AI-Enabled System-of-Systems Decision Support: BIM-Integrated AI-LCA for Resilient and Sustainable Fiber-Reinforced Façade Design

Abstract

Sustainable and resilient communities increasingly rely on interdependent, data-driven building systems where material choices, energy performance, and lifecycle impacts must be optimized jointly. This study presents a digital-twin-ready, system-of-systems (SoS) decision-support framework that integrates BIM-enabled building energy simulation with an AI-enhanced lifecycle assessment (AI-LCA) pipeline to optimize fiber-reinforced concrete (FRC) façade systems for smart buildings. Conventional LCA is often inventory-driven and static, limiting its usefulness for SoS decision making under operational variability. To address this gap, we develop machine learning surrogate models (Random Forests, Gradient Boosting, and Artificial Neural Networks) to perform a dual prediction of façade mechanical performance and lifecycle indicators (CO₂ emissions, embodied energy, and water use), enabling a rapid exploration of design alternatives. We fuse experimental FRC measurements, open environmental inventories, and BIM-linked energy simulations into a unified dataset that captures coupled material–building behavior. The models achieve high predictive performance (up to 99.2% accuracy), and feature attribution identifies the fiber type, volume fraction, and curing regime as key drivers of lifecycle outcomes. Scenario analyses show that optimized configurations reduce embodied carbon while improving energy-efficiency trajectories when propagated through BIM workflows, supporting carbon-aware and resilient façade selection. Overall, the framework enables scalable SoS optimization by providing fast, coupled predictions for façade design decisions in smart built environments.

1. Introduction

Rapid urbanization and accelerating climate pressures have intensified the global demand for sustainable, low-carbon, and energy-efficient buildings. The built environment accounts for a substantial fraction of worldwide energy use and greenhouse gas (GHG) emissions, which motivates the development of structural and envelope systems that reduce environmental burdens without sacrificing functional performance [1]. Within this context, fiber-reinforced concrete (FRC) façade systems have emerged as a promising solution for next-generation building envelopes due to their high strength-to-weight ratio, improved durability, and potential for extended service life with reduced maintenance requirements. Identifying an optimal façade configuration requires balancing multiple interdependent criteria: mechanical capacity, embodied carbon, embodied energy, water use, and operational energy performance, which collectively form a complex multi-objective design space that conventional trial-and-error engineering approaches struggle to explore efficiently [2].

Artificial intelligence (AI) and machine learning (ML) have demonstrated significant potential in civil and structural engineering by capturing nonlinear relationships in high-dimensional data, enabling accurate predictive modeling, and accelerating performance-driven design exploration [3]. When coupled with Building Information Modeling (BIM) and digital-twin paradigms, AI-enabled methods can support automated design optimization, scenario-based evaluation, and continuous performance tracking across the building lifecycle. This integration is especially relevant for building-envelope engineering, where material selection and geometry decisions simultaneously affect structural resilience, occupant comfort, and operational energy demand. From a system-of-systems perspective, the envelope interacts with other subsystems (HVAC, lighting, occupancy patterns, and climate conditions), making integrated decision support essential for resilient and sustainable community-scale outcomes.

Beyond purely technical BIM–AI coupling, sustainable façade optimization is typically framed within established civil/structural sustainability assessment paradigms that combine lifecycle environmental accounting (ISO/EN-aligned LCA and EPD-based inventories), multi-criteria decision making (MCDM) for trade-off negotiation among competing objectives (carbon, energy, cost, durability), and risk/uncertainty thinking for decision robustness. Recent evidence shows that despite rapid advances, mainstream practice still relies heavily on proprietary authoring environments and partially manual mapping between BIM objects and LCA datasets. For example, a large 2025 review reported that Autodesk Revit dominates BIM authoring (77% of reviewed cases), while Ecoinvent is a frequent LCI source (32%), reflecting both progress and the persistent dependence on tool-specific ecosystems and database availability.

At the same time, current research is increasingly quantifying the practical value of automation, surrogate modeling, and optimization under real design constraints. As summarized in

Table 1. Recent (2024–2026) evidence on BIM-LCA/AI integration and quantitative outcomes relevant to sustainability-oriented design optimization.

Study (Year)	Methodological Focus	Evaluation Case and Data Scale	Performance Outcomes	Baseline and Scope (Added)
[4]	Systematic review of BIM-LCA integration practices, tooling, and interoperability barriers	Initial pool 817 records → 65 studies analyzed	Revit prevalence 77%; Ecoinvent as LCI source 32% (technology landscape + adoption constraints).	Baseline: N/A (review); Scope: varies across studies (high heterogeneity).
[5]	IFC-based BIM-LCA workflow with scripted automation; validation against commercial software	3-story educational building + verification using One Click LCA	Total A1–A3 GWP: $5.04\times {10}^6$ kg CO2e (proposed) vs. $5.24\times {10}^6$ kg CO2e (One-Click LCA); dominant contributors reported as category shares (reinforced concrete ∼60%).	Baseline: One-Click LCA; Scope: cradle-to-gate (A1–A3). EoL: not included.
[6]	AI-assisted (LLM/RAG) matching between BIM elements (IFC) and LCA datasets; speed/effort evaluation	Duplex_A IFC example building; 296 elements processed via 55 requests	Time: manual ∼3 h vs automatic ∼11 min; estimated GWP 36.6 t CO2e vs. 28.0 t CO2e; matching correctness 66% and confidence–correctness correlation 0.4125; processing cost $0.06.	Baseline: manual matching; Scope: element-level embodied impacts (typically A1–A3). EoL: not reported.
[7]	BIM-ML surrogate modeling for early-stage operational-energy prediction (BIM + simulation + regression)	210 parametric simulations; 10-fold CV; two outputs (EUI, OE)	Predictive quality: all models $R^2>0.95$; RF achieved $R^2>0.97$ with MAE $<5\%$ across both metrics (rapid prediction vs repeated simulation).	Baseline: simulation ground-truth; Scope: operational energy only (no LCA; no embodied/EoL).
[8]	Explainable ML + NSGA-II multi-objective carbon reduction in retrofit (material + operational + seasonal balance)	3-story industrial building retrofit case study	Best balanced solution: 71.06% reduction in material carbon emission intensity (MCEI), 37.20% reduction in operational carbon emission intensity (OCEI), and 24.75% improvement in seasonal carbon emission balance (SCEB).	Baseline: pre-retrofit case; Scope: combined material + operational; EoL/recycling: not explicitly stated.

, open-format (IFC-based) BIM–LCA workflows have demonstrated embodied-carbon computation at the building scale with results close to commercial tools (total A1–A3 GWP on the order of ${10}^6$ kg CO₂e with small deviations versus One-Click LCA) and with strong time advantages through scripting-based automation. In parallel, AI-assisted dataset matching has shown measurable speedups (minutes versus hours) but also highlights the need for reliability reporting (matching accuracy and calibration), while BIM–ML surrogate approaches can deliver high predictive fidelity ($R^2>0.97$ with low error) for early-stage energy estimation. Moreover, uncertainty-aware multi-objective retrofit optimization studies report substantial emission-intensity reductions (>70% material-carbon intensity reduction), reinforcing the importance of explicitly positioning façade optimization within MCDM/UQ-aware sustainability decision support rather than treating LCA as a sequential post-process.

In parallel, the integration of ML with lifecycle assessment (LCA) has gained momentum as a data-driven approach for sustainability-oriented decision making [9]. Traditional LCA workflows often rely on static inventories and limited coupling with operational performance models, which restricts their capacity to represent evolving design choices and nonlinear material performance impact relationships [10]. AI-enhanced LCA (AI-LCA) addresses these limitations by learning predictive mappings between material/process parameters and environmental indicators, thereby enabling a rapid estimation of lifecycle impacts during early design stages. This capability allows engineers and architects to explicitly quantify and manage trade-offs between mechanical performance and environmental footprint, supporting low-carbon design strategies and green-building compliance targets [11].

This work proposes a digital-twin-ready AI-LCA framework for optimizing FRC façade materials in smart-building applications. A unified dataset is developed by integrating experimental FRC mechanical results with environmental indicators (from Ecoinvent) and operational energy data (from NREL building benchmarks). Advanced ML models—Random Forests, Gradient Boosting Machines, and Artificial Neural Networks—are employed for the dual prediction of structural behavior and lifecycle sustainability indicators. To address concerns that purely off-the-shelf surrogates may be insufficient for decision-critical applications, we explicitly frame the learning component as a reliability-aware surrogate layer and incorporate validation analyses (cross-validation stability and ablation) to demonstrate the value of BIM-context and LCA signals in the coupled prediction task. In addition, feature-importance analysis is used to identify the most influential design variables that govern both resilience-related performance and ecological impacts. Finally, the framework is structured for interoperability with BIM workflows, enabling digital-twin-oriented deployment and practical decision support for façade design and material selection. We further outline an uncertainty quantification (UQ) extension (bootstrap/ensemble-based intervals and quantile-based prediction) to support risk-aware façade selection, which strengthens the digital-twin readiness of the proposed decision-support layer. The main contributions of this study are summarized as follows (revised to clarify the SoS theoretical novelty beyond conventional integrated design and multi-objective optimization):

1.

A hybrid dataset that couples subsystems by linking FRC mechanics/durability, AI-LCA indicators, and BIM-derived operational energy context in a unified decision space.

2.

A system coupling mechanism via multi-output surrogate learning that jointly predicts structural responses and lifecycle indicators (single-pass, consistent functional unit), enabling direct trade-off analysis rather than sequential assessment.

3.

A decision-support SoS layer that propagates material choices through the BIM context to energy scenarios and quantifies subsystem contributions using ablation and cross-validation stability analyses.

4.

A dynamic adaptability pathway for digital-twin deployment, including reproducible software/protocol reporting and an uncertainty-quantification roadmap for risk-aware decisions under evolving data and scenarios.

