BIM-Integrated Multi-Objective Optimisation of Prefabricated Construction Configurations: A WBS-Based Generative Decomposition Framework

Abstract

Building Information Modelling (BIM) workflows for prefabricated construction lack mechanisms that generate and compare alternative component configurations directly from a design model. Existing approaches define the optimisation search space manually and outside the model, address only one or two criteria, and treat the Work Breakdown Structure (WBS) as a post-design planning tool. This paper reinterprets the WBS as a generative decomposition mechanism. A WBS Engine decomposes the geometry of an existing BIM model into prefabricated subsystems before design decisions are fixed, producing the search space for optimisation without manual parametrisation. A Scenario Evaluator queries a database of 47 prefabricated components, and NSGA-II evaluates 60 configurations against four objectives. These are total cost, embodied carbon, assembly factor and number of lorry trips. Applied to a residential case study implemented in Dynamo, the prototype identified 16 non-dominated solutions. The best compromise configuration achieved a total cost of £150,444.01, 127,731.00 kgCO₂e, an assembly factor of 0.190 and 10 lorry trips. Wall module size accounted for 17.4% of cost variation, while floor module size governed assembly complexity. The findings show that BIM-WBS integration with multi-objective optimisation supports informed early-stage decisions in industrialised construction.

1. Introduction

Productivity, cost overruns and environmental impact remain persistent issues in the architecture, engineering and construction (AEC) sector. Industrialised construction, including prefabrication, modular construction and manufacturing-based delivery, addresses these issues by transferring activities to controlled factory environments, improving quality and reducing waste [1,2]. Realising these benefits depends on informed decisions at early design stages, where the choice of construction system simultaneously determines cost, embodied carbon, assembly complexity and logistics. In current practice, these criteria are evaluated sequentially by separate disciplines, without a mechanism that links them to specific design alternatives within the Building Information Modelling (BIM) model [1].

BIM provides the representational and computational infrastructure to integrate these decisions [2]. In current applications, BIM remains predominantly descriptive, supporting documentation and coordination but not the systematic generation and comparison of alternative construction configurations [1,3]. Multi-objective optimisation (MOO) methods such as NSGA-II have been applied in BIM and adjacent construction contexts [4,5,6,7], yet the search space is defined manually and external to the model. The Work Breakdown Structure (WBS) has been integrated with BIM for asset lifecycle management [8] and digital collaboration [9], although it has been treated consistently as a planning instrument that organises work already decided. The early design stages where these decisions matter most therefore remain unsupported by an integrated tool.

This paper reinterprets the WBS as a generative decomposition mechanism rather than a post-design planning instrument. The WBS Engine decomposes the geometry of an existing BIM model into prefabricated subsystems before design decisions are finalised. This decomposition produces the discrete configuration space that an optimisation algorithm requires, removing the need for project-specific manual parametrisation. The configuration space adapts automatically to changes in the BIM model, which allows the tool to operate during the schematic phase when the model is still evolving.

The paper makes three contributions. First, it demonstrates that WBS-based decomposition of a BIM model can automatically generate a discrete construction configuration space for multi-objective optimisation, an application not identified in the reviewed literature. Second, it shows that WBS decomposition, component database queries and NSGA-II optimisation can be integrated within a single Dynamo–Revit workflow using standard BIM tools and the open-source pymoo library, without external platforms. Third, the resulting Pareto front provides quantitative evidence that wall module size is the primary determinant of cost and carbon variation in prefabricated construction, and that the best compromise solution occupies an intermediate position on the front rather than the individual extreme of any objective. These contributions are interdependent. The automatic generation of the search space from the WBS is what allows the Pareto front to communicate trade-offs directly linked to constructive decisions in the BIM model. The prototype is implemented and validated through a residential case study following Design Science Research (DSR) criteria [10].

2. Literature Review

2.1. BIM and DfMA for Prefabricated Construction

Modular construction and prefabricated components offer well-documented advantages in quality, labour efficiency, cost and environmental impact [11]. DfMA formalises the design principles that exploit these advantages by rationalising components for factory production and systematic assembly [12]. RIBA [13] and Abrishami and Martín-Durán [14] have highlighted the potential of BIM-DfMA integration to improve manufacturability assessment, reduce material over-specification and facilitate component-based decision-making. Doan et al. [1] identified that, in current practice, cost, constructability and environmental performance are addressed separately, and that both professional practice and academic literature lag behind the available technological capabilities.

BIM object libraries have been developed to support component-based design aligned with prefabrication [15], and ontological frameworks have been proposed to formalise the semantic relationships between BIM objects and DfMA constraints [16]. These contributions establish the conceptual infrastructure for BIM-DfMA integration, but do not provide the computational mechanism for the automated and multi-criteria generation and evaluation of configurations. Getuli et al. [15] advanced in this direction by proposing a parametric methodology for BIM object libraries, though focused on the parametrisation of a single attribute rather than on multi-criteria optimisation. Laovisutthichai and Lu [17] provided an analytical framework for DfMA in modular buildings, identifying the need to systematically evaluate prefabricated component configurations against multiple criteria.

2.2. WBS Integration with BIM

The WBS has been integrated with BIM for asset lifecycle management [8] and for decentralised collaborative platforms using blockchain and IoT [9]. In both cases, the WBS structures work that has already been decided during the design phase. Jalilzadehazhari et al. [18] demonstrated how BIM and multi-criteria methods can be combined for the evaluation of design alternatives, but without WBS-based decomposition. The use of the WBS as a generative design mechanism, transforming BIM model geometry into a discrete set of prefabricated configurations that feed an optimisation algorithm; represents a distinct and unexplored application. The distinction between WBS as a planning tool and WBS as a generative design mechanism is analogous to the broader evolution of BIM from documentation platform to active analytical environment: in both cases, the same underlying data structure is reinterpreted to serve a fundamentally different purpose.

