], whereas in winter-dominated/high-latitude contexts, moderate WWR increases can enhance daylight and passive gains without breaching glare limits [77,81]. Finally, constructability and operational behavior often mediate the gap between models and realized performance; reporting geometric degrees of freedom, material/boundary assumptions, and operational hypotheses, together with elasticity ranges and triggering boundaries for key interactions, strengthens comparability and transferability across projects.

Taken together, the envelope and facade evidence suggests that PDGD should treat glazing properties, shading geometry, and opening control as a coupled design package rather than as isolated variables. From our perspective, the most transferable results arise when studies report not only performance gains but also the specific combinations of projection depth, orientation, control logic, and material properties that deliver robust improvements across climates. Future work should therefore emphasize families of facade strategies that can be parameterized and reused and should document comfort and visual quality outcomes alongside energy metrics so that design teams can judge tradeoffs more transparently.

4.2.3. Roofs/BIPV

Research on roofs/BIPV concentrates on roof properties and slope/tilt, occasionally extending to skylights and BIPV integration. Compared with facades, roof variables couple more directly with full-sky irradiance, obstruction factors, and seasonal solar altitude, thereby shaping context-dependent trade-offs among peak-load mitigation, passive gains, annual BIPV yield, and glare risk. Evidence shows that in high-irradiance or lightly obstructed settings, well-chosen tilt and azimuth can raise effective BIPV output while tempering summer peaks; yet the benefits are highly sensitive to local shading, roof albedo, and maintainability [9,88,89]. Where urban morphology induces stronger shading–reflection interactions, the attainable BIPV window narrows; studies therefore emphasize controlling self-shading via roof geometry and assessing coupling with adjacent masses to avoid gaps between nameplate capacity and realized yield [43,90,91,92].

For skylights, multiple studies report a dual role in daylight delivery and buoyant/stack-driven exhaust: with careful control of aperture area, optical properties, and shading construction, skylights can substantially improve daylight in deep-plan spaces and relieve overheating via night purging; however, low-altitude glare and summer solar gains can quickly flip the balance unless sun-control is effective, requiring precise coordination of daylighting targets, shading design, and roof thermal bounds [19,93].

In short, the relatively small but growing set of roofs and BIPV studies points to an important frontier for PDGD. Roof variables affect both peak loads and on-site generation potential, and their interactions with facade design are still underexplored. Practical guidance will be strengthened if future PDGD work on roofs and BIPV consistently reports solar resource assumptions, system efficiencies, and maintenance constraints and if it compares alternative roof and facade packages under common climate and load scenarios. This would allow designers to treat roof and facade as a joint design space rather than as separate add-ons.

4.2.4. Room/Floor Plan

At the room/floor plan scale, space organization, open area, and ventilation paths act jointly on daylight distribution, airflow patterns, and peak loads. Relative to massing/facade controls, this scale adds topological and discrete decisions (e.g., corridor–room connectivity, placement of shared/service spaces, number and height of operable openings), which yield threshold-like, non-smooth responses: when combinations of opening networks, spatial depth, and orientation cross certain critical conditions, daylight availability and overheating hours often shift abruptly rather than gradually [94]. Consistent findings across typologies show that in classrooms and offices with high occupant density, opening orientation and sill/head height largely determine pressure pathways and local velocities for natural ventilation [86]; in deeper plans, coupling atria/lightwells with corridor daylighting can raise sDA/UDI and reduce uniformity risks, bounded by glare thresholds and noise/smoke management requirements [95].

Because occupancy behavior and operation of openings/shading strongly mediate outcomes at this scale, studies should, beyond fixing weather files and material properties, specify occupancy schedules and internal gains alongside operational assumptions for openings, and where feasible, report elasticity under alternative occupancy scenarios to avoid over-extrapolation [21,76,96].

From a user-centered perspective, the room and floor plan scale highlights the need to integrate behavioral uncertainty more explicitly into PDGD. Space topology, door and window placement, and internal buffer zones influence not only energy and comfort but also privacy, circulation, and use patterns. We interpret the current evidence as a call for multi-scenario plan optimization, where occupancy schedules, window operation habits, and internal partition options are varied within realistic bounds. Reporting results across several plausible behavior scenarios will make plan-level PDGD more credible for practitioners and will reduce the risk of overfitting designs to a single assumed pattern of use.

4.2.5. District/Urban

Although fewer in number, district/urban studies consistently show that development intensity/density, urban form, and street-canyon geometry jointly condition block-level energy use, outdoor thermal comfort, and BIPV potential through their effects on sky-view factor, ventilation corridors, and radiative exchanges [97]. Consequently, identical single-building strategies can yield divergent energy–comfort trade-offs under different urban contexts [75,92,98].

In high-density, pronounced heat-island settings, surface/facade albedo, vegetation, and water bodies markedly influence mean radiant temperature and nocturnal heat release [99]; when aspect ratios and prevailing-wind alignment are unfavorable, local cooling demand and overheating hours increase substantially [100,101]. The spatio-temporal availability for BIPV is likewise constrained by block morphology: under obstruction-dominated conditions, roof/facade strategies at the building scale must be coordinated with block-level massing–spacing–orientation to achieve both peak-load relief and generation yield [91].

Accordingly, studies at this scale should report urban-form metrics (e.g., density, mean/max heights, floor-area ratios, canyon openness), meteorological context, and surface/facade material parameters, and where possible, provide scenario elasticity across density and openness to avoid boundary-free extrapolation from single-building evidence. Overall, the District/Urban scale foregrounds a morphology–microclimate–load/generation triad, with the primary challenge being transparent characterization of exogenous drivers and obstruction/ventilation pathways.

Taken together, the object and scale analysis shows that the current PDGD evidence base is strongest at the single-building massing and facade scales, where boundary conditions are well controlled and reusable, and where trade-offs between energy, daylight, and comfort are clearly exposed. Roof and BIPV applications and especially district or urban studies are still emerging and rely on smaller samples and heavier assumptions about external drivers. From our perspective, results at these less mature scales should be read as directional hypotheses rather than ready-made design rules, and they call for stronger reporting of context, uncertainty, and validation. We also see an opportunity for cross-scale workflows that reuse knowledge from the more established building and facade scales to inform district and urban interventions instead of treating each scale in isolation.

4.3. Objectives and Metrics System

In PDGD, the choice of objectives and metrics defines the performance space, the simulation setup, and the way design alternatives are judged. We organize the objectives into five core families based on recent studies, namely energy, daylighting, thermal comfort, carbon, and economics, and examine how the associated metrics are used, combined, and reported across the corpus [5,21,102].