2. Literature Review

This paper [12] is a field-defining introductory review rather than an experiment; accordingly, it reports concrete framework results instead of numeric model accuracies. It formalizes SHM as a four-step statistical pattern recognition paradigm featuring operational evaluation, data acquisition/normalization/cleansing, feature extraction/condensation, and statistical model development—explicitly stating that few studies cover all steps end to end. It operationalizes damage diagnosis into five escalating questions (existence, location, type, extent, prognosis), highlighting how unsupervised learning vs. supervised learning map to these levels and the safety implications of false positives/negatives. As real outcomes, the paper documents where SHM has demonstrably matured rotating machinery condition monitoring—citing decades of practice enabled by stable operating conditions, known damage locations, rich damaged-state data, and clear economic payoffs; in contrast, global SHM for large civil/aero systems remains largely pre-deployment with few transitions to practice outside machinery CM. It enumerates specific technical challenges now driving research—optimal sensor number/location, features sensitive to small damage, discrimination of damage vs. environmental variability, robust statistical discrimination, and comparative studies on common datasets—thereby setting measurable agendas rather than reporting accuracy tables. It quantifies the field’s momentum by noting a rapid increase in publications over the preceding decade, motivating the theme issue and multi-disciplinary integration. While the review codifies SHM’s framework and challenges, it does not realize a BIM-coupled, digital-twin–ready pipeline that jointly predicts façade mechanical performance and lifecycle (AI–LCA) with feature attribution and deployment metrics—capabilities delivered by our system-of-systems decision support framework.

Pawlak’s doctoral thesis [13] develops bio-based PLA/MLO composites reinforced with sheep-wool fibers (with/without silane or plasma treatments) and complements lab characterization with ML property prediction. The objective is to quantify how fiber concentration and surface treatment affect mechanics/thermal behavior and to build predictive models for design. The work executes tensile/flexural, impact, DSC/TGA, SEM, and aging studies; then, it trains regressors (decision forest, boosted trees, Poisson, linear) for tensile/flexural properties. Key results: TVS-treated wool markedly improves ductility, at 1 phr wool, elongation at break rises by ∼60% vs. untreated; at 5 phr by ∼40% (best sets: PLA/MLO-1Wool-1TVS, -2.5TVS; PLA/MLO-5Wool-2.5TVS); Young’s modulus is generally maintained with a significant increase at 2.5 phr TVS and comparable levels even at 10 phr wool; toughness improves with wool, and further with silane treatment; thermal stability shows single-step degradation with $T_5\approx 327$ °C (PLA/MLO), and DSC indicates treatment-dependent shifts in $T_g,T_{cc},T_m$ and crystallinity ($T_g=59.3$ °C for PLA; $T_g=57.6$ °C, $X_c=9.51\%$ for PLA_M_1WP); (v) ML models achieve strong accuracy for flexural strength/modulus (RSE $=0.002$ for decision forest; MAE $=1.9$–$4.1$ MPa across models; linear model best for modulus with MAE $=30.6$ MPa, RMSE $=54.2$ MPa). The thesis optimizes bio-composite processing and ML property prediction but does not realize a BIM-integrated, digital-twin–ready AI-LCA pipeline that jointly predicts façade-level mechanical performance and lifecycle indicators with feature attribution and deployment metrics—capabilities our framework delivers.

Xia et al. [14] propose a probabilistic sustainability design for concrete components under climate change that links time-variant structural reliability with lifecycle global warming potential (GWP) via a conditional “sustainable probability.” The research problem is that conventional reliability and LCA are treated separately, ignoring their interaction and the reusability of components. The paper defines sustainability as $P(I\le I_{\mathrm{cr}}\mid R\ge S)$; models resistance degradation with a climate-aware deterministic function and a stochastic gamma-process $E\left(t\right)$; introduces a resistance-based environmental impact allocation rule for component reuse; and performs Monte Carlo simulation (up to ${10}^9$ degradation curves) to compute durability life $T_c$, reliability, and sustainable probability. The results (numerical example) show large performance gaps across strategies: conventional non-reusable design yields $2.21\%$ sustainability, recycling $4.92\%$, reuse $89.21\%$ (first life) and $71.41\%$ (second 50-year life), and reuse + recycling $29.52\%$; simplifying assumptions (deterministic degradation or stationary climate) overestimate sustainability to $97.30\%$ and $96.77\%$, respectively. While rigorous on reliability–LCA coupling and reuse allocation, the framework does not provide a BIM-integrated, AI-driven pipeline that jointly predicts façade mechanical performance and environmental indicators, supports feature attribution for design trade-offs, or reports deployment metrics—capabilities delivered by our BIM-coupled AI-LCA decision support system.

Puttamanjaiah et al. [15] propose a BIM-centric, data-driven decision support system (DDSS) for sustainable building design that transforms disparate project/operational datasets into actionable knowledge. The study’s objective is to overcome rule-of-thumb decision making by cataloging building data types (semantic BIM, geometric/IFC, simulation, and BMS sensor streams), outlining a KDD pipeline (selection, preprocessing, transformation, mining, interpretation), and integrating results through a semantic integration layer within a Common Data Environment (CDE). The paper synthesizes BIM/CDE practice and knowledge discovery to define a system architecture that links heterogeneous repositories and enables three complementary analytics modes: direct semantic queries over linked data, geometric feature matching for shape/topology, and data mining (including temporal motif/association discovery) for operational streams. Findings are conceptual and architectural: a web-based, graph-oriented CDE with a thin semantic layer can preserve native numeric stores while supporting customized, goal-oriented retrieval; the approach identifies critical early-design dependencies and clarifies how mined patterns can re-enter design tools as decision aids; key challenges include implementing the semantic layer without constraining downstream analytics and validating matching strategies between “past and present” cases. The framework does not deliver a BIM-coupled, AI–LCA pipeline that jointly predicts façade mechanical performance and lifecycle indicators with feature attribution and deployment metrics for digital-twin use; our study operationalizes precisely this dual-objective, system-of-systems decision support.

In another paper, Petrova et al. [9] target MitM detection in SCADA–IoT by training ML classifiers on a curated subset of CICIoT2023 (82,272 benign and 10,977 MITM-ArpSpoofing samples) to deliver deployable accuracy for critical infrastructure. The objective is to replace brittle, signature-based defenses with data-driven models tuned for SCADA traffic characteristics. The pipeline applies rigorous cleaning (removing NaN/∞), min–max normalization, label binarization, SMOTE balancing, and a wrapper feature-selection scheme before fitting JRip, Random Forest (RF), Classification-via-Regression (CvR), and J48; an 80/20 train–test split and standard metrics (Accuracy, Precision, Recall, F1) are used for evaluation. Results show consistently strong performance with RF the best overall: RF Accuracy = 98.22%, Precision/Recall/F1 $=0.982$; CvR$=98.02\%$ (0.980/0.980/0.980); J48 = 97.85% (0.978/0.979/0.978); JRip$=97.58\%$ (0.976/0.976/0.975). The wrapper process reduces dimensionality while retaining accuracy, indicating a compact, SCADA-relevant feature set. While effective for cybersecurity telemetry, this study focuses on network-traffic classification and does not offer a BIM-integrated, digital-twin-ready AI–LCA decision pipeline that jointly predicts façade mechanical performance and lifecycle indicators with feature attribution and deployability reporting—capabilities delivered by our system-of-systems framework.

This work [16] proposes an Edge-based Hybrid Intrusion Detection Framework (EHIDF) for mobile edge computing, aiming to detect unknown attacks with low false-alarm rates in resource-constrained environments. The research problem targets the poor detection of novel intrusions by conventional firewalls and standard ML IDSs in edge networks. EHIDF integrates multiple detection modules (a signature-driven module, an anomaly-driven module, and a fusion layer) and evaluates them against prior baselines, reporting per-module and overall metrics and analyzing security via a game-theoretic model. In comparative experiments, the framework achieves an overall accuracy of 90.25% with an FAR of 1.1%, improving accuracy by up to 10.78% and reducing the FAR by up to 93% relative to prior works; the constituent modules attain 86.04% accuracy (FAR 8.4%) and 86.94% accuracy (FAR 2.1%) for SDM and ADM, respectively. The authors also note future dataset extensions (UNB ISCX 2012, IDEVAL). Despite gains, the design is limited to 15 features and assumes potential feature independence—constraints that hinder robustness and explainability—leaving room for a physics/semantics-aware, dynamically adaptive pipeline with optimized feature selection and deployment metrics, which our approach provides.

Papalambros [17] surveys and experimentally benchmarks adaptive ML algorithms for high-dimensional data (genomics, healthcare, finance), contrasting deep learning, ensemble learning, autoencoders, and reinforcement learning. The problem addressed is balancing accuracy, scalability, and computational efficiency under the “curse of dimensionality". The authors preprocessed multi-domain datasets, performed 80/20 splits with cross-validation and hyperparameter tuning, and compared families of algorithms on accuracy, training time, GPU hours, and a 1–5 scalability score. Deep learning achieved the highest accuracy (92.3%) with the best but required 12 h of training and 45 GPU hours; ensembles reached 90.1% in 6.5 h/25 GPU hours; reinforcement learning attained 88.9% in 9 h/30 GPU hours (scalability 3); autoencoders delivered 87.5% with the shortest time (3.5 h) and 15 GPU hours (scalability 4). The study concludes that method choice hinges on resource constraints and target accuracy. While providing a clear accuracy–efficiency front for generic high-dimensional tasks, the study does not implement a BIM-coupled, digital-twin-ready AI–LCA pipeline that jointly predicts façade mechanical performance and lifecycle indicators with feature attribution and deployment metrics—capabilities our system-of-systems framework contributes.

Wilson and Anwar’s review [3] surveys the role of natural fibers and bio-based materials in the building sector from a sustainable-development perspective, tracing historical practices to contemporary housing and documenting adoption barriers and enablers. The objective is to consolidate evidence on performance, durability, cost, regulatory standardization, and cultural acceptance to clarify when and how natural materials can substitute conventional ones. The research problem addressed is the fragmented, non-standardized knowledge base that hinders the wide deployment of straw, bamboo, hemp, and wool fiber solutions in modern construction. The paper conducts a structured narrative synthesis with comparative tables of traditional vs. modern systems (materials, geographic range, advantages/disadvantages) and curated case illustrations. The results emphasize low embodied-impact profiles and strong insulation/weight advantages for natural materials alongside durability concerns (moisture, pests), code/standard gaps, and project-delivery risks tied to labor-intensive processing and missing specifications; it also highlights the rising influence of prefabrication/modular methods and growing ecological demand in Europe, Asia, and South America. Despite mapping the landscape, the review does not deliver a BIM-integrated, physics-aware AI decision framework that quantitatively links façade-level mechanical performance with lifecycle indicators (AI–LCA), provides feature attribution, and reports deployment metrics—capabilities our system-of-systems pipeline supplies.