2.3. Multi-Objective Optimisation in BIM Environments

NSGA-II [19] has been applied in contexts adjacent to BIM, evaluating objectives such as cost, carbon, logistics and sustainability [4,5,6,7]. Moradabadi et al. [4] (the closest technical precedent to this work) implemented NSGA-II within Dynamo for lifecycle cost optimisation but defined the search space manually and externally to the model, preventing the space from adapting automatically to model geometry. Padala and Kumar [5] integrated multi-objective optimisation with BIM for sustainable construction management but limited the approach to energy and structural objectives without connecting to prefabrication-specific metrics. Zhang et al. [20] optimised prefabricated component logistics using two-stage methods, demonstrating that logistics objectives are sensitive to module size in ways that cost-only analyses fail to capture. Peng et al. [6] and Zhang and Wang [7] demonstrated that intermediate solutions on the Pareto front and inflexion points on the front provide the greatest practical value for decision-making, a result consistent with the findings of this study. Li et al. [21] addressed the carbon footprint of prefabricated buildings from a circular economy perspective, reinforcing the relevance of including carbon as an explicit optimisation objective.

2.4. Identified Research Gap

The literature shows growing interest in construction industrialisation, particularly around research lines such as modular construction, prefabrication and DfMA, and their integration with BIM. Prior studies have investigated the use of WBS in BIM for planning and collaboration, as well as the application of multi-criteria optimisation techniques in BIM environments. However, these studies have developed independently rather than in an integrated manner. Existing BIM-DfMA projects focus on representation, coordination and standards verification, with limited capacity to support automated evaluation at early design stages. Similarly, whilst the WBS has been integrated with BIM to improve digital planning and collaboration, its use as a generative mechanism for early design decomposition in prefabricated construction projects remains largely unexplored. Furthermore, multi-criteria optimisation approaches in BIM tend to focus on specific performance aspects and rarely connect with component-based design or DfMA principles. As a result, there is a gap in BIM-integrated decision-making: practical tools are lacking that can combine automatic WBS-based design decomposition, the use of prefabricated components, and multi-criteria evaluation and optimisation to generate and compare alternative construction scenarios from an existing architectural BIM model during early design stages. Addressing this gap is essential to facilitate informed decision-making and to fully exploit the potential of construction industrialisation strategies.

Building on this, the proposed prototype addresses three principal limitations present in existing BIM-MOO approaches. First, manual definition of the search space in BIM-MOO tools [4,5,18] requires a high level of parametrisation expertise and must be redefined for each project, limiting utility at early design stages where alternatives need to be explored rapidly and flexibly. Second, the absence of a WBS-based constructive decomposition causes optimisation to operate on geometric or abstract parameters rather than on prefabricated components linked to measurable performance attributes such as cost, carbon, assembly and logistics. Third, many BIM-MOO approaches focus on one or two evaluation criteria, reducing their capacity to represent the multi-criteria nature of decisions in industrialised construction. This prototype therefore adopts three principal design decisions: it automatically generates the search space from the WBS decomposition of the BIM model; it connects evaluation to a component database containing cost, carbon, assembly and transport attributes derived from standardised sources; and it enables simultaneous evaluation against four objectives. Whilst the system operates on a representative component database rather than on verified industrial catalogues, it demonstrates the technical feasibility of the approach and establishes the basis for future extension to real contractual data.

3. Methodology

3.1. Research Design

This research adopts an applied, design-oriented methodology aligned with the principles of Design Science Research (DSR) [10]. This approach is appropriate because the study does not merely analyse an existing phenomenon but develops and evaluates a digital artefact: a BIM-integrated framework and prototype to support early-stage decision-making in prefabricated construction projects. The prototype combines automatic WBS-based design decomposition, a prefabricated component database and a multi-objective evaluation and optimisation process. The research strategy corresponds to a single exploratory–descriptive case study [22], in which an existing architectural BIM model is used as input to generate, evaluate and compare alternative construction configurations without modifying its original geometry. The methodology is primarily quantitative in nature, drawing on data extracted from the BIM model and the component database, including dimensions, technical constraints, cost, embodied carbon, assembly and logistics. The results therefore carry comparative value within the case study and are not intended as absolute project-level predictions.

This design-oriented approach rests on a pragmatist paradigm, which Design Science Research adopts because it judges knowledge by its usefulness in solving a practical problem rather than by correspondence to an observed reality. Epistemologically, valid knowledge is generated through the construction and evaluation of a working artefact, so the demonstrated capability of the prototype is the primary evidence, while the quantitative inputs of cost, embodied carbon, weight and dimensions are treated objectively as measurable properties from standardised sources. Ontologically, the study takes a pragmatic realist position: the prefabricated components and their performance attributes are real, measurable entities that exist independently of the observer, whereas the construction configurations the framework generates are designed constructs brought into being through the WBS-based decomposition of the model. The reasoning is therefore deductive in evaluation, assessing the artefact against predefined criteria, yet exploratory in construction, where the decomposition mechanism is developed and refined iteratively.

Prototype evaluation was structured around three DSR-derived criteria. First, technical operability was assessed by running the complete Dynamo graph with different random seeds, verifying that all outputs were generated correctly and without errors. Second, output coherence was evaluated by validating the Pareto front, confirming that the 16 resulting solutions satisfied the strict non-dominance principle, that the front represented logical trade-offs between cost, carbon, assembly and logistics, and that min–max normalisation enabled identification of a balanced solution corresponding to scenario P10. Third, decision utility was verified through the automatic generation of a PDF report, a Bill of Materials (BOM) table and colour overrides in the BIM model, ensuring that the results were comprehensible and usable by a design team. In this way, the research design enables demonstration of the technical feasibility of the prototype and its capacity to convert an existing BIM model into an active basis for comparing prefabricated scenarios at early design stages.

3.2. Case Study

The case study is a residential building model developed using BIM software, comprising 68 wall elements (external and internal), floors and roof. All plan dimensions are multiples of the standard prefabrication module of 1200 mm, which eliminates infill panels in floors and roof. Wall thicknesses are 315 mm, 414 mm and 440 mm (external) and 125 mm (internal). Standard panel heights are 3500 mm (external walls) and 3000 mm (internal walls). The model geometry is not modified during optimisation; it acts as a fixed geometric input.

The choice of a regular, modular geometry is deliberate for a proof-of-concept validation. Regular geometries eliminate infill panels from floors and the roof, allowing the structure of the Pareto front to be interpreted exclusively in terms of module size choices. This boundary condition is acknowledged as a limitation (Section 6.4) and does not restrict the generality of the framework, which handles irregular geometries through the INFILL and CLOSURE component logic within the Component_Library. The residential typology was selected for its high relevance to volumetric and panelised prefabrication discussed in the BIM-DfMA literature [13,14,17].