Based on Figure 9, the use of objectives/metrics exhibits a clear hierarchy: energy is most prevalent (135), including EUI (116), and cooling/heating loads both (41); daylight (52), including UDI (41), sDA (18), and ASE (8); thermal comfort is comparable (52), with PMV (29), PPD (27), and overheating (9); carbon (27); and economics (22) including Net Present Value (NPV = 16), and payback (9). This distribution aligns with recent emphases on energy and loads, climate-based daylight metrics, and the co-optimization of comfort–carbon–cost. The specific data can be found in Supplementary Materials C’s Supplementary Data.

For energy, EUI is the dominant integrative efficiency indicator, complemented by cooling/heating load to capture peak and seasonal responses under climate–envelope–system coupling—consistent with trends in ML-augmented MOO, interpretable modeling, and load decomposition.

For daylight, the combination of UDI, sDA, and ASE has become mainstream: UDI reflects annual illuminance usability, while sDA/ASE embody climate-based adequacy and overexposure criteria, enabling joint optimization of energy–glare–visual quality.

For thermal comfort, PMV/PPD remain core metrics, with overheating indicators added to represent heat-stress risks; comfort is often co-optimized with energy as a key trade-off dimension.

For carbon and economics, carbon emissions objectives are paired with NPV/payback to evaluate decarbonization feasibility and investment efficiency, supporting life-cycle performance–cost decisions.

4.3.1. Energy

Within the energy objective system, we adopt a two-tier structure of EUI and cooling/heating loads: EUI indicates integrated annual energy intensity per unit floor area [103], while cooling/heating loads capture peak and seasonal demand under climate–envelope–system coupling [104].

From the synthesis in Figure 9, EUI is the most frequently used indicator, and cooling/heating loads complement it by informing capacity sizing and operational risk, thereby enabling measurable trade-offs between “overall efficiency” and “peak safety margins [105].” In cross-objective coupling, energy objectives are routinely considered alongside thermal comfort, daylighting, carbon, and economics: (1) energy and comfort are constrained by set-points, airtightness, and ventilation [74]; (2) daylight strategies bidirectionally affect lighting demand and solar gains [106]; and (3) energy relates to life-cycle carbon and NPV/payback, which bound decision spaces [107].

At design objects levels, shape factor, orientation, WWR, U-value/SHGC of transparent elements, and shading geometry exert high sensitivities on EUI–peak loads, producing seasonal trade-offs of heat gains and diurnal/seasonal peak shifts, hence the need to co-measure annual energy with extreme-hour loads [85,108].

Accordingly, the core aim of the energy subsection is to minimize EUI while bounding peak loads under indoor-environment and multi-performance constraints, ensuring both annual efficiency and capacity adequacy.

4.3.2. Daylighting

Daylighting adopts a Climate-Based Daylight Modeling (CBDM) triad—maximize UDI/sDA and minimize ASE [109]—capturing usability, adequacy, and overexposure while naturally coupling with energy and visual-comfort boundaries [110].

At design objects levels, WWR, aperture position/orientation, shading geometry (fixed/active), and skylight/light-well forms provide elastic controls over UDI–sDA–ASE: vertical apertures or top lighting increase UDI/sDA but may inflate ASE without adequate shading, hence coordinated upper/lower bounds maintain adequacy without overexposure [88,111,112].

At room-to-building scales, plan depth, floor-to-floor height, and atrium/void configurations shift the side-light vs. top-light balance, reshaping the spatial distribution of UDI–sDA and the peak locations of ASE, which necessitates clear declarations of scale and dominant aperture types in the objective statement [113,114].

Coupling with EUI/loads is physically consistent: greater useful daylight reduces lighting energy and can aid winter gains, yet increases summer cooling and glare risks; standard layering maximizes UDI/sDA while bounding ASE and jointly balances seasonal reversals with EUI/loads [80].

For specific programs (e.g., education and offices), acceptable bands for UDI/sDA and upper limits for ASE are tuned to occupancy schedules, workstation density, and task visual demands to ensure scenario-fit under comparable metric conventions [56,115,116].

4.3.3. Thermal Comfort

We structure thermal comfort objectives as a triad of steady-state comfort, overheating risk, and temporal consistency, with PMV/PPD serving as the primary vocabulary for comparable assessment across spaces and seasons [117]. To bound tail events, overheating hours/degree-hours (e.g., >26 °C or >28 °C) are adopted as risk metrics that complement annual comfort goals [118].

Program-specific fit matters: classrooms vs. offices differ in bandwidth tolerance and continuity, with education space more sensitive to peaks and scheduling [86,119].

Window operation, shading, and envelope features shape radiant asymmetry, vertical stratification, and draft, so the target layer should combine PMV/PPD, overheating hours, and bandwidths to avoid compensatory risks from single metrics [120,121].

Comfort and energy interact seasonally: tightening comfort bands reduces dissatisfaction yet often raises energy use, whereas moderate widening aids savings subject to overheating caps and protection of sensitive groups [122,123].

At population scales, thresholds and temporal priorities should reflect high-density occupancy vs. low-metabolic tasks so that objectives remain attainable, maintainable, and interpretable [78].

4.3.4. Carbon

Within the early-stage context of PDGD, notably, when carbon is considered, it is almost always operational carbon only—i.e., emissions due to building energy use. Early-stage generative studies exclude embodied carbon of materials because precise material choices and quantities are not yet defined at the concept design stage, avoiding uncertainties that would come from a full LCA. This approach is recommended by sustainability scholars [124], who argue that attempting a full carbon life-cycle analysis in early design can mislead comparisons.

Consequently, most early-stage work frames the boundary around energy mix, system efficiency, and envelope performance, which are actionable and decision-sensitive before bills of materials are fixed [74]. Under this convention, operational carbon is expressed as annual intensity (kgCO₂e/m²·yr) or annual totals, forming a one-to-one mapping with energy indicators (EUI/loads) at the yearly scale for trade-off analysis [125,126,127].

For programs with high occupancy (e.g., education, offices), objectives should reference occupancy schedules and equipment timetables to distinguish winter benefits vs. summer penalties, improving seasonal consistency in design evaluation [128]. When projects require explicit renewal horizons, it is useful to annotate lifetime milestones (e.g., 5/10/20-year cumulative emissions) while retaining the operational-only accounting, thus avoiding uncertainty from the materials side [129].

Alignment with policy and economics arises when carbon price/allowances or RECs enter the boundary, providing consistent price signals so energy savings and emission reductions register coherently in NPV/payback assessments [130].

4.3.5. Economics

In our corpus, economic objectives are led by NPV and complemented by payback, forming a dual-indicator pattern of life-cycle value and capital recovery speed that aligns with annual-scale performance trade-offs.