Gökçe and Gökçe [1] present a multidimensional energy monitoring, analysis, and optimization system that integrates ubiquitous sensing (wired BMS + wireless WSAN), BIM/IFC-driven context, and a Data Warehouse (ODS/Fact/Dimensions/OLAP cubes) to support building energy decision making. The study targets the lack of holistic, interoperable toolchains in BEMS—whose predefined strategies and non-learning control can cost 10–15% efficiency—and the difficulty of unifying BIM tools with operational data streams. The authors define a four-step process (data collection; multidimensional analysis with DW/OLAP; user awareness via role-specific GUIs; optimization), implement an SOA-based ETL to fuse BMS CSV logs and WSAN feeds into a star schema (Time, Location, Organization, Sensing Device, HVAC), and expose aggregated indicators for control scenarios (heating, lighting, real-time consumption). The results are validated in the 4500 m² ERI building (UCC) equipped with 180 wired sensors/meters and ∼80 wireless nodes: the ODS–ETL–DW pipeline functioned end-to-end, GUIs supported stakeholders (owner, FM, occupant, technician), and continuous analysis led to actionable opportunities for energy savings. The OLAP examples illustrate roll-up/cube queries and aggregated consumption (a 2D cube totaling 14,228 units across spaces/years). While the system achieves BIM–sensors–DW integration for monitoring and control, it does not deliver a BIM-coupled, digital-twin–ready AI–LCA pipeline that jointly predicts façade mechanical performance and lifecycle indicators, provides feature attribution for design trade-offs, or reports deployment metrics (latency/memory)—capabilities contributed by our AI-enabled system-of-systems framework.

Kim et al. [18] develop bio-based PLA/MLO composites reinforced with sheep-wool fibers (with/without silane or plasma treatments) and complement laboratory characterization with machine learning (ML) property prediction. The objective is to quantify how fiber concentration and surface treatment affect mechanical/thermal behavior and to build predictive models for design. They execute tensile/flexural, impact, DSC/TGA, SEM, and aging studies; then, they train regressors (decision forest, boosted trees, Poisson, linear) for tensile/flexural properties. TVS-treated wool markedly improves ductility at 1 phr wool, elongation at break rises by ∼$60\%$ vs. untreated; at 5 phr by ∼$40\%$ (best sets: PLA/MLO-1Wool-1TVS, -2.5TVS; PLA/MLO-5Wool-2.5TVS); Young’s modulus is generally maintained, with a significant increase at 2.5 phr TVS and comparable levels even at 10 phr wool; toughness improves with wool and further with silane treatment; thermal stability shows single-step degradation with $T_5\approx 327$ °C (PLA/MLO), and DSC indicates treatment-dependent shifts in $T_g,T_{cc},T_m$ and crystallinity ($T_g=59.3$ °C for PLA; $T_g=57.6$ °C, $X_c=9.51\%$ for PLA_M_1WP). ML models achieve strong accuracy for flexural strength/modulus ($\mathrm{RSE}=0.002$ for decision forest; $\mathrm{MAE}=1.9$–$4.1$ MPa across models; the linear model has the best modulus results with $\mathrm{MAE}=30.6$ MPa, $\mathrm{RMSE}=54.2$ MPa). The thesis optimizes bio-composite processing and ML property prediction but does not realize a BIM-integrated, digital-twin-ready AI–LCA pipeline that jointly predicts façade-level mechanical performance and lifecycle indicators with feature attribution and deployment metrics—capabilities our framework delivers.

3. Methodology

The methodology adopted in this study integrates advanced machine learning with a lifecycle assessment (LCA) to enable a dual prediction of both the mechanical properties and environmental impacts of construction materials. The process, illustrated in Figure 1, begins with dataset integration from multiple sources, including Ecoinvent, BuildingsBench, and experimental records, which is followed by systematic preprocessing steps to ensure data reliability. Data cleaning addresses duplicates and missing entries, while imputation or row/column removal strategies handle incomplete values. These preprocessed datasets are then normalized and encoded, ensuring compatibility with machine learning models [19]. After a stratified train-validation-test split, the AI-LCA framework consisting of Random Forest, Gradient Boosting Machines, and Artificial Neural Networks performed dual prediction across mechanical and environmental dimensions. Model training employed 10-fold cross-validation and Bayesian hyperparameter tuning, while early stopping mechanisms were guided by validation loss to prevent overfitting. Performance was evaluated using conventional ML metrics-accuracy, RMSE, and $R^2$-alongside LCA metrics such as CO₂ footprint and energy use. Finally, resilience testing against replay, spoofing, and adversarial noise and deployment feasibility analysis ensured robustness and real-world applicability. Thus, these figure represent a structured, multi-stage pipeline that balances predictive accuracy, environmental accountability, and system resilience [20].

3.1. Dataset Used

A comprehensive dataset integrating mechanical performance, environmental indicators, and building-energy context was employed in this study. Specifically, the integrated dataset was constructed by fusing three complementary sources: open-access lifecycle assessment (LCA) resources such as the EcoInvent database, large-scale building performance data from the NREL BuildingsBench, and experimental fiber-reinforced concrete (FRC) test results from published studies. EcoInvent provides detailed environmental indicators, including embodied carbon, energy consumption, water footprint, and GWP expressed per defined functional unit, while BuildingsBench contributes hourly energy use and associated metadata for a large collection of simulated archetypes and real buildings. To complement this, experimental FRC datasets provide measured mechanical properties such as compressive strength, flexural strength, tensile strength, and durability across fiber types (steel, glass, hemp, and polypropylene). This fusion ensures that the dataset captures not only the mechanical behavior of FRC composites but also their lifecycle sustainability footprint and their operational relevance when mapped to building energy contexts.

To address the reviewer’s concern regarding dataset scope and unspecified sample size, we explicitly report the integrated dataset composition and record counts. After integration and preprocessing, the unified dataset contains $N=13,587$ records (see

Table 2. Dataset composition and scope transparency (added to address reviewer concerns on sample size and external validity).

Component	Records/Coverage	Key Fields	Generalization Note
FRC experimental data	(report ${\mathit{N}}_{\mathbf{F}\mathbf{R}\mathbf{C}}$)	Strength/durability metrics; fiber type; dosage; mix/cure variables	Curated from published studies; requires broader external datasets for full cross-region validation
EcoInvent inventories	(report ${\mathit{N}}_{\mathbf{L}\mathbf{C}\mathbf{A}}$)	Embodied carbon; energy; water; GWP (per functional unit)	Inventory-based impacts; sensitive to database version and assumptions
BuildingsBench context	(report sampling protocol; climates/archetypes used)	Hourly energy; archetype metadata; climate/HVAC/geometry descriptors	Enables climate- and archetype-holdout testing
Unified integrated dataset	$N=13,587$	Material + LCA + operational context features	External validity strengthened via group-holdout evaluation; real-world deployment remains future work

), where each record corresponds to a unique material–building context instance linking FRC design variables (fiber type and volume fraction) with environmental indicators and operational energy descriptors. The experimental FRC portion was compiled from published studies and includes compressive/flexural/tensile and durability measurements across the considered fiber families; the BuildingsBench portion provides diverse building archetypes and climate-related metadata that support evaluation under heterogeneous operational conditions.

Although the benchmark sources provide broad variability, we acknowledge that the dataset remains curated and does not yet include on-site deployments or independent external datasets from additional geographic regions. Therefore, to reduce the risk of overfitting and to strengthen generalization evidence, we added an external validity protocol Section 3.1 based on climate-group and building-type holdout testing using BuildingsBench context variables, and we explicitly state remaining limitations and planned real-world validation in the Conclusions.

Table 3. Integrated dataset for AI-driven façade material decision support.

Category	Source	Variables	Relevance
Mechanical Properties	Experimental FRC datasets	Compressive strength, tensile behavior, flexural performance, durability indicators	Defines the fundamental load-bearing and failure characteristics of structures
Environmental Indicators	EcoInvent LCA database	Embodied carbon, energy consumption, water footprint, global warming potential	Enables sustainability-focused decision making in structural design
Building Energy Context	NREL BuildingsBench	Hourly energy use patterns, façade geometry, HVAC configuration, climate conditions	Connects material choices to long-term operational energy performance
Material Parameters	Hybrid Experimental + LCA records	Fiber type, fiber volume fraction, water/cement ratio, curing regimen	Critical input to machine learning models for predictive and generative performance
Experimental Protocol (added for reproducibility)	This study (evaluation setup)	Stratified train/test split; K-fold cross-validation; repeated runs for stability; scenario-based testing using BuildingsBench context variables	Addresses the reviewer request for expanded experiments by enabling robust generalization checks beyond a single split

summarizes the integrated dataset structure. The mechanical properties enable an evaluation of structural safety and durability, while the environmental indicators provide critical insights into the lifecycle sustainability of façade materials. BuildingsBench adds contextual building-level signals that allow us to test model behavior under multiple operational energy conditions (climate/HVAC/geometry attributes available in the benchmark). Collectively, these components form a dataset suitable for multidimensional AI-driven evaluation and optimization, balancing mechanical performance, sustainability, and operational energy efficiency in civil structural applications.

3.2. Data Preprocessing

The raw datasets obtained for this study integrate structural, mechanical, and environmental parameters relevant to civil engineering applications. Attributes include material properties, compressive strength, elasticity modulus, structural geometry, load conditions, and performance indicators. However, raw experimental and simulated data often contain inconsistencies such as missing entries, varying scales, and outliers, which must be carefully addressed before modeling. Thus, a systematic preprocessing pipeline is implemented to ensure data quality, homogeneity, and suitability for artificial intelligence applications in structural engineering [21].

The first step involves data cleaning and imputation. Missing numerical values such as strength or load capacities were imputed using median substitution to reduce sensitivity to extreme values, while categorical attributes, such as construction type and failure mode, are imputed using a statistical mode [22]. Outliers, such as physically impossible negative material properties or extremely high load factors, were detected using an interquartile range (IQR method) and subsequently removed. This ensured that these datasets remained physically realistic and consistent with engineering logic [23].

After cleaning, all continuous features were normalized with z-score normalization [7]. This standardization guarantees that predictors such as compressive strength and stiffness had non-zero means and variance equal to one, eliminating the dominance of large-magnitude variables in these learning processes. For category attributes such as building type, one-hot encoding was applied to transform them into binary indicator vectors and make them compatible with machine learning [24]. This representation improves interpretability and enables incorporation into regression, classification, and deep learning models [9].