4. Development and System Framework

4.1. Component Database

The system uses an Excel workbook with five interrelated sheets (Component_Library, Material_Performance, Component_WBS_Map, Components DATABASE, Component_Performance) containing the properties of 47 active prefabricated components. The WBS hierarchy has three levels: the building at level 1, four subsystems at level 2 (floors 1.2, external walls 1.3, internal walls 1.4 and roof 1.5), and eight component categories at level 3, which are the leaf nodes that hold the components. These eight leaf categories are external walls (WBS 1.3.1, n = 21), internal walls (WBS 1.4.1, n = 5), three floor layers (WBS 1.2.1 to 1.2.3, n = 4 each) and three roof layers (WBS 1.5.1 to 1.5.3, n = 3 each). It is important to note that, whilst the WBS 1.1 Structure branch, encompassing foundations, beams and columns, is identified as a fundamental building component within the theoretical WBS framework, it is not considered in the system implementation. This is because the project focuses on systems that can be industrialised through prefabrication, where primary structural elements are typically constrained by prior design decisions and offer less flexibility at later stages. The implemented WBS therefore covers only the envelope and roof subsystems amenable to modular optimisation at early design stages. Figure 1 shows the WBS hierarchy applied to the case study. External walls and internal walls are treated as separate subsystems at WBS levels 1.3 and 1.4 rather than as a single wall category. This separation reflects their different functions and physical properties. External walls use 315 mm, 414 mm and 440 mm thicknesses with a 3500 mm standard panel height, whereas internal walls use a 125 mm thickness with a 3000 mm standard panel height, and the two families carry different cost, carbon and assembly attributes in the component database. The internal subsystem at WBS 1.4 contains internal wall components only. No additional element types are grouped within it.

Within this structure, the Component_Library sheet constitutes the technical core of the database, containing 44 attributes per component structured into four groups. (1) Identity and geometry: Component code, family, type, function and dimensions in mm; variable-dimension components (INFILL and CLOSURE pieces) are identified via Is_Variable and their dimensional ranges. (2) Material composition: Up to five-layer references to Material_Performance (density kg/m³, unit cost £, embodied carbon kgCO₂e/kg) and layer thicknesses. (3) Calculated performance attributes: Total weight (kg), embodied carbon (kgCO₂e) and cost (£) summed across layers using Excel formulae. For example, PF-EXT-600-3500-315 (Timber 100 mm + OSB 12 mm + Plasterboard 13 mm + Mineral Wool 150 mm + Fibre Cement 40 mm) yields 170.15 kg, £182.28 and 111.66 kgCO₂e. (4) Constructive assessment indices: Interface_Score (joint complexity with adjacent elements), Standardisation_Score (degree of standardisation within the constructive family), Assembly_Factor (composite indicator, objective f3) and Transport_Factor_Panel (factor proportional to panel volume, used in the lorry calculation f4). These four indices were assigned directly by the authors to each component using an ordinal 0–1 scale, calibrated against British prefabrication practice and the reviewed literature [23,24,25,26,27,28], given that no standardised metric exists for these attributes in the context of industrialised construction.

It is essential to note that the 47 components in the database are bespoke designs created specifically for this research; they do not correspond to commercially available products. The geometries and layer compositions were defined by the authors based on British light-timber-frame prefabrication practice. Unit costs were derived from BCIS (Building Cost Information Service) material supply rates [29]; embodied carbon factors from the ICE v3.0 (Inventory of Carbon and Energy) database [30]; and material densities from Eurocode 1 (EN 1991-1-1) [31]. The database excludes manufacturing costs, labour, installation and preliminary costs. This scope is consistent with the comparative purpose of the prototype: all 60 configurations are evaluated using the same methodology, so the relative performance differences between solutions are valid for early-stage decision-making, even though they do not reflect absolute project-level accuracy. To support full transparency and replicability, the complete component database is provided as Supplementary Material. The accompanying file lists all 47 components with their family, function, dimensional values, variable-dimension ranges for INFILL and CLOSURE pieces, the material composition and layer thicknesses of each component, and the derived weight, embodied carbon, cost, Interface_Score, Standardisation_Score, Assembly_Factor and Transport_Factor_Panel values. The derivation rules are those stated above: cost from BCIS material supply rates [29], embodied carbon from the ICE v3.0 database [30] and densities from Eurocode 1 [31], with the composite indices assigned on the ordinal 0 to 1 scale described in Section 4.3. This allows the full evaluation basis for every configuration to be reproduced.

From an operational perspective, this five-sheet organisation separates raw material data (Material_Performance) from component geometry and composition (Component_Library, Components DATABASE), hierarchical classification (Component_WBS_Map) and optimisation-ready performance summaries (Component_Performance). Each layer can be updated independently without affecting the others. The workbook is loaded once at the start of the Dynamo graph execution and stored as a shared Python (Dynamo Core 3.5) dictionary (db) between nodes, avoiding repeated file access during the 10,000 evaluations (50 individuals × 200 generations).

4.2. System Architecture

Building on the existing BIM model and the component database described above, the prototype organises its computational workflow through a modular architecture implemented in Dynamo. The system comprises four principal Python nodes: (1) Wall Extractor, which extracts the geometry and parameters of walls from the BIM model; (2) WBS Engine, which decomposes walls, floors and roof into constructive subsystems using modular division rules and queries the component database; (3) Scenario Evaluator, which defines alternative component configurations; and (4) MOO Engine, which executes the multi-objective optimisation process using the NSGA-II algorithm.

Figure 2 shows the general system framework, structured into five stages: BIM model input, WBS decomposition, database query, scenario evaluation and multi-objective optimisation. This structure enables an existing architectural BIM model to be transformed into a set of alternative prefabricated configurations without modifying the original model geometry. The detailed implementation of the system in Dynamo is shown in Figure 3, where the interaction between the different nodes and the functional separation between geometric processing, configuration generation and optimisation can be observed.