NPV (maximize) quantifies the net benefit of early design alternatives under a unified horizon and discount rate, drawing inputs from simulation-derived operating costs and design-dependent incremental capital, and serves as an economy-wide objective aligned with energy intensity in the generative search space [107]. Payback (minimize) supplies an intuitive threshold for capital recovery and is commonly declared with NPV to avoid pursuing short paybacks at the expense of life-cycle value; together they bound the cost–benefit envelope of performance improvements [131].

For comparability, statements should fix analysis horizon and discount rate, merging incremental capital expenditure (CAPEX) (from envelope/system choices) with operational cost deltas driven by performance within a single cash-flow frame, thereby creating a one-to-one mapping to energy and (operational) carbon objectives [132,133]. For education/public programs with high occupancy and budget rigidity, studies commonly adopt “NPV-led ranking + Payback cap” as the basic generative objective layer, i.e., NPV as the ordering spine with an acceptable payback upper bound for financial executability [130].

At the decision tier, we recommend NPV as the primary objective with payback as a complementary bound: the former captures long-term value and path dependence, while the latter manages capital-recovery and liquidity risk—together ensuring actionability within a four-dimensional frame of annual intensity, peak capacity, carbon, and cost.

4.3.6. Variations in Objectives Across Building Typologies

Figure 10 summarizes the performance deltas (Δ) of twelve commonly used indicators when optimization is applied to different building typologies. For each metric–typology cell we computed the median Δ, interquartile range (IQR), and n. The specific data can be found in Supplementary Materials E’s Supplementary Data.

• Daylight performance:

Across typologies, optimization consistently improves daylight performance. UDI tends to increase in all building types, with the largest typical gains in commercial buildings and smaller but still positive gains in public, other, and residential stocks. sDA rises in the limited number of cases where it is reported. ASE generally decreases, which indicates that glare and overexposure risks are reduced alongside higher daylight availability.

2.

Energy:

EUI shows broad reductions across residential, commercial, public and other buildings. Typical EUI decreases cluster around 20–30 percent for residential, commercial, and other buildings, with slightly lower but still meaningful reductions in public and special types. Cooling and heating energy or load also fall substantially in most typologies, and public buildings often exhibit the largest relative cooling and heating reductions because of their high internal gains and long occupancy schedules.

3.

Thermal comfort:

Thermal comfort indicators show a consistent trend toward reduced discomfort. The PPD generally decreases across typologies, and the fraction of time with overheating also tends to decline, which is compatible with the observed reductions in cooling loads. PMV values move closer to the neutral range in most cases, confirming that the optimization schemes do not only reduce energy demand but also improve comfort conditions.

4.

Carbon and economics:

Carbon emission reductions broadly track the energy results, with the largest typical savings in residential and other buildings and smaller but still positive reductions in public and commercial stocks. Economic indicators show more variability: several studies report shortened payback periods that are consistent with energy savings, while NPV results are heterogeneous because they depend strongly on local tariffs, discount rates, and assumed lifetimes. This variability suggests that energy and carbon gains are relatively portable across contexts, whereas economic metrics should be interpreted with greater caution.

To summarize, these patterns suggest that PDGD would benefit from a more standardized and transparent objective system. A pragmatic structure is to treat EUI and peak loads as the default energy pair; to rely on climate-based daylight metrics that combine useful daylight illuminance, spatial daylight autonomy, and annual sunlight exposure; to couple annual comfort indicators with explicit overheating or risk metrics; and to frame carbon and economics within consistent yearly boundaries and financial assumptions. We recommend that future studies clearly state these choices, fix analysis horizons and discount rates, and report both performance deltas and absolute levels. This would make results more comparable across typologies and climates and would also help practitioners interpret whether reported gains are meaningful for their own projects.

4.4. Algorithms and Tools

4.4.1. Main Algorithms and Tool Clusters

Based on Figure 11 (the specific data can be found in Supplementary Materials D Supplementary data), the field shows a concentrated head with a diverse long tail. Evolutionary MOO drives the core, with NSGA-II alone accounting for 26.7% of mentions and being directly coupled to high-fidelity energy/daylighting simulations to construct Pareto fronts [134,135]. GA and GA-BP are frequently used for parameter search and surrogate training [110,136]. The coverage curve (Figure 11c) indicates that the top 20 algorithms explain roughly 75% of all mentions, highlighting practitioners’ preference for robust, tool-ready solvers. Overall, contemporary practice is anchored in direct multi-objective search, with surrogates and data post-processing accelerating evaluation, compressing the search space, and providing mechanistic interpretation.

As shown in Figure 12 (the specific data can be found in Supplementary Materials E’s Supplementary data), frequency mirrors the concentration of algorithm. Rhino/Grasshopper functions as the parametric workbench (~21.9%), typically paired with EnergyPlus (~15.5%) and Ladybug/Honeybee (~15.5%) to form the predominant front-end modeling–mid-stream evaluation–visualization triad. The coverage curve (Figure 12c) indicates that the top 8–10 tools account for ≈80–85% of usage, implying de facto standardization. The co-occurrence heatmap (Figure 12d) reveals stable cliques: Grasshopper, Ladybug/Honeybee, EnergyPlus for single-loop optimization at building/early-stage scales; EnergyPlus, OpenStudio and Radiance, Daysim for batch computation and model checking; Python/MATLAB act as the glue layer for job orchestration, hyper-parameter search, and multi-API integration.

4.4.2. Algorithm–Tool Ecosystem: Taxonomy and Roles

The ecosystem can be mapped bidirectionally between solver–evaluator–decision layers and workbench/engine–bridging/management–scripting layers. The solver layer centers on evolutionary MOO (NSGA-II/III, SPEA2, PSO and variants) [137], augmented by Bayesian Optimization or hybrid evolutionary search to explore global (or near-global) trade-offs [44]. The evaluator layer comprises physics-based engines for energy and daylighting (EnergyPlus, Radiance/Daysim) together with their surrogate counterparts—GPR [138], tree ensembles (RF/XGBoost/LightGBM/CatBoost) [5], kernel methods (SVM/SVR/LSSVM/GLSSVM) [139,140], and deep networks (MLP/CNN/LSTM) [141]—which approximate expensive simulations and support evaluate–guide–resample loops. The decision layer provides clustering and dimensionality reduction for structural synthesis, Multiple Criteria Decision Making (MCDM) (AHP/ANP/TOPSIS, etc.) for multi-criteria ranking [142], and XAI (eXplainable Artificial Intelligence)/sensitivity (SHAP/LIME/Morris) for mechanism-level interpretation and robustness auditing [143].