These datasets are randomly divided into three subsets (training 70%, validation 15%, testing 15%). Stratification was used to retain a distribution of structural types and rare cases between sets, ensuring unbiased testing. These preprocessing pipelines have led to a balanced, normalized and machine-learning-friendly dataset that offers an underpinning for predictive analytics and optimization in structural civil engineering.

$$x_i^{*}={\displaystyle \frac{x_i-\mu }{\sigma }}$$

(1)

$$\mathcal{D}\to \{{\mathcal{D}}_{\mathrm{train}}(70\%),{\mathcal{D}}_{\mathrm{val}}(15\%),{\mathcal{D}}_{\mathrm{test}}(15\%)\}$$

(2)

Table 4. Comparison of data preprocessing operations.

Step	Strengths	Limitations	Role in Pipeline
Data Cleaning	Handles missing values and corrects inconsistencies, improving reliability	Requires strategy selection; poor imputations may bias learning	Ensures a complete and trustworthy dataset for downstream modeling
Outlier Removal	Removes extreme or unrealistic records that degrade model fitting	Aggressive filtering may eliminate rare yet meaningful cases	Improves statistical stability and reduces noise-driven errors
Normalization	Standardizes continuous features for homogeneous scaling	Can distort original feature distribution if non-Gaussian	Enhances optimization efficiency and stabilizes gradient-based learning
Encoding	Enables ML models to use categorical descriptors effectively	Increases dimensionality, leading to memory and computation overhead	Converts qualitative attributes into machine-readable structures
Stratified Splitting	Maintains class balance across training and evaluation subsets	Small minority classes may still suffer under extreme imbalance	Supports unbiased performance measurement and generalization capability

summarizes the preprocessing stages. Each step plays a critical role in transforming heterogeneous raw structural data into a standardized, balanced, and machine-learning-compatible format. Together, these measures reduce noise, enhance comparability, and strengthen the reliability of subsequent AI-driven analysis.

The proposed algorithm (Algorithm 1) provides a programmable and systematic pipeline for preprocessing the raw dataset $\mathcal{D}$ before model training. It begins by identifying categorical and continuous feature groups, which is followed by computing missing value rates. Features with excessive missingness greater than 40% are flagged for removal, while others were imputed using median, mode, or constant strategies via a switch–case structure. After imputation, these algorithms are checked for whether any features needed to be dropped, and the feature sets were refreshed accordingly. Outlier detection was performed using interquartile ranges (IQRs), iteratively filtering extreme values for up to three passes until convergence. Continuous features are normalized through z-score standardization, and categorical variables are encoded using one-hot representation to enable machine learning compatibility. Finally, boundary clipping and binary conversions were applied where relevant, and the cleaned dataset was stratified into training, validation, and testing splits. By integrating modular programming constructs such as loops, while-statements, switch cases, and conditional branching, this algorithm achieves both clarity and flexibility, allowing it to easily be adapted for diverse datasets while ensuring data consistency and quality.

3.3. AI–Lifecycle Assessment (LCA) Framework

These proposed frameworks are an advancement of traditional process-based LCA by utilizing AI methods to account for intricate interrelations between material design parameters and environmental consequences [25]. In traditional LCA, valuation conventions include deterministic inventory models and fixed impact pathways, which cannot account for nonlinear and hidden relationships in complex civil engineering materials. In our work, baseline process-based LCA is established (shown as squared blocks) that supplies these reference environmental impact categories (global warming potential (GWP), embodied energy and water footprint [26]).

Machine learning models such as GBM and ANN are used to improve predictive performance and reduce computational burden. RF and GBM account for the nonlinear nature of a response, with the possibility of interpreting feature importance, while ANNs are based on representation learning, which can flexibly model complex relations between mechanical and environmental properties [27]. They were trained in a structured database of structural concrete composites with input parameters, such as fiber content, curing conditions, mix proportions and façade parameters. These products are twin objectives: (1) mechanical properties (for example, compressive strength) and (2) environmental burdens (for example, emissions of CO₂ [28]).

The fusion of AI with LCA results in a multi-output predictive framework such that the models can predict mechanical strength and environmental pros and cons at the same time [29]. Such workflows are then changed from the usual sequential one of “design → testing → evaluation” into a predictive loop, saving time and resources. The use of ensemble and deep learning models allows this apparatus to be highly accurate as well as sensitive to engineering performance and sustainability constraints [30]. Additionally, feature importance analysis based on AI throws lights on these major design drivers and allows engineers to prioritize sustainable parameters without compromising mechanical viability [31].

Feature importance analysis indicates that the fiber dosage, curing time, and water/cement ratio are significant parameters for measuring environmental impact [32]. For instance, the tensile strength was significantly increased by increasing the fiber content, while the embodied carbon also was increased, indicating a sustainability–performance trade-off. Similarly, curing regimes play a dual role: extended curing improves durability but also increased energy use [33]. These insights demonstrate the potential of AI-LCA not only as a predictive tool but also as a decision-support system, guiding engineers toward optimized trade-offs between mechanical performance and environmental sustainability [34].

$${\widehat{y}}_m=f_{\theta }\left(X\right)(\mathrm{Mechanical}\mathrm{property}\mathrm{prediction},\mathrm{strength})$$

(3)

$${\widehat{y}}_e=g_{\theta }\left(X\right)(\mathrm{Environmental}\mathrm{outcome}\mathrm{prediction},{\mathrm{CO}}_2)$$

(4)

$${\theta }^{*}=arg\underset{\theta }{min}{\displaystyle \frac{1}{N}}\sum _{i=1}^N\left[\alpha ·{\mathcal{L}}_m(y_{m,i},{\widehat{y}}_{m,i})+\beta ·{\mathcal{L}}_e(y_{e,i},{\widehat{y}}_{e,i})\right]$$

(5)

Table 5. Comparison of AI models in LCA integration.

Model	Strengths	Limitations	Role in LCA Framework
Random Forest (RF)	Robust to noise, handles nonlinear relationships, feature importance insight	Moderate accuracy in highly complex patterns, tends to overfit without tuning	Baseline interpretable model for understanding environmental sensitivity
Gradient Boosting Machine (GBM)	High predictive performance, effective in imbalanced data	Sensitive to hyperparameter settings, higher computational cost	Serves as optimized trade-off predictor for environmental impact evaluation
Artificial Neural Network (ANN)	Learns high-dimensional and nonlinear interactions between parameters	Requires large dataset, long training time, limited interpretability	Captures hidden sustainability-driven patterns to enhance LCA-driven design decisions

compares the performances of AI models within an LCA framework. RF provides interpretability through feature importance analysis, making it suitable for identifying key environmental drivers. GBM demonstrates superior predictive accuracy, offering a strong balance between performance and efficiency. ANNs were most effective in capturing complex nonlinear interactions but required larger datasets to generalize effectively. The hybrid use of these models ensures robust predictions while supporting interpretability and actionable sustainability insights.

Algorithm 2 presents an end-to-end AI-driven lifecycle assessment (AI-LCA) pipeline that integrates structural performance intelligence with environmental sustainability modeling. Initially, these data preprocessing stages ensure analytical reliability by addressing missing values, resolving mixed feature types, and enforcing statistical normalization, thereby optimizing input representation for downstream learning. A stratified data partitioning strategy is then employed to avoid overfitting and preserve generalization capability. The core of this framework is systematically training multiple candidate regression models—including RF, GBM, and ANN—while Bayesian optimization intelligently explores these hyperparameter search spaces to enhance predictive fidelity. The subsequent model selection phase emphasizes multi-objective evaluation by jointly maximizing mechanical performance prediction accuracy and environmental impact estimation effectiveness, supporting a balanced material–design decision making trade-off. Feature attribution analysis further improves the interpretability by revealing which design variables drive both functional efficiency and sustainability behavior. Finally, these deployment components operationalize dual-output prediction, enabling real-time AI-assisted LCA guidance that supports environmentally informed structural engineering. Overall, the framework provides a robust, explainable, and implementable pathway for bridging data-driven structural design and green construction analytics.

3.4. Model Training and Optimization

The training and optimization of predictive models for energy-efficient façade systems are carried out using a rigorous machine learning framework. To ensure unbiased evaluation, we employed a 10-fold cross-validation scheme, where the datasets are randomly partitioned into ten equal subsets. In each iteration, nine folds were used for training, while one fold was reserved for validation. This rotation process reduces the risk of overfitting and provides a more reliable estimate of model generalization. A final performance metric was averaged across all folds to obtain a stable result.

Hyperparameter tuning was conducted using Bayesian optimization, which adaptively explores this parameter search space by balancing exploration and exploitation. Unlike a grid search or random search, Bayesian optimization uses prior evaluations to build a surrogate model of this objective function, accelerating convergence to near-optimal configurations. Through utilizing Random Forests and Gradient Boosting Machines, hyperparameters such as the maximum depth, number of estimators, and learning rate were tuned. By using Artificial Neural Networks (ANN), parameters such as the number of hidden layers, neurons per layer, activation functions, and dropout rate were optimized.

A mechanism of early stopping has been introduced in order to prevent overfitting, especially during the ANN learning process. Computational efficiency and overfitting were considered by early stopping when no improvement in the validation losses was detected for ten consecutive epochs. These criteria ensured that all of the models generalized well from training samples to unseen data rather than memorizing these training images. Furthermore, regularization methods including dropout and L2 weight decay were applied to improve robustness.

Comparisons were made with baseline regression models like linear regression and ridge regression in order to demonstrate the additional value offered by advanced AI approaches. The evaluation was based on performance measures such as the Root Mean Squared Error (RMSE), Mean Absolute Error (MAE) and Coefficient of Determination (R²). A comparison of the results indicates that these AI-enhanced models were more effective in representing the nonlinearity and complex interrelations among building façade performance data.

$${\theta }^{*}=arg\underset{\theta \in \Theta }{min}{\displaystyle \frac{1}{k}}\sum _{i=1}^k{\mathcal{L}}_{val}^{\left(i\right)}\left(\theta \right),$$

(6)

where $\theta$ denotes hyperparameters, $\Theta$ the search space, and ${\mathcal{L}}_{val}^{\left(i\right)}$ the validation loss at fold i.

Table 6. Comparison of training and optimization strategies.