The five-stage architecture in Figure 2 represents the data flow from BIM model input to optimisation output. The optimisation itself is iterative rather than single-pass. Within the MOO Engine, NSGA-II repeats evaluation, non-dominated sorting, selection, crossover and mutation across 200 generations, so each candidate configuration is assessed and refined repeatedly before the Pareto front is returned. The iteration occurs inside the optimisation stage and operates on the discrete configuration space generated by the WBS Engine. The framework does not feed results back into the BIM geometry, because the model geometry is held fixed as the design input and only the component configuration is varied. If the designer modifies the BIM model, the WBS decomposition regenerates the configuration space automatically and the optimisation is repeated, which constitutes an outer design iteration loop at the discretion of the design team.

The system execution workflow follows a defined sequence. First, the Wall Extractor reads the geometry of the BIM model and extracts parameters such as length, height, thickness and element type for each selected component. This information constitutes the geometric basis upon which the rest of the system operates.

Then the WBS Engine uses this data to decompose each element into prefabricated components using modular assignment rules. This node constitutes the core of the system, transforming the geometry of the model into a discrete set of constructive elements. For each component, an integer division is applied according to predefined module sizes (3600 mm, 1200 mm and 600 mm), assigning panels in priority order. Where element height exceeds the standard size, CLOSURE components are incorporated as vertical closing elements, whilst residual lengths are resolved using INFILL components as horizontal closing elements within the ranges defined in the database. For example, an external wall of 25,800 mm × 4000 mm with configuration [3600; 1200; 600] mm receives 7 × PF-EXT-3600-3500-440, 7 × PF-CLOSURE-500-440 (3600 mm module), 1 × PF-EXT-600-3500-440 and 1 × PF-CLOSURE-500-440 (600 mm module), 16 units in total.

The modular division rule applied by the WBS Engine follows a deterministic greedy assignment, summarised in the pseudocode below (Algorithm 1).

Algorithm 1: WBS Engine Modular Division

Input: element length L, element height H, ordered module list M = [3600, 1200, 600], standard panel height H_std, database db Output: list of assigned prefabricated components 1: remaining = L 2: components = empty list 3: for each module m in M (largest to smallest): 4:    count = floor(remaining/m) 5:    if count > 0: 6:      append count standard panels of size m from db 7:      remaining = remaining minus (count times m) 8: if remaining > 0 and remaining within [INFILL_min, INFILL_max]: 9:    append one variable INFILL component of length = remaining 10: if H > H_std: 11:    for each assigned panel: 12:      append one CLOSURE component sized to (H minus H_std) 13: return components

This rule is applied independently to every wall, floor and roof element extracted from the BIM model. The greedy assignment from largest to smallest module guarantees the minimum standard panel count for a given module set, while the INFILL and CLOSURE logic resolves residual lengths and heights that are not exact multiples of the available modules.

Once the WBS-based decomposition has been generated, the Scenario Evaluator constructs component configurations representing different modularisation strategies. These configurations are used as initial solutions for the optimisation process, as shown in

Table 1. NSGA-II seed scenarios.

	Walls (mm)	Floors (mm)	Roof (mm)	Strategy
S1	3600; 1200; 600	3600; 1200	4800	Minimum component count, lowest cost and carbon
S2	1200; 600	6000; 3600; 1200	3600; 1200	Balance between cost, carbon and assembly factor
S3	600	1200	1200	Maximum assembly flexibility, highest component count

The three predefined scenarios define the initial seed solutions and their module configurations.

.

Finally, the MOO Engine evaluates these configurations using the NSGA-II algorithm, generating a set of non-dominated solutions that form the Pareto front. Each solution is evaluated against four objectives: cost, embodied carbon, assembly and logistics. The system therefore returns not a single optimal solution, but a set of comparable configurations that enable visualisation of the trade-offs between the different performance criteria.

In addition to numerical results, the system produces supplementary outputs designed to facilitate interpretation. These include comparative performance charts, an automated PDF report and a visualisation within the BIM environment via colour overrides associated with WBS levels (Figure 4). These outputs allow the optimisation results to be reviewed in both tabular and visual format within the BIM environment.

The system architecture establishes a clear separation between data extraction, WBS decomposition, configuration generation, optimisation and results visualisation. This separation allows each component of the workflow, for example, the database, modularisation rules or optimisation parameters, to be modified without affecting the rest of the system. Under controlled conditions, that is, using the same input data and system parameters, the prototype produces consistent results across runs, satisfying the DSR reproducibility criterion [10].

4.3. Multi-Objective Optimisation

This section describes the four optimisation objectives, the justification for the assembly factor, and the algorithm parameters. NSGA-II [19] implemented via pymoo 0.6.0 [32] in CPython 3.9 minimises the four objectives simultaneously.

Table 2. Prototype optimisation objectives.

Objective	Unit	Definition
f1—Total cost	£	∑ Cost × number_panels across all WBS levels and building elements.
f2—Embodied carbon	kgCO2e	∑ Embodied_Carbon_kgCO2e × number_panels across all WBS levels.
f3—Assembly factor	-	AF = 0.35·W_norm + 0.25·A_norm + 0.25·Interface_Score + 0.15·Standardisation_Score; W_norm and A_norm are weight and area normalised to [0, 1] within each constructive family.
f4—Lorries	trips	Physical load model: vehicle 13.6 m × 2.55 m × 4.0 m, 24 t payload; wall panels vertical, floor/roof panels stacked. Lorries = max (⌈V_total/V_vehicle⌉, ⌈W_total/24,000⌉).

Definition of the four optimisation objectives, units and formal expressions. Cost (f1): BCIS rates, manufacturing and labour excluded. Carbon (f2): ICE Database v3.0 factors. AF (f3): composite indicator defined by the authors; see Section 4.3. Components are bespoke designs for this research.

formally defines each objective.

It is important to note that the assembly factor (AF) was developed as a composite indicator to quantify the relative installation complexity. In the absence of a standardised metric in the literature, the AF was formulated as a weighted linear combination of four attributes. Panel weight received the highest weighting (0.35), in accordance with Masood et al. [23] and Xu et al. [24], who identify lifting equipment requirements as the primary determinant of the installation method [24,25]. Interface complexity and panel area were weighted equally (0.25 each) for their influence on connection time and spatial coordination on site [24,25]. Standardisation received the lowest weighting (0.15), as it operates as an enabling factor rather than a direct constraint [25]. Interface and standardisation scores were assigned per component using an ordinal 0–1 scale calibrated to British prefabrication practice, following the approach of Padala and Kumar [5]. The AF ranges from 0 to 1: lower values indicate simpler assembly.