On the tooling side, Grasshopper orchestrates geometry and process; Honeybee/Ladybug bridge weather, geometry, and engines; EnergyPlus and Radiance/Daysim deliver thermal/energy and optical/daylighting evaluation [144]; OpenStudio supports model management and batch runs [145]; and Revit/Dynamo and DesignBuilder assist BIM-to-simulation mapping and checking [18,146]. Meanwhile, parametric tools for form generation based on performance optimization, such as EvoMass, have also emerged [147]. This taxonomy affords replaceable, extensible compositions across objects and scales (Figure 8) and the objective–metric relationship (Figure 9).

In practice, PDGD has coalesced around several “de facto” algorithm–tool pipelines. First, Rhino/Grasshopper serves as the geometric parameterization front end, linked to Ladybug/Honeybee for environmental boundary conditions and solver settings, while EnergyPlus and Radiance/Daysim perform coupled energy–daylight/glare simulations; multi-objective evolutionary optimization is then carried out with plugins such as Octopus and Wallacei or Galapagos for single-objective scenarios to yield Pareto fronts and design candidates [71,148,149,150]. Second, DesignBuilder/OpenStudio functions as an integrated modeling–simulation platform, combined with built-in or external GA/NSGA-II workflows for early-stage co-optimization of energy use, thermal comfort, and daylighting [135,140,151,152]. Third, a “surrogate-model + global optimization” pipeline trains RF/XGBoost/SVR/ANN surrogates or applies Bayesian optimization with Kriging/Gaussian processes, iterating at low cost on samples generated via Grasshopper or jEPlus to rapidly approximate multi-objective Pareto sets in large design spaces [153,154]. Fourth, a TRNSYS-driven pipeline targets system-level problems (e.g., HVAC and thermal storage), often combined with evolutionary algorithms/surrogates to co-optimize component and operational parameters [155,156,157]. These pipelines can be flexibly mixed and matched across scales and design objects, but they share a closed-loop structure: parametric geometry/boundaries-high-fidelity simulation, kernels-multi-objective global optimization-Pareto-front decision making, and visualization-(optional) surrogate/sensitivity/uncertainty analyses.

4.4.3. Task–Scale–Objective Alignment

For single-building, three-objective problems (energy × daylight × comfort), parametric geometry is authored in Grasshopper; Honeybee/Radiance/Daysim and EnergyPlus deliver annual daylight and thermo-energy evaluation; NSGA-II/III or hybrid evolutionary methods generate the Pareto front; clustering/DR and MCDM synthesize and rank candidates; and XAI/sensitivity analyses articulate variable effects and robustness [44,91,144].

For early-stage facade/shading generation, the optical pipeline (Radiance/Daysim) coupled with parametric geometry captures structure–optics–thermal interplay inside the optimization loop, with SHAP/Morris decomposing geometric parameter influence and stability [131].

For urban/block-scale microclimate or aggregate energy, SBO–simulator loops built on OpenStudio batch modeling and parallel scheduling permit tractable screening across multi-scenario, multi-weather, and multi-morphology settings [158].

For operations and control, uncertainty-aware surrogates (LSTM/Bayesian) combined with MPC or evolutionary/swarm strategies balance efficiency, comfort, and flexibility in a design-to-control workflow [141].

In BIPV and low-carbon techno-economics, two-stage frameworks co-optimize carbon/cost with energy, daylight, and comfort, typically retaining energy/optical physics cores while leveraging surrogates and MCDM/XAI for interpretability and communicability [133].

4.4.4. Quality Control and Threats to Validity

Geometry–physics inconsistency is a primary threat; we therefore recommend “unified operating conditions and comparable scopes” up front—explicit assumptions for weather files, occupancy, energy, lighting, ventilation strategies, systems, and control logic—before multi-objective search and trade-off analysis [110,159]. This improves scientific comparability and provides reproducible baselines for optimization and validation. By integrating

Table 3. Comparative matrix of leading building-performance tools and capabilities.

Software	Engine/Role	Expertise	Optimization	Parametric	Energy	Daylighting	Airflow	LCC	Carbon	Interop
Rhino/Grasshopper	Parametric workbench	Medium	External (scripts/components)	Native	via EnergyPlus (Honeybee)	via Radiance/Daysim (Honeybee)	AFN via EnergyPlus	Custom/External	Operation (external plug-ins)	Native GH; exports via Honeybee/IFC
Honeybee/Ladybug	Bridging toolkit (GH)	Medium	External (GH/Python)	Native (GH components)	✓ (EnergyPlus)	✓ (Radiance/Daysim)	AFN via EnergyPlus	Custom	Operation (workflow add-ons)	GH components; Radiance/EnergyPlus bindings
EnergyPlus	Physics engine (thermal/energy)	Medium–High	External	Scripts (Param sets/OpenStudio Measures)	✓	via Radiance (preferred)	✓ (AFN)	via add-ins/external	Operation	IDF/gbXML; via OpenStudio
Radiance	Physics engine (optical/daylighting)	High	External	Scripts	—	✓	—	—	—	File exchange
Daysim	Daylighting (Radiance-based)	Medium–High	External	Scripts	—	✓	—	—	—	File exchange
OpenStudio	Model mgmt and batch runs (EnergyPlus)	Medium	External	Measures	✓ (via EnergyPlus)	via Radiance (measures)	✓ (AFN via EP)	External	Operation	OSM/IDF; gbXML
DesignBuilder	GUI for EP/Radiance; QA/visualization	Low–Medium	Built-in (GA module)	Built-in	✓	✓	Limited	Module	Operation	IDF/gbXML
Revit/Dynamo	BIM and scripting	Medium	External	Built-in (Dynamo)	via Insight/EP plugins	via plugins/analysis add-ins	—	External	Operation (plugins)	IFC/gbXML
Python	Scripting/orchestration	Medium–High	External (BO/EA libs)	Scripts	via APIs	via APIs	via APIs	External	External	APIs/CLI (EP, Radiance, OS)
MATLAB	Scripting/numerical	Medium–High	External (toolboxes)	Scripts	via co-sim/APIs	via file exchange	via EP AFN	External	External	APIs/File I/O

Note: The symbol “✓” indicates that the software possesses native capabilities or built-in modules for the specified analysis.

more interpretable and engineering-transferable results can be obtained across the three dimensions of “performance-cost-risk”.

Looking across algorithms and tools, our assessment is that no single stack is universally best. Parametric and evolutionary methods remain a robust baseline because they preserve geometric transparency and align with the way designers reason about form and constraints, but they can be slow for high dimensional or tightly coupled physics. Surrogate and Bayesian approaches give substantial efficiency gains when simulations are expensive, yet they demand careful sampling, validation, and reporting of uncertainty in order to avoid overconfident conclusions. Data- and model-driven tools open new possibilities for rapid feedback and complex pattern discovery, but they also introduce opacity and dependencies on data quality, training choices, and tool interoperability. For engineering ready PDGD, we therefore advocate explicit reporting of algorithm choices, tool chains, validation steps, and failure modes so that readers can judge not only what was achieved but also how reliable and transferable the reported workflows are.