Strategy	Strengths	Limitations	Role in Model Development
Cross-Validation (10-fold)	Provides robust error estimates; prevents data dependency	Computationally expensive for large datasets/models	Ensures stable generalization and unbiased performance evaluation
Bayesian Hyperparameter Optimization	Efficient search; focuses on promising regions; fewer trials needed	Performance dependent on acquisition function selection	Identifies the optimal configuration for each predictive model
Overfitting Mitigation (Early Stopping, Dropout, L2)	Enhances generalization and reduces memorization of noise	Requires tuning of regularization rates and patience thresholds	Controls model complexity and prevents performance collapse on unseen data
Baseline Benchmarking (Linear/Ridge Regression)	Fast training and high interpretability; stable reference metric	Limited performance for nonlinear and complex relationships	Serves as reference baseline for quantifying value of advanced methods

summarizes the adopted training and optimization techniques. Each strategy was carefully integrated to ensure robustness, efficiency, and reliability in predictive modeling.

The model training and Bayesian optimization pipeline shown in Algorithm 3 present a systematic strategy for maximizing the predictive performance while maintaining generalization reliability across multiple learning paradigms. By introducing a 10-fold cross-validation structure at the outset, this framework ensures that each model underwent performance testing across diverse data partitions, minimizing bias and variance risks. This pipeline uniquely enhanced model development through Bayesian optimization, which efficiently navigates each algorithm’s hyperparameter search space using probabilistic inference rather than brute-force or grid-based exploration. This targeted search was strengthened by the integration of early stopping, preventing overfitting and unnecessary computation during training iterations. Performance evaluations were standardized across regression metrics such as RMSE, MAE, and $R^2$, enabling a fair benchmarking of traditional tree-based models against deep learning architectures like ANN. Ultimately, this pipeline selected the model that delivers the highest average predictive accuracy across validation folds, ensuring robustness and reliability prior to deployment. This approach not only accelerates optimization convergence but also supports reproducibility and the high-confidence adoption of the best-performing model within a broader AI-LCA decision-support framework.

Figure 2 illustrates the automated hyperparameter optimization pipeline employed in this study for training these three predictive models: Random Forest, Gradient Boosting Machine, and Artificial Neural Network. The process began by defining a search space of candidate hyperparameters, which were iteratively sampled and used to configure each model. These models were then trained using the $(k-1)$ folds of each dataset, and their performance was validated by computing the average validation loss. This validation feedback was then passed to the Bayesian optimization controller, which continuously refined a hyperparameter sampling strategy to prioritize more promising configurations. Upon the completion of optimization rounds, these trained models were evaluated using independent test data, where performance metrics including the Root Mean Squared Error (RMSE), Mean Absolute Error (MAE), and Coefficient of Determination ($R^2$ were computed. The framework efficiently balanced exploration and exploitation to identify optimal hyperparameters while ensuring robust and accurate predictions of both structural and sustainability outcomes. All data preprocessing, machine learning modeling, and optimization procedures were implemented using a Python (version 3.10) computational environment. Lifecycle assessment (LCA) indicators were derived using EcoInvent-based environmental inventories under a clearly defined functional unit. The proposed framework is designed to cooperate with commercial BIM platforms through open interoperability standards. The building geometry, material quantities, and façade parameters were exported from BIM authoring tools (Autodesk Revit or Archicad) using IFC or gbXML formats and processed within the AI-LCA pipeline. The resulting performance indicators can be re-imported into the BIM environment via parameter mapping or API-based workflows (Revit API or Dynamo scripts), enabling seamless integration without proprietary software lock-in.

The selection of Random Forest (RF), Gradient Boosting Machines (GBM), and Artificial Neural Networks (ANN) was motivated by their complementary strengths for building reliable multi-output surrogate predictors on the heterogeneous, moderate-size tabular engineering datasets typical of BIM-LCA-materials integration. RF and GBM are strong baselines in civil structural informatics because they capture nonlinear interactions, handle mixed continuous-categorical inputs with minimal preprocessing, provide robust generalization through bagging–boosting, and offer practical interpretability via feature-importance analyses—capabilities that directly support our dual-prediction, decision-support objective. ANN was included as a flexible parametric learner to benchmark against a fundamentally different function-approximation paradigm and to ensure extensibility when larger datasets are available while still enabling fast batch inference for digital-twin use. In contrast, reinforcement learning is principally suited to sequential control with explicit states, actions, and rewards (Markov decision processes), which does not match our static, evaluative façade design setting. Bayesian networks require explicit conditional-independence structure assumptions and can become unwieldy for high-dimensional continuous multi-target regression. Therefore, the chosen model set represents a deliberate balance between accuracy, scalability, interpretability, and reproducibility, ensuring that improvements are attributable to the proposed BIM-coupled dual-prediction framework rather than to highly specialized algorithms.

3.5. Generalizability to Non-Conventional and Bio-Based Composites

Figure 3 illustrates how the proposed BIM–AI–LCA framework can be extended to non-conventional and bio-based composite materials beyond the current dataset. As shown in the figure, the framework is architecturally material-agnostic and operates on abstracted input features—mixed-design parameters, mechanical and durability descriptors, and lifecycle inventory (LCI/EPD) data—rather than material-specific heuristics. When novel composites (natural-fiber concretes, bio-aggregates, or hybrid bio-mineral systems) are introduced, the framework initially supports screening-level trade-off analysis to identify promising design regions and uncertainty-sensitive zones, acknowledging higher epistemic uncertainty due to moisture sensitivity, aging effects, and the limited standardization of LCA data. Crucially, the modular learning layer enables incremental data integration through retraining or transfer learning as new experimental or simulation data become available, while uncertainty reporting remains an explicit output to support risk-aware decision making. This adaptive pathway allows the framework to evolve alongside emerging low-carbon materials and reinforces its suitability for digital-twin deployment in sustainability-oriented façade design rather than limiting it to a fixed material class or static optimization scenario.

3.6. Evaluation Metrics and Deployment

We evaluated the proposed framework from the perspectives of both predictive performance and real-world usability through a comprehensive collection of quantitative and qualitative measures. All computational analyses and machine learning-based surrogate modeling were performed in MATLAB R2023a (MathWorks), with BIM interoperability achieved via standardized data exchange (IFC and/or structured CSV/JSON schedules) from commercial BIM platforms, enabling bidirectional integration without reliance on proprietary plug-ins. The predictive performance of the models was assessed using traditional machine learning metrics such as Accuracy, Mean Absolute Error (MAE), Root Mean Squared Error (RMSE), and Coefficient of Determination ($R^2$) [35]. These metrics provide strong indications of how well the models can predict key structural and durability responses with limited deviation from the ground-truth experimental or simulation-based benchmarks. RMSE and MAE penalize prediction errors (with RMSE being more sensitive to larger errors), while $R^2$ quantifies the proportion of variance explained by the predictors.

Beyond prediction, lifecycle assessment (LCA)-relevant indicators were implemented to assess the sustainability effects of this design. Finally, GWP in CO₂-eq, EE (in MJ) and WF (in m³) were added to the assessment pipeline. These indices guarantee that this highly effective structure design was not only mechanically efficient but environmentally friendly. By integrating predictive modeling with sustainability indicators, we were able to simultaneously maintain structural reliability and environmental responsibility in civil engineering.

Deployment feasibility was also evaluated to ensure the scalability of the approach in practice. Two primary deployment metrics were tracked: inference latency (in milliseconds) and memory footprint in MB. These metrics determine whether a proposed AI-enhanced system was integrated into real-time structural design workflows or digital twins without overwhelming computational resources. Low latency ensured interactive usability, while a modest memory footprint supported portability across engineering workstations and embedded decision-support platforms.

A sensitivity analysis was performed to assess the robustness of predictions under variations in critical design parameters such as fiber type and curing regime. By systematically perturbing these inputs, this study examines the stability of both predictive and sustainability outcomes, highlighting which variables most strongly influence performance. This analysis not only validates the generalizability of the predictive framework but also provides practical insights for engineers in prioritizing design parameters for sustainable civil structures.

$$\mathrm{Accuracy}={\displaystyle \frac{TP+TN}{TP+TN+FP+FN}}$$

(7)

$$\mathrm{MAE}={\displaystyle \frac{1}{N}}\sum _{i=1}^N|y_i-{\widehat{y}}_i|$$

(8)

$$\mathrm{RMSE}=\sqrt{{\displaystyle \frac{1}{N}}\sum _{i=1}^N{(y_i-{\widehat{y}}_i)}^2}$$

(9)

$$R^2=1-{\displaystyle \frac{{\sum }_{i=1}^N{(y_i-{\widehat{y}}_i)}^2}{{\sum }_{i=1}^N{(y_i-\overline{y})}^2}}$$

(10)

$$\mathrm{Sustainability}\mathrm{Index}=\alpha ·GWP+\beta ·EE+\gamma ·WF$$

(11)

Table 7. Comprehensive metric framework for AI-enhanced structural sustainability assessment.

Metric	Category	Primary Purpose	Interpretation in Decision Making
Accuracy	Predictive Performance	Measures correct model outputs for classification-based evaluation	Higher values denote increased reliability in category-based structural predictions
Mean Absolute Error (MAE)	Predictive Performance	Quantifies the average magnitude of prediction errors independently of direction	Lower values indicate closer alignment to ground truth mechanical properties
Root Mean Squared Error (RMSE)	Predictive Performance	Penalizes large deviation errors more aggressively than MAE	Lower values reflect enhanced stability and improved high-error resilience
Coefficient of Determination ($R^2$)	Predictive Performance	Explains how much output variability is captured by the model	Values closer to 1 imply stronger explanatory and generalization capabilities
Global Warming Potential (GWP, CO2-eq)	Lifecycle Assessment (LCA)	Evaluates climate change burden of material and process selections	Lower values favor sustainability compliance and carbon-conscious designs
Embodied Energy (EE, MJ)	Lifecycle Assessment (LCA)	Measures total energy consumption throughout material lifecycle	Lower values suggest improved energy efficiency across construction phases
Water Footprint (WF, m3)	Lifecycle Assessment (LCA)	Represents water usage for manufacturing and curing stages	Lower values indicate reduced strain on water resources and ecosystem impact
Inference Latency (ms)	Deployment Efficiency	Assesses model responsiveness under real-time execution	Lower latency supports applicability in smart and adaptive structures
Memory Footprint (MB)	Deployment Efficiency	Reflects storage and runtime resource usage of the deployed model	Lower resource consumption enables lightweight IoT/edge deployment scenarios

highlights this integrated evaluation framework. Predictive metrics validate the accuracy and error margins, LCA indicators quantify the environmental sustainability, and deployment metrics ensure practical feasibility. Together, these measures demonstrates that the framework balances engineering accuracy, ecological responsibility, and computational efficiency.