The weight and area terms in the assembly factor are normalised within each constructive family rather than across the whole database, so that components are compared against peers of the same type. For each family f (external walls, internal walls, floor layers and roof layers), the normalised weight of component i is W_norm,i = (W_i minus W_min,f) divided by (W_max,f minus W_min,f), where W_min,f and W_max,f are the minimum and maximum component weights within family f. The normalised area A_norm,i is computed in the same way using component face area. This family-level normalisation maps both terms to the interval [0, 1] within each family, which prevents large external wall panels from dominating the assembly factor of small floor or roof components and keeps the interface and standardisation scores, already defined on a 0 to 1 ordinal scale, on a consistent footing. The composite assembly factor is then computed using the fixed weighting in Table 2, and the family-level results are combined at configuration level in proportion to the panel count of each family.

Algorithm parameters: Population size n = 50, generations g = 200, and total evaluations = 10,000. The initial population includes the three seed vectors (S1–S3) plus 47 randomly sampled configurations. Pareto solutions are ordered by decreasing wall module size (P01 = largest, P16 = smallest). The best compromise solution is identified by min–max normalisation over the five performance indicators (f1–f4 plus panel count):

Sk = ∑i [(fi(θk) − fi^mrn)/(fi^max − fi^mrn)]; θ* = argmink Sk

This aggregation function follows the standard min–max normalisation method documented in the multi-objective optimisation literature [19,32], where the solution with the minimum global score represents the best compromise across all evaluated criteria. NSGA-II was selected over alternative MOO algorithms for three reasons. First, its fast non-dominated sorting mechanism is computationally appropriate for the discrete, low-dimensional search space of this problem. Second, its crowding distance operator preserves solution diversity on the Pareto front. Third, pymoo [32] provides a well-documented and actively maintained implementation in Python that meets the DSR reproducibility requirements. With 50 individuals and 200 generations, each of the 60 configurations is evaluated on average approximately 167 times, providing robust statistical coverage of the search space.

In terms of computational cost, the complete workflow including BIM geometry extraction, WBS decomposition, the full NSGA-II run of 10,000 evaluations and the generation of the PDF report, the Bill of Materials and the BIM colour overrides execute within the Dynamo environment on a standard desktop workstation without specialised hardware. The dominant cost is the repeated database query during evaluation, which is mitigated by loading the workbook once into a shared Python dictionary at the start of execution, as described in Section 4.1, so that the 10,000 evaluations operate on in-memory data rather than repeated file access. The discrete, low-dimensional search space of 60 configurations keeps the optimisation tractable, and the run completes as a single batch process suitable for early-stage use. Precise runtime and memory profiling across hardware configurations was not recorded in this proof-of-concept study and is identified as part of the reproducibility work in Section 7.

The project outputs allow the simultaneous evaluation of multiple construction configurations derived from the WBS decomposition of the BIM model. These outputs include both quantitative results, such as the Pareto front and associated performance indicators, and visual representations that facilitate their interpretation within the BIM environment.

The following section presents the results obtained from the application of the prototype to the case study, focusing on the structure of the Pareto front, the identification of optimal solutions and the evaluation of the relationships between cost, embodied carbon, assembly complexity and transport logistics across the generated configurations.

5. Results

Running the prototype on the case study generates a set of non-dominated solutions that define the Pareto front of the problem. From these results, it is possible to analyse how the different construction configurations influence cost, embodied carbon, assembly complexity and transport logistics.

This section is structured into four parts. First, the Pareto front and the range of variation in performance indicators are described. This is followed by analysis of the front structure and the behaviour of the principal design variables. Subsequently, the best compromise solution is identified and its performance evaluated against reference configurations. Finally, the Bill of Materials associated with that solution is presented.

5.1. Pareto Front

The NSGA-II algorithm identified 16 non-dominated solutions within the search space of 60 configurations.

Table 3. Pareto front: 16 non-dominated solutions.

	Walls (mm)	Floors (mm)	Roof (mm)	Cost (£)	CO2e (kg)	AF	Lorries	Panels
P01	3600, 1200, 600	6000, 3600, 1200	4800	£149,498.75	126,541.88	0.3968	11	694
P02	3600, 1200, 600	6000, 1200	4800	£149,592.65	126,660.00	0.3548	11	784
P03	3600, 1200, 600	6000, 1200	1200	£150,056.64	127,052.98	0.3216	10	874
P04	3600, 1200, 600	4800, 1200	4800	£149,674.03	126,918.25	0.3259	11	862
P05	3600, 1200, 600	4800, 1200	3600, 1200	£149,828.69	127,049.24	0.3160	11	892
P06	3600, 1200, 600	4800, 1200	1200	£150,138.02	127,311.22	0.2981	11	952
P07	3600, 1200, 600	3600, 1200	4800	£149,655.25	126,738.75	0.3319	11	844
P08	3600, 1200, 600	3600, 1200	3600, 1200	£149,809.91	126,869.75	0.3216	11	874
P09	3600, 1200, 600	3600, 1200	1200	£150,119.24	127,131.73	0.3030	11	934
P10	3600, 1200, 600	1200	4800	£150,444.01	127,731.00	0.1900	10	1600
P11	3600, 1200, 600	1200	3600, 1200	£150,598.67	127,862.00	0.1871	10	1630
P12	3600, 1200, 600	1200	1200	£150,908.00	128,123.98	0.1816	10	1690
P13	3600, 600	1200	1200	£153,272.93	128,752.41	0.1804	11	1771
P14	1200, 600	1200	1200	£159,806.20	130,458.58	0.1774	11	1962
P15	600	1200	3600, 1200	£175,209.08	134,326.92	0.1758	13	2391
P16	600	1200	1200	£175,518.41	134,588.91	0.1722	13	2451

Pareto front ordered by decreasing wall module size. P01: minimum cost/carbon. P10: best global compromise (min–max normalisation). P16: minimum assembly factor. CO₂e in kgCO₂e; AF = Assembly Factor.

presents the complete front.