4.5. Comparative Evaluation of Workflows

Drawing on Figure 13 (the specific data can be found in Supplementary Materials D’s Supplementary data), we compare three workflow archetypes: PEMOO, SBO, and Data/Model-Driven Generation (DMDG). Meanwhile, Hybrid/Engineering Optimization (HEO) (the combination of these three methods) is also a common practice. In terms of prevalence (Figure 13a), PEMOO dominates with 139 assignments (65.9%), followed by DMDG 55 (26.1%), and SBO 17 (8.1%). Regarding depth of adoption (Figure 13b), 63.2% of studies use a single workflow, while 34.9% combine two and 2% combine three—evidence of an emerging hybridization paradigm. Figure 13c illustrates the correlation between the three types of workflows and the design object. This is particularly relevant as the majority of existing research focuses on applying PEMOO to optimize the envelope/facade and single-building massing/layout.

4.5.1. Parametric + Evolutionary Multi-Objective Optimization

PEMOO remains the most deployable baseline for PDGD: designers expose a compact, interpretable set of geometry/construction variables (e.g., massing, orientation, WWR, shading), evaluate candidates with EnergyPlus/Radiance, and drive population-based search (NSGA-II/III, GA) to form stable Pareto fronts that support transparent trade-off selection and scheme screening [160]. This pattern has repeatedly delivered concurrent gains in energy and comfort/glare on envelope and single-building tasks [155]. Figure 14 lists the common processes of this workflow.

Across climate zones and building types, NSGA-II and its variants frequently deliver competitive fronts for cold-region envelopes and mixed-mode strategies, confirming the method’s robustness to context and objective scaling [131,145]. For residential and mixed-use cases, MOO systematically balances heating/cooling loads with daylight and cost, yielding non-dominated sets that enable transparent post hoc selection by design teams [161]. Parallelized or coordinated MOO implementations further reduce wall-clock time without sacrificing solution quality, which is crucial when hundreds to thousands of simulations are required [4]. Finally, recent pipelines enrich parametric MOOs with lightweight predictors to prioritize promising regions of the design space while retaining the interpretability of explicit parametric controls [12].

4.5.2. Surrogate/Bayesian Optimization

SBO targets the core bottleneck of expensive, tightly coupled simulations by closing the loop “small-sample DoE → surrogate fitting (RF/GP/boosting) → sequential acquisition (EI/UCB) → infill simulation.” Across office, education, and mixed-use cases, SBO reduces the number of high-fidelity evaluations needed to approach the Pareto set, while maintaining solution quality under multi-objective formulations [18,128,162]. Where public-building retrofits require strict compute budgeting and carbon co-objectives, Bayesian optimization with learned surrogates has proven sample efficient and deployment friendly [74,128]. Figure 15 lists the common processes of this workflow.

SBO closes the loop “high-fidelity simulation–small-sample sampling–surrogate fitting–sequential search” and is particularly suited to high-dimensional, expensive tasks. Response surfaces/Kriging/Gaussian Processes provide uncertainty-aware predictions from limited samples; acquisition functions such as Expected Improvement (EI) enable fast exploration–exploitation balance, markedly reducing the samples needed to approach the Pareto front [163]. For cross-scale problems, surrogates are often paired with multi-objective search and ML predictors to compress sampling and stabilize extrapolation—this pairing links energy [50], daylight [46], and comfort [20] targets to form/morphology and material/assembly variables, thereby extracting interpretable design regularities on limited budgets. Compared with the baseline evolutionary route, SBO’s distinctive strengths are sample efficiency and uncertainty management: (1) it can discover promising optima under very small samples especially in coupled co-simulation with Radiance/CFD/energy, and (2) it natively supports confidence interval reporting and risk control for engineering delivery. Its weaknesses are extrapolation risk and surrogate degradation under strong non-stationarity or regime shifts (e.g., opening-state switches)—two issues that call for adaptive resampling and multi-surrogate ensembling [113,164]. Practically, ML regressors (GP/Kriging, RF, XGBoost, MLP) can serve as ‘surrogate evaluators,’ while BO with EI/UCB or hybrid heuristics conducts the search—specifically, this combination substantially reduces expensive simulation calls in small-sample, high-cost energy/comfort optimization.

4.5.3. Data/Model-Driven Generation

Positioning data as the generative engine, DMDG provides near-real-time feedback by learning geometry to performance mappings and, increasingly, by directly synthesizing candidate solutions. Interpretable ML (RF/XGBoost/LightGBM with SHAP) predicts EUI/UDI and exposes variable contributions that guide parametric edits, enabling rapid pruning of dominated regions before physics-in-the-loop verification [158,165]. Data-driven pipelines also co-optimize with evolutionary search, for example, learned predictors inside NSGA-II/III, to sustain Pareto quality under tight budgets [166]. For end-to-end generation, emerging work combines generative models (e.g., GANs) with reinforcement learning to synthesize morphology/facade options conditioned on performance cues, then filters them via surrogate or simulation checks [167]. Moreover, the introduction of Artificial Neural Network (ANN) models for optimal design has also become a common practice [168,169]. Figure 16 lists the common processes of this workflow.

A key advantage is speed and breadth; the liabilities are data dependence and controllability. Domain shift across climates/building types can erode accuracy, motivating online transfer/continual learning to keep predictors calibrated as projects move from design to operation [170]. Consequently, DMDG works best when coupled with physics checks, rule/constraint injection, and explicit uncertainty reporting.

4.5.4. Hybrid and Engineering-Oriented

Hybrid/engineering-oriented workflows act as the “connective tissue.” A total of 56 out of 153 studies combined more than two workflows. One class employs two- or multi-stage frameworks: parametric + evolutionary methods serve as a front-end coarse search, while surrogate/ML models provide back-end refinement and uncertainty quantification (UQ) to balance cost–performance–risk. Another class builds multimodal/multiphysics pipelines that unify climate–site–geometry–materials–systems, bringing comfort–energy–carbon–cost jointly into objectives/constraints; these pipelines are often more robust in “both–and” scenarios but demand stronger process governance (versions, data baselines, validation protocols) [88,171].

Decision sequence (three steps) and operational guidance

Table 4. Comparison matrix of three generative workflows.