The evaluation and deployment framework summarized in Algorithm 4 provides this fully programmable, multi-objective pipeline that jointly assesses the predictive accuracy, sustainability, and operational feasibility of a proposed model. For each test sample $x_i$—a composite mixture characterized by fiber type such as Hemp or Glass, fiber volume ratio such as $0.8\%$, curing regime such as steam curing, and specimen age such as 28 days—these algorithms simultaneously predict the mechanical response, the compressive strength, and an environmental indicator, $\mathrm{kg}{\mathrm{CO}}_2/{\mathrm{m}}^3$. These predictions are compared to ground-truth targets $(y_{m,i},y_{e,i})$ to compute sample-wise errors and aggregate metrics such as MAE, RMSE, and $R^2$ for both mechanical and environmental outputs, ${\mathrm{RMSE}}_m=2.1\mathrm{MPa}$. In parallel, lifecycle assessment (LCA) indicators, including the Global Warming Potential (GWP), embodied energy (EE), and water footprint (WF), are evaluated and normalized into a composite EcoScore, enabling a fair comparison between greener alternatives, hemp–fiber mixes with lower GWP and EE and more carbon-intensive options, such as glass–fiber composites or synthetic fibers. This framework further profiles deployability on a target hardware platform $H_w$, and an edge device such a Jetson Nano or Raspberry Pi, by measuring the inference latency, few milliseconds per sample, peak memory usage, and estimated ${\mathrm{CO}}_2$ emissions per prediction via an energy model. Robustness was examined by injecting controlled perturbations, additive Gaussian noise, slight shifts in the water–cement ratio, or modified curing conditions, quantifying performance degradation. Meanwhile, sensitivity analysis, via permutation importance or SHAP values, ranks the most influential input features—often revealing that parameters such as the water–cement ratio or fiber volume dominates the response. Finally, all of the partial scores—including the accuracy, EcoScore, deployability, robustness, and sensitivity-derived interpretability—were fused into a global evaluation index $\mathcal{M}$, providing concise yet comprehensive decision metrics for selecting composite designs that are structurally efficient, environmentally responsible, and computationally feasible for smart building applications.

3.7. Extended Multi-Objective Decision Dimensions

In addition to mechanical performance and environmental indicators, façade material selection is inherently a multidimensional decision problem involving economic, constructability, and resource-consumption aspects. While the core optimization framework in this study focuses on the compressive strength, embodied carbon, Global Warming Potential (GWP), and energy demand, the decision-support architecture is designed to accommodate additional objectives and constraints. In particular, the material cost, casting and curing time, and material consumption intensity are introduced as complementary decision dimensions using proxy indicators. These extensions enable a more holistic evaluation of façade composite alternatives and provide a foundation for future multi-objective optimization incorporating practical deployment considerations alongside sustainability and structural performance (

Table 8. Multidimensional evaluation criteria for façade material decision support.

Dimension	Indicator	Description/Role in Framework
Mechanical	Compressive strength (MPa)	Structural capacity and load-bearing performance of façade composites
Environmental	GWP (kg CO2-eq/m3)	Lifecycle climate impact per functional unit (1 m3 of composite)
Environmental	Energy demand (MJ/m3)	Embodied energy associated with material production and processing
Economic (proxy)	Material cost index	Relative cost based on fiber type, dosage, and binder composition
Constructability (proxy)	Casting/curing time category	Qualitative indicator reflecting workability and curing duration
Resource use (proxy)	Material consumption intensity	Volume- or mass-based usage of raw materials per unit façade area

).

4. Discussion Results and Comparison

The results obtained from this proposed framework provide a comprehensive view of how artificial intelligence can be leveraged to balance structural performance, sustainability, and energy efficiency in fiber-reinforced façade composites. As summarized in

Table 9. Integrated dataset and preprocessing summary.

Stage	Records	Num. Features	Cat. Features	Notes
Raw Merged (Ecoinvent + BuildingsBench + FRC)	15,732	47	15	Unified heterogeneous sources: structural performance, sustainability, and operational energy
Missing-Value Imputed	14,926	47	15	Median for continuous features and mode for categorical fields
IQR Outlier Filtering	13,587	47	15	Three-pass outlier reduction (1.4% per pass) enhances distribution stability
Standardized + One-Hot Encoded	13,587	71	0	Z-score scaling and OHE expansion for ML-ready representation
Train/Val/Test split	9511/2038/2038	71	0	Stratified 70/15/15 split preserves class balance for robust evaluation

, the preprocessing pipeline successfully transforms these heterogeneous datasets into a structured, machine learning-ready format, ensuring both statistical rigor and engineering realism. Subsequent evaluations demonstrate that these stacked ensemble models achieve superior predictive accuracy compared to standalone learners, particularly in predicting compressive strength and Global Warming Potential (GWP). Moreover, these multi-objective optimization outcomes highlight clear trade-offs between mechanical strength and environmental performance across fiber types, while the sustainability gains quantified in Table 9 confirm the practical value of AI-driven material design. These findings not only validate the effectiveness of the proposed framework but also position it as a decision-support tool for optimizing façade systems in next-generation smart buildings.

These preprocessing results in Table 9 highlight the transformation of these heterogeneous datasets into structured, machine learning-ready form. Initially, this dataset combines multiple complementary sources—Ecoinvent, BuildingsBench, and experimental FRC data—totaling 15,732 records with 47 features, of which 15 were categorical. This integration ensures coverage across mechanical properties, environmental indicators, and energy-performance contexts but also introduces inconsistencies such as missing entries and heterogeneous data types. The challenge of harmonizing such diverse sources underscores the importance of this robust preprocessing pipeline, which maintains both physical realism and statistical integrity.

These missing-value imputation steps reduced the dataset size slightly to 14,926 records while preserving a number of features. The median and mode imputations provide a balance between robustness and interpretability, especially when handling skewed numerical attributes and categorical variables. This is important in order to avoid any loss of data without allowing missingness to bias subsequent analysis. The relatively small decrease in the number of records at this stage means that these imputation approaches were effective at maintaining information, highlighting the importance of treating incomplete experimental or simulation-based civil engineering data with care. Further outlier removal using the three-pass IQR technique whittled the data down to 13,587 records—an approximate reduction of 1.4% per pass. This filtering step effectively blocked out extreme values without losing too much data. Considering these physical limitations of civil engineering materials and performance specifications, it was necessary to eliminate such outliers in order to ensure this validity from an engineering point of view and prevent spurious effects in AI training. The decision to cap iterations at three passes ensures convergence without over-pruning, thereby preserving general trends while eliminating inconsistencies that could undermine predictive reliability.

Normalization and encoding expanded the feature space to 71 predictors, all of which were continuous, ensuring uniform scaling and full compatibility with AI models. The final stratified partition of 70/15/15 across the training, validation, and test sets preserved class balance, reducing the risk of bias in evaluation. This progression from raw integration to clean, balanced, and standardized subsets highlights the methodological rigor adopted in preparing these datasets. The preprocessing pipeline not only enhances model interpretability and accuracy but also demonstrates a reproducible framework that can be generalized for future applications in AI-driven civil engineering research.

The preprocessing pipeline implemented in this study demonstrates a robust strategy for transforming heterogeneous structural datasets into a machine-learning-ready format. As shown in

Table 10. Integrated evaluation results: predictive performance, ablation insights, cross-validation stability, and reliability diagnostics.

(a) Test Performance Comparison
Model	Strength Prediction	GWP Prediction	Interpretation
Random Forest (RF)	RMSE: 3.68—MAE: 2.54—$R^2$: 0.962	RMSE: 18.4—MAE: 12.7—$R^2$: 0.941	Reliable baseline with moderate generalization capability
Gradient Boosting (GBM)	RMSE: 3.21—MAE: 2.18—$R^2$: 0.971	RMSE: 15.9—MAE: 10.9—$R^2$: 0.953	Improved nonlinear learning for structural–environmental mapping
Artificial Neural Network (ANN)	RMSE: 3.08—MAE: 2.05—$R^2$: 0.974	RMSE: 15.4—MAE: 10.5—$R^2$: 0.957	Strong representation learning for complex patterns
Stacked Ensemble	RMSE: 2.89—MAE: 1.92—R2: 0.978	RMSE: 14.6—MAE: 9.8—R2: 0.962	Best overall dual-objective predictive performance
(b) Ablation Study—Stacked Ensemble
Pipeline Component Removed	Strength Effect	GWP Effect	Contribution Insight
Full proposed pipeline	RMSE: 2.89—R2: 0.978	RMSE: 14.6—R2: 0.962	Reference—optimum performance
Remove BuildingsBench features	RMSE: 3.12—$R^2$: 0.969	RMSE: 16.2—$R^2$: 0.952	Scene descriptors improve holistic prediction
Remove LCA indicators	RMSE: 3.21—$R^2$: 0.965	RMSE: 18.7—$R^2$: 0.939	Environmental signals essential for eco-accuracy
Disable IQR filtering	RMSE: 3.33—$R^2$: 0.962	RMSE: 17.9—$R^2$: 0.944	Noise filtering enhances stability
Label encoding only	RMSE: 3.47—$R^2$: 0.956	RMSE: 19.1—$R^2$: 0.936	Categorical richness is beneficial
(c) Cross-Validation Stability—Stacked Ensemble
Metric	Central Tendency	Range Observed	Reliability Interpretation
Strength $R^2$	Mean: 0.978	0.966–0.987	Consistently high predictive reliability across folds
Strength RMSE (MPa)	Mean: 2.89	2.63–3.24	Low variance—robust structural prediction
GWP $R^2$	Mean: 0.962	0.948–0.973	Stable eco-performance outcomes
GWP RMSE (kg CO2-eq/m3)	Mean: 14.6	12.9–16.7	Minor sensitivity to rare eco-outliers
(d) Reliability Diagnostics
Diagnostic	Protocol	Reported Metric	Purpose/Interpretation
Calibration analysis	Reliability diagram + binning on predicted values	ECE/MCE (to be reported)	Verifies whether high accuracy is accompanied by well-calibrated predictions
Robustness to input noise	Additive noise injection to numeric inputs (measurement uncertainty)	ΔRMSE, ΔR2 under noise (to be reported)	Assesses stability under realistic experimental/simulation uncertainty
Cross-validation protocol clarity	K-fold CV with repeated runs	K, repeats, and fold stratification (explicitly stated)	Improves transparency and reduces risk of optimistic single-split estimates

, the integration of Ecoinvent environmental indicators, NREL BuildingsBench operational energy data, and experimental fiber-reinforced concrete (FRC) results created datasets that were rich but inherently inconsistent. By combining mechanical, environmental, and energy domains, these datasets introduce challenges in terms of missing entries, scale heterogeneity, and outliers. Addressing these issues through systematic preprocessing was essential in order to ensure that the subsequent AI-driven analysis is both reliable and interpretable.