Total cost ranges from £149,498.75(P01) to £175,518.41 (P16), representing a difference of 17.4%. Embodied carbon shows a more moderate variation, between 126,541.88 and 134,588.91kgCO₂e (+6.4%). In contrast, the assembly factor exhibits a significantly greater variation, decreasing from 0.397 (P01) to 0.172 (P16), representing a reduction of 56.6%.

The number of lorries remains between 10 and 11 trips for P01–P14, increasing to 13 for P15–P16. This change occurs when the system adopts exclusively 600 mm wall modules, which increases total weight until it exceeds the vehicle payload limit before reaching the volumetric limit. Figure 5 presents a visual comparison of the performance of the Pareto-optimal solutions, enabling identification of the variations between the different configurations.

5.2. Structure of the Pareto Front

Three groups of configurations emerge from the Pareto front. P01–P12 share the wall configuration [3600; 1200; 600] mm (the basis of Scenario S1), with cost variation confined to ±1%, driven solely by floor and roof module choices. Within this group, P10–P12 adopt exclusively 1200 mm floor modules, which increases the panel count by 143.5% (from 694 to 1690) with a cost increase of only 0.9%, an asymmetry documented by Haie et al. [26] and consistent with the lower unit cost of smaller floor panels relative to their greater count. P13–P14 show progressive cost increases of 1.6% and 4.3% as wall modules are reduced. P15–P16 (exclusively 600 mm walls) incur the maximum cost (+17.4%), the maximum carbon (+6.4%) and two additional lorry trips.

The 17.4% figure is the descriptive range between the lowest cost solution (P01) and the highest cost solution (P16), not the output of a variance decomposition or regression model. It is interpreted here as the observed cost spread across the Pareto front attributable to module size choices, given that wall, floor and roof modules are the only variables that differ between solutions. Because the search space is discrete and the solutions differ in known module assignments, the contribution of each subsystem is read directly from the front structure rather than inferred statistically. A formal variance decomposition across a larger and more varied set of geometries is identified as future work in Section 7, and would be required to generalise the relative contributions of wall, floor and roof choices beyond this case study.

The discontinuity at P12/P13 reflects a manufacturing cost threshold: excluding the 1200 mm module from the wall system multiplies joints, increases installation labour and reduces production line efficiency, a threshold invisible to any single-objective analysis. The similar lorry count of 10/11 for P01–P14 indicates that the transport logistics constraint is determined primarily by the total building volume rather than by the degree of panel fragmentation. For P15–P16, the proliferation of 600 mm wall panels activates the weight constraint before the volumetric one, with implications for delivery scheduling and crane availability on site.

5.3. Best Compromise Solution: P10

Solution P10 (walls [3600; 1200; 600] mm, floors [1200] mm, roof [4800] mm) achieves £150,444.01, 127,731.00 kgCO₂e, AF = 0.190, 10 lorries and 1600 panels. The min–max normalisation score is the lowest on the Pareto front, identifying P10 as the best global compromise. P10 preserves large-format wall modularity (the cost and environmental advantages of P01) whilst adopting minimum floor modules (improving AF from 0.397 to 0.190, a reduction of 52.1%), without any increase in lorry count, and with a marginal reduction relative to P01. The 52.1% improvement in the assembly factor without a significant increase in cost or lorries represents the central finding of the Pareto analysis: floor module size drives the assembly factor independently of cost and logistics. In conventional sequential single-criterion optimisation, this solution would not be discovered because the 130.5% increase in panel count would appear as a penalty rather than as an affordable trade-off.

Table 4. Performance comparison.

Indicator	P01 (Min. Cost)	P10 (Best Compromise)	P16 (Min. AF)	Change P01 → P16
Total cost (£)	149,498.75	150,444.01	175,518.41	+17.4%
Embodied carbon (kgCO2e)	126,541.88	127,731.00	134,588.91	+6.4%
Assembly factor (AF)	0.397	0.190	0.172	−56.6%
Lorries (trips)	11	10	13	+18.2%
No. of panels	694	1600	2451	+253.2%

Performance comparison for three key Pareto solutions. Percentage change calculated with respect to P01.

compares the three most informative Pareto solutions.

This configuration represents the best global compromise within the Pareto front, combining near-optimal economic and environmental performance with a substantial improvement in constructability, demonstrating that small variations in cost and carbon can enable significant improvements in assembly ease, improvements that would not be identifiable through single-criterion optimisation approaches.

5.4. Bill of Materials for P10

The Bill of Materials for P10 totals 1600 units distributed across eight WBS leaf categories (

Table 5. Bill of Materials.

WBS Code	Units	Component Types	Modules Used (mm)
1.2.1 Structural floor	422	PF-FLR-xxxx-1200-187	1200 exclusively
1.2.2 Floor insulation	422	PF-FLR-xxxx-1200-218	1200 exclusively
1.2.3 Floor finishes	422	PF-FLR-xxxx-1200-65	1200 exclusively
1.3.1 External walls	249	PF-EXT; PF-CLOSURE-500; PF-EXT-INFILL	600; 1200; 3600 + bespoke infill pieces
1.4.1 Internal walls	55	PF-INT; PF-CLOSURE; PF-INT-INFILL	600; 1200; 3600 bespoke infill pieces
1.5.1 Structural roof	10	PF-RF-4800-1200-245	4800 exclusively
1.5.2 Roof insulation	10	PF-RF-4800-1200-22	4800 exclusively
1.5.3 Roof finishes	10	PF-RF-4800-1200-50	4800 exclusively
Total	1600	-	-

Bill of Materials for solution P10, grouped by WBS level.

). The floor layers (WBS 1.2.1–1.2.3) each contain 422 units of 1200 mm modules. The roof (WBS 1.5.1–1.5.3) uses exclusively the 4800 mm module (10 units per layer). The external walls (WBS 1.3.1) present the greatest typological diversity: 39 standard PF-EXT-1200 panels, 76 PF-EXT-3600 panels, 21 PF-EXT-600 panels, 78 CLOSURE pieces of varying lengths, and 35 bespoke INFILL pieces covering 24 distinct residual lengths. The internal walls (WBS 1.4.1) comprise 30 standard panels, 21 CLOSURE and 4 INFILL pieces across three module sizes.