Workflow	Data Requirement	Compute Cost	Explainability	Generalization	Implementation Complexity	Typical Toolchain
PEMOO	Low–Medium (simulation-only; explicit variables)	High (many simulations; sensitive to budget)	High (Pareto + variable sensitivity)	Medium (scenario-specific; robust within space)	Medium (plug-in chains; standard)	Rhino/Grasshopper + Honeybee/Ladybug + EnergyPlus/Radiance + NSGA-II/PSO
SBO	Medium–High (initial DoE + sequential sampling)	Medium (expensive sims → cheap surrogate)	Medium–High (feature effects, UQ)	Medium (valid near sampled design space)	Medium–High (fit, UQ, sampling policies)	OpenStudio/E+ or Radiance/CFD + DoE + Kriging/GP + EI/PI + Multi-objective BO
DMDG	High (curated labels/paired geometry–performance)	Low–Medium (fast inference; training may be high)	Medium–Low (needs XAI/constraints)	Low–Medium (depends on data domain)	High (data curation, pipelines, QA)	BIM/Parametric model + Feature eng. + ML/DL/RL + XAI + physics checks

.

• Set physical fidelity and coupling depth by object/scale (e.g., EnergyPlus/Radiance loose coupling vs. EnergyPlus fine-step co-simulation).
• Under compute/time constraints, prioritize SBO for expensive segments (annual daylight, co-simulation) or use ML for rapid generation of design alternatives or partial substitution of high-fidelity simulations.
• Use parametric + evolutionary to provide an interpretable Pareto set and a structured variable space, serving as the front-end coarse search within hybrids.

For engineering delivery, researchers should perform sensitivity screening and dimensionality reduction as a pre-step; during optimization, include UQ measures (surrogate-error cross-validation, hypervolume improvement threshold ε, confidence-interval visualization) and convergence governance (sample budgeting, early-stopping rules). In closing, follow a strict reproducibility checklist (tools/versions; weather and operating conditions; control logic; mesh and time step; random seeds; data/script paths) and consolidate evidence with representative Pareto solutions + external validation. In this way, one builds a reusable, verifiable, and object/scale-adaptable design decision asset within the three-dimensional space of ‘performance–cost–risk’.

From our perspective, the three workflow archetypes should be seen as complementary rather than competing. PEMOO workflows are well suited to problems where interpretability, explicit constraints, and dialog with stakeholders are central, such as envelope design in regulated contexts. SBO workflows are most valuable when simulation cost or uncertainty is the main bottleneck and when the design team can invest in careful experiment planning and model checking. DMDG excels in early exploration and in tasks where many design variants must be screened quickly, but it requires explicit strategies to encode rules, constraints, and risk. In practice, we anticipate that mature PDGD applications will increasingly combine these archetypes within the same project, moving from transparent parametric baselines through surrogate acceleration to selective use of data driven generation where it adds clear value.

4.6. Data and Validation

Robust data practices and validation techniques are crucial for performance-driven design given the complexity of simulation outputs and the need for credible results. We observed that PDGD studies can be viewed in an “evidence pyramid” structure.

At the base of this pyramid are simulation-generated datasets—the large volumes of data produced by parametric energy simulations, daylight renderings, CFD analyses, etc. These provide abundant information under controlled conditions and are the primary data source for training surrogates or evaluating design options. However, they carry the risk of simulation bias: if the model assumptions are off or if a scenario was not simulated, the conclusions may not hold in reality. Many studies explicitly acknowledge this by validating their simulation models against known benchmarks or simpler analytic cases before trustfully using them in optimization [152,172].

The next tier of the pyramid involves surrogate model validation—when machine learning models (ANNs, regression, etc.) are used to approximate simulation results, researchers use techniques like cross-validation or test-train splits to ensure the surrogate predictions are accurate within the design space. For example, a study might train a neural network to predict energy use from design parameters and report that it achieves, say, 5% error on a held-out test set, giving confidence that the surrogate can guide the optimizer. Many papers provide metrics such as R², RMSE, or mean absolute error for their predictive models, demonstrating high fidelity [21].

Above the surrogate validation comes cross-case or multi-case validation. This is less common: it refers to applying the generative workflow to multiple different cases (e.g., different climates, different building types) to test its generality. Only a minority of studies do this, but those that do provide valuable evidence of robustness. For instance, a study optimized a facade design in several distinct climate zones and found consistent energy-saving patterns, thereby strengthening confidence that the method works beyond a single scenario [82]. Such cross-case studies are valuable but logistically challenging, which explains their rarity [163].

Finally, at the top of the pyramid is real-world validation, meaning comparisons with actual measured data from buildings or experimental field studies. This is the scarcest category in current PDGD literature. A few notable exceptions exist; for example, a study might apply a generative design solution to an actual building and then monitor its performance post-implementation, or use monitored building performance data to calibrate their simulation model (as in an optimization study that calibrated a building energy model with a year of utility data before optimizing retrofit measures). Real-world calibration is more common in building operations studies than early-stage design, but some generative design research does incorporate calibration to ensure the baseline model is accurate. Overall, the field still leans heavily on simulation-based evidence [173].

In short, the evidence pyramid is “simulation–surrogate/prediction–multi-case–field,” as shown in Figure 17; the closer to real-world measurements and multi-project comparisons, the greater the reusability and engineering credibility of findings.

4.6.1. Calibration, Uncertainty, and Sensitivity

Another critical aspect of validation is uncertainty analysis and sensitivity analysis, which only a subset of PDGD studies currently include. Uncertainty analysis evaluates how sensitive outcomes are to variations in inputs or assumptions. Techniques like Monte Carlo simulation or bootstrapping are used to quantify result variability. For example, a design might be optimized under a typical climate year—uncertainty analysis would test it under a hotter year or different user behavior to see if it still performs adequately. We found relatively few studies explicitly doing this, but those that did provided confidence intervals or error bars on their performance metrics, which is extremely useful for decision-makers (e.g., showing that a certain facade design saves 20 ± 5% energy across weather variability) [174]. Sensitivity analysis, on the other hand, identifies which input parameters most strongly affect the outcomes. Methods range from simple one-at-a-time sensitivity tests to more rigorous global sensitivity indices. Several studies incorporate a sensitivity analysis prior to optimization—for instance, to reduce the dimensionality of the problem by fixing insensitive parameters [121]. Others do a post-optimal sensitivity check to interpret the Pareto solutions (which aligns with the trend of using explainable AI (XAI) tools like SHAP values to rank the importance of design features) [175]. Including these analyses greatly enhances the transparency and reliability of PDGD results [158]. We encourage future studies to routinely include UQ and sensitivity analysis, as they guard against overconfidence in a single optimal solution and reveal how robust a solution is to change.

4.6.2. Reproducibility Checklist and Data Governance

To support replication and validation, we propose a reproducibility checklist (see

Table 5. Reproducibility and data governance checklist.