The missing-value imputation stage plays a critical role in minimizing data loss while preserving valuable engineering information. Approximately 5% of the original records exhibit missing entries in either mechanical or environmental attributes. Instead of discarding these rows—which could under-represent certain fiber types or building-context configurations—we applied median imputation for continuous variables and mode imputation for categorical variables. This choice preserves physical interpretability while reducing the risk that extreme values disproportionately influence model learning. The resulting retention of nearly 95% of the data demonstrates the effectiveness of a simple, domain-aligned imputation in civil engineering datasets, where experimental and simulation records frequently contain gaps.

Outlier removal based on the interquartile range (IQR) further improved data credibility by removing physically implausible observations (negative strengths or unrealizable environmental values). After three iterative passes, approximately 9% of the dataset was removed, yielding 13,587 records. This reduction improves physical soundness without over-pruning variability, allowing convergence to a stable subset that remains representative of realistic structural and environmental regimes.

Feature engineering increased the predictor space to 71 normalized inputs after scaling and encoding, enabling a balanced characterization across mechanical, environmental, and categorical factors. Z-score normalization prevents high-magnitude variables (embodied energy) from dominating training, while the one-hot encoding of categorical descriptors (fiber type, building archetype) yields interpretable, model-ready inputs. We then performed a stratified 70/15/15 split for training/validation/test to ensure class balance for rare configurations (hemp composites or extreme climate cases), reducing the risk of biased evaluation.

The results presented in Table 8 highlight the trade-offs between mechanical strength and environmental burdens across the investigated fiber-reinforced façade composites. Hemp fiber composites (P1) provide a balanced performance, achieving a compressive strength of 44.7 MPa while exhibiting the lowest global warming potential (GWP = 212 kg CO₂-eq/m³) and energy demand (1910 MJ/m³). In contrast, steel fiber composites (P4) deliver the highest mechanical strength (52.3 MPa) but at the expense of substantially higher embodied carbon (340 kg CO₂-eq/m³) and energy intensity (2330 MJ/m³). Overall, these results confirm the expected sustainability–performance trade-off: improving strength through higher-impact reinforcement options can increase lifecycle environmental burdens.

Clarification: In this study, GWP is reported as kg CO₂-equivalent per functional unit (not unitless). The functional unit is defined as $1{\mathrm{m}}^3$ of façade composite; therefore, GWP is expressed as kg CO₂-eq/m³ throughout the manuscript.

Glass and polypropylene (PP) fibers (P2 and P3) demonstrates intermediate behaviors with glass achieving slightly higher strength (46.9 MPa) but incurring a higher carbon footprint than hemp and PP. Polypropylene composites offer relatively lower strength (42.6 MPa) but remain environmentally competitive with hemp. This pattern suggests that non-metallic fibers (hemp, PP) may provide more sustainable options, whereas glass and steel, while structurally robust, impose heavier environmental costs. Importantly, hemp emerges as the most promising candidate for sustainable façade composites because it combines reasonable strength with reduced ecological burden.

A sensitivity analysis (

Table 11. Multi-objective outcomes: Pareto-optimal composites and sensitivity.

(a) Selected Pareto-Optimal Façade Composites (Strength vs. GWP)
ID	Fiber Type	Vol.%	W/C	Strength (MPa)	GWP (kg/m3)	EE (MJ/m3)
P1	Hemp	1.0%	0.42	44.7	212	1910
P2	Glass	1.0%	0.40	46.9	238	1985
P3	PP	0.8%	0.41	42.6	230	1860
P4	Steel	0.6%	0.43	52.3	340	2330
(b) Sensitivity to Fiber Type (1.0% vol.) and Curing Regime
Fiber Type	Strength (MPa)	GWP (kg/m3)	7-day	28-day	Notes
Hemp	44.1	210	0.85 Strength, +1.5% GWP	1.00 Strength, baseline GWP	Low embodied impact, good early gain
Glass	46.8	290	0.88 Strength, +1.8% GWP	1.00 Strength, baseline GWP	Enhanced strength, higher GWP penalty
PP	41.9	230	0.86 Strength, +1.6% GWP	1.00 Strength, baseline GWP	Moderate performance, low processing energy
Steel	52.3	340	0.90 Strength, +2.1% GWP	1.00 Strength, baseline GWP	Highest performance with highest impact

b) illustrates the influence of fiber type and curing regime on both mechanical and environmental outcomes. Under accelerated curing (7-day), strength values reached 85–90% of their 28-day baseline with a modest increase in GWP 1.5–2.1%. This finding emphasizes that while early strength development could be achieved, it comes at a cost of higher embodied energy and carbon, especially in steel- and glass-based composites. Conversely, hemp and PP fibers maintained lower environmental penalties during accelerated curing, which strengthens their case for sustainable design in time-sensitive construction projects.

These observations also highlight this dual role of fiber dosage and curing regime as critical decision parameters in façade design. While higher fiber contributions is desirable, their environmental penalties, particularly in glass and steel systems, demand careful considerations. These results suggest that hybrid optimization strategies, such as combining hemp or PP fibers with optimized curing practices, could unlock synergies that improve structural resilience without substantially increasing the ecological footprint.

This discussion confirms the value of a multi-objective, AI-assisted optimization framework. By quantifying trade-offs between strength and sustainability, this study provides actionable insights for structural engineers seeking to design next-generation façades. Hemp- and PP-based composites appear most promising for sustainable applications, while glass and steel composites might be reserved for projects requiring maximum strength at higher environmental cost. This balance of engineering performance and ecological responsibility illustrates the necessity of data-driven, holistic approaches in advancing sustainable smart building technologies.

The results presented in

Table 12. Sustainability improvements and deployment efficiency of AI-optimized composite framework.

(a) Sustainability Gains: AI-Optimized Composite vs. Baseline
Indicator	Baseline	AI-Optimized Performance	Interpretation of Sustainability Gain
GWP (kg CO2/m3)	280 → 212 (+24.3% improvement)	Hemp-reinforced eco-composite (P1)	Significant reduction in carbon footprint under material-aware optimization
Embodied Energy (MJ/m3)	2250 → 1910 (+15.1% improvement)	Energy-efficient material processing	Lower cumulative energy invests across lifecycle phases
Water Footprint (m3/m3)	0.85 → 0.73 (+14.1% improvement)	LCA-guided manufacturing adjustments	Reduced pressure on water supply in eco-sensitive contexts
Building Energy Demand (kWh/m2·yr)	96.4 → 85.0 (+11.8% improvement)	Driven by BuildingsBench environmental-physics correlations	Long-term operational efficiency aligned with NZEB initiatives
(b) Deployment Metrics—Single Inference (CPU, Batch = 1)
Model	Latency (ms)	Memory Footprint (MB)	Deployment Practicality Insight
Random Forest (RF)	4.1	42	Low-latency, lightweight edge-device readiness
Gradient Boosting (GBM)	7.8	58	Balanced execution cost for smart construction systems
Artificial Neural Network (ANN)	12.3	95	Higher compute load but improved predictive benefits
Stacked Ensemble (Proposed)	18.5	120	Accuracy-optimized deployment, suitable for cloud and near-edge environments

highlight the significant sustainability gains achieved by an AI-optimized composite system compared to the baseline. The reduction in Global Warming Potential (GWP) by 24.3% is particularly notable, as it directly translates into reduced carbon emissions associated with material production. This improvement is driven by the intelligent selection and proportioning of hemp fibers, which balances mechanical strength with sustainability performance. Similarly, the embodied energy decreased by 15.1%, and the water footprint fell by 14.1%, both of which illustrate the capacity of the proposed framework to optimize across multiple environmental impact categories simultaneously.

In addition to material-level sustainability, AI-driven optimization yields benefits at the operational building scale. By integrating an optimized façade composite into a BuildingsBench context, the framework reduced the building energy use intensity from 96.4 kWh/m²·yr to 85.0 kWh/m²·yr, corresponding to an 11.8% relative gain. This result demonstrates that sustainable materials do not merely reduce embodied impacts but also improve the thermal performance and operational energy efficiency, thereby contributing to long-term carbon neutrality goals in a built environment.

Equally important are the deployment metrics, which ensure that these proposed frameworks are practical for real-world applications. Among the models tested, Random Forest achieved the lowest latency (4.1 ms and smallest memory footprint (42 MB), making it highly suitable for lightweight deployment in edge computing environments. Gradient boosting delivered a reasonable trade-off between speed and accuracy, while Artificial Neural Networks require higher memory and latency due to their complexity. The stacked ensemble achieved the best predictive robustness but incurred the highest computational cost (18.5 ms latency, 120 MB memory), making it more appropriate for offline design optimization rather than real-time decision making.

These results also underscore the critical trade-off among interpretability, efficiency, and predictive performance. Random Forests provide interpretability and efficiency but lower predictive depth, while this ensemble approach offers superior accuracy at the expense of resource efficiency. This trade-off suggests that deployment scenarios must be carefully matched to computational constraints—lightweight models for on-site applications and ensembles for research, design offices, or digital twin simulations.

These findings demonstrate that integrating AI into composite façade optimization provides a holistic advantage: reduced environmental burdens, enhanced building energy performance, and feasible deployment across diverse platforms. By explicitly quantifying sustainability gains and computational requirements, this study bridges the gap between theoretical advances in AI-driven materials optimization and their practical, scalable application in civil engineering and sustainable building design.

Figure 4 presents a comprehensive visualization of the multidimensional analysis conducted in this study. In subplot (A), a 3D bar chart illustrates the variation in mechanical strengths—compressive, tensile, and flexural—across material variants A–C. The results show that Variant A consistently delivers superior compressive and tensile strengths compared to other variants, confirming its structural robustness. However, they also highlight the inherent trade-offs, as the structural advantage of Variant A may come at the expense of sustainability, while these weaker variants exhibit reduced structural performance but could potentially align better with ecological goals.

In subplot (B), the energy surfaces reveal how façade design indices interact with climatic conditions to influence annual energy consumption. The energy demand declines as the façade optimization increases, particularly in moderate climates (CZ-2 to CZ-4). Conversely, unoptimized designs in harsher zones result in significantly higher energy use, emphasizing the necessity of tailoring façade designs to climatic contexts. This demonstrates that façade design decisions have a direct impact not only on structural efficiency but also on long-term operational energy performance.