The prototype manages INFILL pieces in walls through the Is_Variable field and the Min/Max_Length_mm limits of Component_Library, inserting a custom-dimension INFILL panel where the residual length falls within the permitted range. Traceability between the BOM and the BIM model is ensured internally through the geom_metrics dictionary, which links each assigned component to the ID of the model element from which it originates. This linkage is not exported in the PDF report generated by the prototype but is available within the Dynamo graph environment: the user can consult it by inspecting the WBS Engine node outputs, or by extending the graph with additional nodes that export this information to a spreadsheet or to element parameters in the model. The WBS colour overrides applied visually to the model serve as a verification layer enabling the project team to confirm the correct classification of each element before using the BOM for planning or procurement.

6. Discussion

6.1. Impact of Wall Modularity on Cost and Embodied Carbon

The Pareto front reveals a clear hierarchy of design variables: wall module size is the dominant driver of cost variation, far exceeding the effect of equivalent changes in floor modules. Configurations using exclusively 600 mm wall modules, P15–P16, cost 17.4% more than P01, whilst the full variation in floor and roof modules across P01–P14 produces cost changes below 1%. This implies that DfMA decision-making effort should prioritise wall modularisation over floor and roof choices, a result consistent with Xiang et al. [27], Haie et al. [26] and Liu et al. [28] but expressed here quantitatively and derived directly from the BIM model geometry.

The threshold between P12 and P13 is particularly relevant from a design perspective. The exclusion of the 1200 mm module from the wall system produces a cost increase of £2364, indicating that small changes in the modularisation strategy can generate significant economic jumps when they force the system to rely on more closure pieces or less efficient component combinations. This type of threshold would not be identifiable in a single-criterion analysis focused on a single optimal solution: it only emerges when comparing solutions along the Pareto front. In practical terms, this implies that the Pareto front not only enables comparison of configurations but makes visible the thresholds at which certain wall modularisation decisions produce disproportionate economic jumps relative to the apparent changes in design. This capacity to reveal cost discontinuities linked to specific constructive decisions is a contribution of the multi-objective approach to early-stage DfMA decision-making.

The optimisation strategy in this study operates at the level of component configuration, selecting and combining prefabricated panels to balance cost, carbon, assembly and logistics. Material selection within those components offers a complementary route to performance improvement that the present framework does not address. In structural systems, the strategic assignment of different material grades to different components can improve global performance while reducing total material consumption. Shi, Huang and colleagues demonstrated this principle for steel structures through the triple grade hybrid high-performance steel structure (TGHSS), in which ordinary strength steel beams, high strength steel columns and low yield point steel braces are combined so that each component uses the grade best matched to its structural demand. The concept was established and validated through full-scale cyclic experiments [33,34] and subsequently extended through numerical analysis and design [35,36]. Although the case study presented here is a light-timber-frame system rather than a steel structure, the underlying principle is transferable. Extending the component database to vary material grade and specification within each prefabricated family, alongside module size, would allow the framework to capture material level trade-offs and their associated carbon reductions. This is a promising direction for broadening the optimisation beyond geometric modularisation.

6.2. The Pareto Front as a Decision Instrument

The Pareto front functions not only as an optimisation output but as a communication instrument for multidisciplinary teams. Architects, manufacturers and contractors typically apply different priority criteria; the front allows each stakeholder to identify their preferred region without privileging a single objective. The identification of P10 as the best global compromise demonstrates a characteristic of multi-objective approaches: the solution of greatest value for decision-making is rarely found at the individual extreme of any objective. This result supports the argument that sequential single-criterion analyses optimising cost, then carbon, then constructability, may lead to discarding configurations of high practical value.

The practical benefit of a Pareto-based approach is not merely computational, but also communicative: it transforms an implicit negotiation among stakeholders, where each discipline defends its own priority criteria, into a shared dialogue grounded in quantified trade-offs. In industrialised construction projects involving manufacturers, logistics managers, sustainability consultants, and clients, this shared representation is an indispensable requirement for coordinated decision-making in the early stages.

6.3. The WBS as a Generative Mechanism

The central methodological contribution of this work is the reinterpretation of the WBS as a generative design decomposition tool. In prior BIM research, the WBS has been used primarily to structure work packages, asset management or digital collaboration [8,9,37]. In the prototype presented here, the WBS decomposes the geometry of the existing BIM model into prefabricated component configurations before design decisions are finalised, producing the configuration space that feeds the optimisation algorithm. This reinterpretation is consistent with the broader transition of BIM from documentation platform to active decision-support environment [3,4], but represents an application that is largely unexplored in the reviewed literature.

The practical advantage over existing BIM-MOO methods [4,18] is that the search space adapts automatically to model geometry without manual parametrisation. Any change to the BIM model (wall length, floor depth or component catalogue) propagates automatically to the search space, making the system responsive to design iteration. Compared with Moradabadi et al. [4], the fundamental difference is that the search space is not defined manually but emerges from the WBS decomposition. If BIM models are habitually organised according to a WBS hierarchy consistent with prefabrication systems, as is increasingly required by international BIM information management standards [38], the WBS-based search space generation approach could be applied systematically without project-specific customisation.

The reinterpretation of the WBS as a generative mechanism raises the question of how the approach aligns with mainstream BIM information management practice. The WBS hierarchy used here is compatible in principle with the information management concepts of ISO 19650 [38], in that it organises model content into a structured, hierarchical breakdown that can map to information containers and work packages. The framework does not yet propose a standardised, industry-compliant WBS hierarchy for prefabricated construction, nor does it formalise the link between the generative WBS and conventional work packaging or information delivery protocols. For the approach to be adopted within professional collaborative BIM workflows, the generative WBS would need to be reconciled with the work breakdown conventions already used for planning and procurement, so that the same hierarchy serves both generative decomposition at the design stage and work packaging at the delivery stage. Establishing this alignment, including a candidate standardised hierarchy for panelised and modular systems, is a necessary step towards practical implementation and is identified as future work.