Category	Minimum Fields to Report	Example Format	Storage/Artifact
Software stack and versions	Name; version; build; solver flags; GPU/CPU; OS; date.	EnergyPlus 24.1.0; OpenStudio 3.7.1; Radiance 5.4; Ladybug 1.8.0; Windows 11 23H2.	README.md + requirements.txt + environment.yml; Appendix ‘Versions’.
Weather and climate data	Dataset name; provider; version; URL/DOI; year span; morphing method.	TMYx (2007–2021) EPW, Melbourne Docklands; UHI morphing via UWG v5.4.	data/weather.
Geometry and parametric schema	Variable → bounds; unit; encoding; dependency notes.	WWR_N/S/E/W ∈ [0.2, 0.8]; overhang_d ∈ [0, 1.2] m; ‘if atrium = true then core_depth ≤ 14 m’.	schema.yaml/model.json; Appendix ‘Variables’.
Schedules and controls	Schedule source; setpoints; control type; ventilation; infiltration.	Cooling 26 °C with 1 °C deadband; CO2 control 900 ppm; night set-back 30 °C.	controls.yaml; BPS input snippets.
Simulation settings	Δt; solver tol; rad params (-ab -ad -ar -aa …); grid spacing; material reflectance.	EnergyPlus 10 min; Radiance -ab 5 -ad 4096 -ar 256 -aa 0.1; grid 0.5 m.	sim_settings.md; input templates in repo.
Data splits and cross-validation	Split ratios; k; stratification; seeds; scaler; feature list.	80/20 holdout; 5-fold CV; seed = 42; features: WWR, SHGC, tilt…	ml/README.md; configs/.yaml.
Acquisition and search policy	Algorithm; population/iterations; EI/UCB params; early-stop rules.	NSGA-III, pop = 80, gen = 120; BO with EI, ξ = 0.01, batch = 4; stop if hypervolume < 0.1%.	experiments/logs/.csv.
Carbon and LCA factors	Grid EF (year, source); LCA DB/version; module scope; uncertainty range.	Grid EF 2022 CN North; OneClick vX.Y; A1–A3 + B4; ±15%.	lca/README.md; data/lca/.csv.

) that details information authors should report.

• Version pinning and environment capture. Pin the exact versions of the BPS core and all plug-ins (EnergyPlus/OpenStudio/Radiance/Ladybug–Honeybee, plus OS and hardware) and publish them in both the manuscript and repository. Version drift is the most common cause of reproduction failure in GD–BPS studies; studies explicitly reporting tool versions and weather datasets enable clean re-runs and comparison baselines [133,176,177].
• Provenance of weather, geometry, schedules, and controls. For weather/climate inputs, state dataset type (TMY/AMY), source and version, and whether any morphing/UHI adjustments were applied; mixing years across options should be avoided unless justified [98]. For design variables, publish the parametric schema (bounds/units/constraints) so that surrogate training and evolutionary sampling can be reconstructed [105]. Report occupancy/lighting/HVAC control logic and setpoints, including setbacks and ventilation modes, since hidden GUI defaults frequently override intended assumptions [119].
• Simulation settings and optical/CFD parameters. Record time step, convergence tolerances, daylight grids and Radiance parameters, material reflectances, view definitions, and any mesh or solver options to ensure apples-to-apples comparison—especially along Pareto fronts [150,178].
• Learning, search, and evaluation protocols. For surrogates/ML, declare train/validation/test splits, cross-validation strategy, random seeds, feature lists and preprocessing, and guard against leakage across climates or typologies [45]. For SBO and evolutionary baselines, document acquisition/search policies (e.g., EI/UCB parameters, batch size), population/iterations, and stopping rules; non-repeatable policies should include random seeds and hypervolume/ε-progress criteria [74].
• Carbon and LCA datasets. Declare operational carbon factors (year, source) and the embodied carbon database and version (regional scope and modules), and publish interval estimates to reflect database divergence [138,161].

5. Discussion

Across the corpus, commercial and residential buildings are most prevalent; education/public follow, while industrial and heritage are scarce. This mirrors the global stock and operating hours and aligns with policy priorities for energy and carbon. Meanwhile, case designs are distributed across RB, AB, and SBM with a slight edge for archetype buildings.

Across five design objects, envelope/facade appears most frequently, followed by single-building massing/layout, roofs/BIPV. At the variable level, WWR, glazing/window-opening proportion, and shading geometry dominate. This distribution is sensible: these variables have clear physical couplings and strong interpretability, with tractable dimensionality and crisp constraints—conducive to stable sensitivity structures and a visualizable Pareto front. Examples include explorations of the WWR × shading geometry × operability triad for glare–load–daylight trade-offs [80]; under complex geometry or urban obstruction, the “optimal window” for WWR and opening ratios narrows markedly [81].

Objectives exhibit a layered structure: energy (EUI and peak/seasonal loads) is most frequent, followed by daylight and thermal comfort, then carbon and economics. EUI, as an annual systems-efficiency indicator, complements heating/cooling loads to form a two-tier backbone for optimization and trade-off analysis. In practice, EUI is often co-optimized with UDI/sDA/ASE (daylight) and PMV/PPD/overheat hours (comfort), forming a three-objective frame [179]. At early design stages, carbon typically focuses on operational carbon, keeping a one-to-one mapping to energy and sidestepping uncertain embodied inputs [132].

Algorithm usage shows a “strong head and long tail”: MOO is central, with NSGA-II cited in most of studies. On tools, Rhino/Grasshopper + Ladybug/Honeybee is the high-frequency workstation and evaluation stack, with Python/MATLAB as orchestration glue. The typical pattern is parametric geometry in Grasshopper coupled to EnergyPlus/Radiance, NSGA-II builds a Pareto front, followed by clustering/dimension reduction and MCDM for synthesis/ranking [180].

Comparing three general routes shows PEMOO excels at interpretable Pareto sets and variable sensitivity, making it a practice-ready baseline; SBO offers sample-efficient search in high-fidelity, multi-physics segments; DMDG supports rapid feedback and inverse generation. A characteristic pattern is a two-stage/hybrid flow: evolutionary global exploration up front, then surrogate/Bayesian refinement or ML acceleration downstream.

Evidence can be organized into four tiers: simulation—surrogate/prediction—cross-case replication—field measurement; moving upward increases reusability and credibility. Accordingly, adopt a reproducibility and data-governance checklist.

Despite the progress observed, several challenges and limitations persist across the surveyed PDGD studies. Figure 18 summarizes the key cross-cutting challenges together with the corresponding recommendations proposed in this review, and the following paragraphs discuss each point in more detail.

(1) Objective and metric inconsistencies.

There is a lack of standardization in how performance objectives are defined and reported, which undermines comparability between studies. For example, “energy savings” might be reported as total kWh in one study but as a percentage reduction or an EUI in another, making direct comparison difficult [21,155]. Similarly, daylight performance might be variously measured by UDI, sDA, or illuminance levels, even though the daylighting community has proposed more unified dynamic metrics such as sDA and UDI for some time [181,182].

In the comfort realm, some studies report improvements in PMV, which is a unitless index, while others report PPD percentage or hours of exceedance [82,134]. These heterogeneous conventions can even invert the apparent benefit direction; for instance, one study treats an increase in PMV toward 0 as “good” since comfort improves, whereas another looks at a decrease in PPD as “good.” Without clear, consistent reporting—such as always stating whether a higher metric value is better or worse—it is easy to misinterpret results across papers. A specific example is glare: some papers optimize to minimize ASE, but others maximize daylight glare probability (DGP) compliance, where a higher percentage of time below a threshold is better [160,178]. If not carefully described, one might think they conflict when they actually agree.

Recommendation: Future research should adopt harmonized metrics, such as always including standard metrics like sDA for daylight and PMV/PPD for comfort, and explicitly state the “direction” of improvement for each. Doing so would enable more direct aggregation of evidence.

(2) Computational expense and convergence issues.

High-fidelity multi-objective optimization remains computationally intensive. Many studies using brute-force evolutionary algorithms report long run-times or heavy computational loads, especially when simulations are detailed. There is a risk that some reported “optimal” solutions are not truly converged Pareto fronts but rather the best found under a limited simulation budget. Surrogate-assisted optimization can reduce the number of simulations needed but introduces other complexities such as model training time and the risk of surrogate inaccuracy [20,45]. Similarly, one-shot generative models, which we term data/model-driven generation, produce solutions quickly but rely on the quality of training data and may miss extremes.

In our review, we saw few studies discussing convergence criteria or performing a posteriori check for optimality; many simply stop at a fixed generation count or when computing time is exhausted [125]. This means some published Pareto sets could be sub-optimal if more iterations were run. Methodological work on sequential and multi-stage schemes, such as metamodel-based or sequential optimization workflows, shows promise in balancing exploration and exploitation but is still relatively rare in the PDGD corpus [177].

Recommendation: Adopt robust stopping criteria, such as monitoring hypervolume improvement and stopping when it plateaus, and report them. Additionally, hybrid strategies are promising. For instance, run a global evolutionary search for a few generations, then switch to a local optimizer or surrogate refinement for convergence [114]. Such hybrid schemes that combine global and local search were suggested in a few studies to manage the balance between exploration and exploitation, but they need to be more widely implemented [19]. Simply put, the field must tackle the trade-off between solution quality and computational effort more explicitly, to ensure we are genuinely pushing performance boundaries and not just approximating them under time constraints.

(3) Limited validation and reproducibility.

As noted, relatively few studies go beyond simulation to validate results in other contexts or in real buildings. Cross-case studies, which apply the same method to multiple building types or climates, are sparse [82], and field validation, which involves monitoring a built project, is even rarer [183]. This raises questions about the external validity of some optimized designs: will they perform as expected outside the simulation environment? This concern echoes broader discussions in the building performance gap literature, which highlight discrepancies between predicted and measured performance and call for more systematic validation frameworks [184].

Moreover, many papers do not provide enough information for independent reproduction. Crucial details like weather files, occupancy assumptions, or even which version of a tool was used are sometimes omitted. This lack of transparency can lead to “black box” results that practitioners may distrust, and mirrors reproducibility concerns raised more generally in building energy benchmarking and cross-tool validation studies [185]. We also observed potential threats to validity such as reliance on default settings that are not reported. For example, using a default HVAC control in EnergyPlus without mentioning it could affect outcomes, and known limitations in specific modeling domains can strongly influence results if they are not acknowledged [156]. Another issue is mixing data sources, such as using typical year weather for energy but an extreme year for comfort without clarifying. Without meticulous reporting, it is hard to know if an impressive result is due to a real design improvement or an artifact of an assumption.

Recommendation: Increase the emphasis on reproducibility: Publish supplemental data and scripts when possible, and at minimum follow a checklist such as the one provided in Table 5 to document all key parameters. Additionally, the community could establish common test cases or benchmarks—for example, standardized EnergyPlus models or prototype buildings used across different optimization studies, similar to existing benchmarking efforts in residential and cross-tool studies [186]. A standard office building model used across different optimization studies would facilitate apples-to-apples comparisons. Encouragingly, some recent reviews and methodological contributions have called for exactly this kind of standardization and cross-study verification [184,187].

(4) Toolchain bias and interoperability issues.

The heavy use of a relatively uniform set of tools such as Rhino/Grasshopper, Ladybug Tools, and EnergyPlus, often on Windows OS, means the field might be overfitting to a particular toolchain. Many PDGD workflows in our corpus rely on BIM–parametric environments tightly coupled to a single simulation backend [140,141], which has enabled great progress but may bias research toward problems that are easily solvable with those tools. For instance, they facilitate early-stage design of building form and envelope, as opposed to complex retrofits involving control systems or materials, which are harder to parametrize in typical Grasshopper-based workflows [169].

There is also a concern that reliance on specific software versions without backward compatibility can make studies obsolete or irreproducible over time. This phenomenon is known as “version drift” in computational research [188]. Only a few works experimented with different tools for the same task. Examples include comparing EnergyPlus with an alternative engine, or using both Rhino/Grasshopper and a BIM-based tool to generate geometry [186]. Without such comparisons, we cannot be sure that the chosen tools are not introducing unintended biases. For example, EnergyPlus has known limitations or modeling choices in certain natural ventilation scenarios, which can affect any PDGD result in that domain if the limitations are not recognized [189].

Recommendation: Promote tool interoperability and cross-validation. Where feasible, authors should test their workflow with an alternate tool or at least ensure key assumptions such as weather data and material properties are consistent if transferring models between platforms. Using open standards such as gbXML, IFC, and Modelica for systems could help make workflows more tool-agnostic. Containerization through Docker images has been suggested as a means to capture the computing environment and share it, reducing issues of setup differences. Ultimately, increasing the diversity of software and techniques tested will strengthen confidence that PDGD findings are not just quirks of a specific digital workflow but are truly general insights.

In conclusion, while PDGD for buildings has seen rapid advancements, addressing these limitations will be key to its maturation as a field. Consistent metrics, efficient yet rigorous optimization, thorough validation, and robust toolchains will collectively ensure that PDGD research translates into reliable, real-world-ready solutions for high-performance buildings.