6.4. Limitations

Four limitations define the scope of the results. First, the case study has regular geometry with all floor and roof dimensions as multiples of 1200 mm, which eliminates infill panels in floors and roof and allows the Pareto front structure to be attributed exclusively to module size choices. The framework handles irregular geometry through the INFILL and CLOSURE component logic described in Section 4.2, which resolves residual lengths and non-standard heights at the element level. This capability has not been validated on a building with non-orthogonal angles or variable spans, where a higher proportion of INFILL and CLOSURE components would be generated and the Pareto front structure may differ. The performance of the framework under such conditions is therefore established in principle through the component logic but not yet demonstrated empirically, and is identified as a priority for validation in Section 7. Second, the AF weightings (0.35 weight, 0.25 area and interface, 0.15 standardisation) were defined by the authors from the literature on prefabrication, modularisation and component performance [23,24,25,26,27,28] and are not empirically calibrated against on-site installation time, labour cost or modular construction practice records. The factor therefore provides a comparative ranking between configurations rather than an absolute estimate of installation time or labour, and the relative ordering of solutions on the assembly objective should be read as indicative rather than predictive until calibration data are available. Third, the database consists of bespoke components designed for this research using BCIS rates for cost, ICE Database v3.0 for embodied carbon and Eurocode 1 for density; manufacturing, labour and installation are excluded. Results should not be used for cost planning or carbon reporting without supplementary analysis. Fourth, the prototype has been validated as a proof of concept in a single case study; external validity requires replication across different building typologies and scales. These limitations define the conditions under which the results should be interpreted. At the same time, they establish a clear agenda for future research aimed at validating the practical applicability of the framework.

6.5. Positioning Relative to Prior Work

The prototype occupies a differentiated position relative to the closest technical precedents. Moradabadi et al. [4] demonstrated NSGA-II within Dynamo for lifecycle cost optimisation of commercial buildings; the fundamental difference is that their search space was defined manually and externally to the model, whereas in the prototype presented here it emerges automatically from the WBS decomposition. This distinction determines whether the tool can be used in early schematic design (where the model is still evolving) or only in detailed design (where parameters are fixed). Qi and Costin [16] provided the most comprehensive ontological framework for BIM-DfMA integration with rigorous semantic modelling; the present prototype is less comprehensive ontologically but offers a functional implementation within standard BIM tools. Doan et al. [1], reviewing BIM-offsite construction integration between 2020 and 2025, identified the absence of automatic configuration generation as a persistent gap; the prototype addresses this directly with a functional implementation.

7. Conclusions

This paper has demonstrated three principal contributions. First, WBS decomposition of an existing BIM model can automatically generate a discrete construction configuration space for multi-objective optimisation, without manual parametrisation, an application not identified in the reviewed literature. Second, a single Dynamo–Revit workflow integrating WBS decomposition, component database queries and NSGA-II optimisation is technically feasible using standard BIM tools and the open-source pymoo library, without additional external platforms beyond the BIM environment used. Third, the resulting Pareto front provides quantitative evidence that wall module size is the primary determinant of cost and carbon variation (with a range of 17.4%) in prefabricated construction, and that the best compromise solution occupies an intermediate position on the front rather than the individual extreme of any objective. These three contributions are interdependent: it is the automatic generation of the search space from the WBS that enables the Pareto front to communicate trade-offs directly linked to constructive decisions in the BIM model, converting the architectural model into an active basis for early-stage scenario comparison.

These findings have direct implications for early-stage DfMA practice. First, they justify prioritising wall modularisation decisions over floor and roof choices, given that the wall module accounts for 17.4% of cost variation compared with less than 1% attributable to floors. Second, they demonstrate that the identification of cost thresholds (such as that observed between P12 and P13) requires the simultaneous comparison of multiple solutions, which conventional single-criterion approaches do not permit. Third, the Pareto front enables teams with different priority criteria (architects, manufacturers, contractors) to identify their preferred configurations within a space of quantified trade-offs, without the need for implicit negotiations or sequential iterations between disciplines. The methodology further demonstrates that reinterpreting the WBS as a generative mechanism (rather than a post-design planning tool) is a conceptual shift with concrete computational consequences: the search space adapts automatically to any change in model geometry, making the system useful during the phases of greatest design iteration.

Nevertheless, the results should be interpreted within the limits of the research design. The case study has regular geometry with dimensions that are multiples of 1200 mm, a condition that eliminates infill panels in floors and roof and allows the structure of the Pareto front to be attributed exclusively to modularisation choices. Irregular geometries, with non-orthogonal angles or variable spans, may produce different front structures and will require the system to manage a greater number of INFILL and CLOSURE components. Similarly, the assembly factor (AF) weightings were defined based on the literature and are not empirically calibrated against real site scheduling data; the AF provides a comparative ranking between configurations but not an absolute estimate of time or labour. The component database, constructed using BCIS rates, ICE Database v3.0 and Eurocode 1, excludes manufacturing, labour and installation, so the absolute cost and carbon values are not directly transferable to cost plans or sustainability reports without supplementary analysis. Finally, the prototype has been validated as a proof of concept in a single residential case study; external validity requires replication across different typologies, scales and construction systems.

Future work should address these limitations along three lines. First, the empirical calibration of the AF using site scheduling data and real assembly time studies from light-timber-frame prefabrication projects in the British market, with the aim of transforming the indicator from a comparative ordinal scale to a metric with predictive validity. Second, validation of the framework on models with irregular geometry, multiple typologies, multi-family residential, educational, healthcare, and different building scales, to establish the generalisation limits of the WBS-based search space generation approach. Third, integration with verified manufacturer data through real-time API connections to digital catalogues, which would enable replacement of the bespoke component database with real contractual data and extend the utility of the prototype in professional contexts. Additionally, extension of the prototype to complementary prefabricated systems like MEP modules, ventilated façades, flat roof systems and the evaluation of NSGA-III for objective spaces of five or more criteria represent natural directions for development. Finally, evaluation of the impact of the prototype on real decision-making through studies with industry project teams would enable determination of whether Pareto front representation effectively modifies design decisions and under which organisational conditions it proves most useful. In addition, future work will record detailed runtime and memory profiling across hardware configurations to support full computational reproducibility, conduct a formal sensitivity analysis of the NSGA-II operator settings together with a comparison against NSGA-III and MOEA/D for higher-dimensional objective spaces, and undertake variance decomposition of the design variables across multiple geometries to quantify the independent contribution of wall, floor and roof module choices to each objective.