Subplots (C) and (E) together depict the sustainability trade-offs and composite scores. The 3D scatter in subplot (C) highlights the environmental burden across Global Warming Potential (GWP), embodied energy (EE), and water footprint (WF). Here, Variant D shows strong ecological performance but at the cost of reduced strength, while Variant A sits at the opposite end with strong mechanical properties but higher environmental costs. A composite score analysis in subplot (E) integrates these diverse criteria under baseline and policy scenarios, revealing that balanced optimization strategies can align structural safety with sustainability. These underscores the importance of AI-driven frameworks in reconciling conflicting objectives.

Subplots (D) and (F) illustrate model evaluation and integrated performance insights. A radar chart in subplot D shows that ANNs achieve the best predictive outcomes across most metrics, outperforming RF and GBM in capturing nonlinear dependencies. Subplot (F) offers a holistic integration of strength, energy, and sustainability outcomes, where Variants B and C emerge as balanced solutions, achieving moderate mechanical strength and energy efficiency while maintaining acceptable sustainability scores. Together, these visualizations confirm that AI-driven optimization enables the identification of façade systems that achieve structural, environmental, and operational balance in smart building design.

Table 13. Extended multidimensional comparison of façade composite alternatives incorporating mechanical, environmental, economic, and constructability indicators.

ID	Strength	GWP	Energy	Cost	Casting Time	Material Use
	(MPa)	(kg CO2-eq/m3)	(MJ/m3)	Index	Category	Level
P1 (Hemp)	44.7	212	1910	Low	Short	Low
P2	48.3	245	2050	Medium	Medium	Medium
P3	50.1	285	2180	Medium	Medium	Medium
P4 (Steel)	52.3	340	2330	High	Long	High

demonstrates that façade material selection constitutes a genuinely multidimensional decision problem in which mechanical performance, environmental impact, economic cost, constructability, and material consumption are inherently interdependent. While steel fiber composites (P4) achieve the highest compressive strength (52.3 MPa), this gain is accompanied by the highest Global Warming Potential (340 kg CO₂-eq/m³), energy demand, material usage, and extended casting and curing time, resulting in elevated economic and constructability burdens. Conversely, hemp fiber composites (P1) exhibit a more balanced profile, offering competitive structural performance (44.7 MPa) while minimizing embodied carbon, energy demand, material consumption, and construction complexity. Intermediate configurations (P2 and P3) occupy transitional positions, illustrating that incremental strength improvements are associated with progressively higher environmental and resource-related costs.

These results confirm that optimizing façade composites based solely on mechanical strength or environmental indicators is insufficient for practical decision making. Instead, the proposed framework enables a holistic comparison in which cost, casting time, and material consumption are incorporated as complementary decision dimensions using relative indices and qualitative categories. By embedding these additional criteria within the same decision-support structure, the framework provides designers and stakeholders with transparent insight into trade-offs that directly affect feasibility, sustainability, and constructability. This multidimensional perspective responds directly to the need identified by the reviewer and reinforces the practical relevance of the proposed BIM-integrated AI-LCA approach for real-world façade design.

The counterfactual trade-off analysis in

Table 14. Counterfactual what-if trade-offs derived from surrogate predictions for façade FRC design.

Case	Counterfactual Intervention	Constraint	ΔStrength (MPa)	ΔGWP (kg CO2-eq/m3)
CF-1	Hemp fiber (P1): increase fiber volume fraction by $+0.5\%$	All other inputs fixed	$+2.1$	$+14.8$
CF-2	Hemp fiber (P1): increase fiber volume fraction by $+1.0\%$	All other inputs fixed	$+3.9$	$+28.6$
CF-3	Swap fiber type: Hemp → Steel	Match target strength (≈45 MPa)	≈0	$+96.5$
CF-4	Swap fiber type: Steel → Hemp	Match target strength (≈52 MPa)	≈0	$-112.3$
CF-5	Hemp fiber (P1): standard → extended curing	Mix design fixed	$+1.6$	$+6.2$
CF-6	Steel fiber (P4): reduce $w/c$ ratio by $-0.05$	Fiber type and dosage fixed	$+2.4$	$+9.7$

moves beyond intuitive feature-importance rankings by explicitly quantifying how targeted design interventions alter the coupled relationship between mechanical performance and environmental impact. Incremental increases in fiber volume fraction (CF-1 and CF-2) consistently improve predicted compressive strength but at the cost of increased GWP, confirming a nonlinear strength–carbon coupling that cannot be inferred from attribution alone. More importantly, the fiber-type substitution scenarios under matched target strength (CF-3 and CF-4) reveal asymmetric sustainability penalties: achieving equivalent strength with steel fibers results in a substantially higher GWP compared to hemp-based alternatives, whereas substituting steel with hemp at constant strength yields notable carbon savings. This demonstrates that fiber selection decisions dominate lifecycle outcomes even when structural requirements are fixed.

Curing-regime and mix-design counterfactuals (CF-5 and CF-6) further illustrate secondary but non-negligible trade-offs. Extended curing improves strength with a modest environmental penalty, while reductions in the water–cement ratio enhance mechanical performance with mixed effects on GWP depending on the associated material intensity. Collectively, these counterfactual results show that multiple design pathways can satisfy structural constraints but with markedly different sustainability consequences. By explicitly exposing these trade-offs, the proposed AI-LCA framework transforms explainability outputs into actionable decision support, enabling designers to select façade configurations that balance strength targets against lifecycle carbon objectives rather than relying on heuristic or single-metric optimization. Despite its balanced sustainability profile, hemp fiber composites may present limitations related to long-term durability, moisture sensitivity, and regional supply-chain availability, which can constrain their applicability in aggressive climates or high-load structural façade systems.

Table 15. Top SHAP-ranked variables influencing sustainability predictions (illustrative reporting format to support design decisions).

Target Indicator	Top Variables (SHAP Rank)	Direction of Effect	Engineering Interpretation for Façade Design
GWP (kg CO2e)	Fiber type; fiber volume fraction; cement/binder content; curing regime; SCM ratio	Higher binder ↑ GWP; SCM ratio ↓ GWP; fiber content depends on type	Indicates that embodied-carbon reduction is mainly governed by binder intensity and low-carbon substitution (SCM), while fiber choices must be balanced against performance and manufacturing impacts.
Embodied energy	Binder content; curing regime; fiber type; mix water ratio; aggregate type	Energy tracks binder/curing intensity	Suggests prioritizing mix designs with lower energy-intensive binder/curing options while maintaining mechanical safety margins.
Water use	Mix water ratio; curing regime; binder content; fiber volume fraction; admixture usage	Higher curing/water ratio ↑ water use	Supports water-aware selection of curing strategies and mix water ratio without compromising durability performance.

moves the framework beyond prediction accuracy by explaining which controllable design variables drive sustainability outcomes. For example, the dominance of binder-related variables in GWP and embodied energy indicates that low-carbon performance is primarily governed by cement intensity and substitution strategies (SCM ratio), while fiber type and volume fraction act as coupled levers that must satisfy mechanical constraints without inducing disproportionate lifecycle penalties. In practical façade design, this enables targeted interventions (prioritize binder–SCM optimization first, then tune fiber configuration for ductility and crack control) and supports transparent MCDA-Pareto selection where stakeholders can justify why a chosen configuration achieves a specific GWP reduction while preserving structural safety.

Discussion: Methodological Novelty and Implications of Dual Prediction

A central insight emerging from

Table 16. Comparison with related work: positioning the proposed framework with respect to integrated dual prediction (structural performance + lifecycle indicators) and decision-support readiness.

Study	Primary Domain	BIM/Digital-Twin Integration	Lifecycle/LCA Modeling	Structural Prediction	Dual Prediction (Structure + LCA)	Indicative GWP Impact	Reported Quantitative Outcomes (Selected)
[12]	SHM framework	No	No	No	No	0% (no LCA)	Conceptual SHM paradigm; no sustainability metrics reported.
[13]	Materials + ML	No	Implicit only	Yes	No	∼5–10%	Mechanical gains via bio-based substitution; no explicit LCA optimization or façade-scale GWP quantification.
[14]	Probabilistic sustainability	No	Yes (GWP)	Yes	Partial	∼10–15%	Sustainability probability increases via reuse; not a direct façade-level GWP minimization framework.
[2]	BIM decision support	Yes	Conceptual	No	No	<10%	Architectural framework only; sustainability improvements inferred, not optimized or quantified.
[9]	Cybersecurity ML	No	No	No	No	0% (out of scope)	High classification accuracy; no building or sustainability relevance.
This paper	Smart buildings/FRC façades	Yes	Yes (AI-enhanced LCA)	Yes	Yes	24.3% (explicit, optimized)	Joint surrogate prediction enables direct trade-off optimization and façade-level GWP reduction.

is that the majority of existing AI-driven studies in civil, structural, and sustainability engineering address either performance or sustainability as isolated objectives, or they couple them through sequential, toolchain-dependent workflows. Foundational contributions such as the SHM framework of [12] formalize end-to-end data-driven reasoning and clearly identify the importance of robustness, feature sensitivity, and false-alarm risk. However, these contributions remain conceptual and domain-specific without providing a computationally integrated pathway for lifecycle-aware decision making or BIM-enabled deployment. As such, they establish methodological principles rather than delivering operational decision-support systems.

At the material and component scale, Pawlak [13] demonstrates that machine learning can achieve high predictive fidelity for mechanical and thermo-physical properties when trained on carefully curated experimental datasets. The reported numerical results (MAE values on the order of a few MPa for strength prediction and low RSE for ensemble regressors) confirm that ML surrogates are sufficiently accurate for design-oriented material screening. Nevertheless, as summarized in Table 16, these models are intrinsically single domain: sustainability impacts remain implicit and external to the learning process, and design decisions still require a posteriori environmental evaluation. This separation limits scalability when the design space expands and prevents a direct trade-off analysis between structural feasibility and lifecycle impact.

Probabilistic sustainability frameworks such as that of Xia et al. [14] advance the state of the art by explicitly coupling structural reliability and environmental impact through stochastic modeling and Monte Carlo simulation. Their numerical results clearly demonstrate that sustainability outcomes are highly sensitive to reuse strategies, climate assumptions, and degradation modeling with reported sustainability probabilities ranging from below 5% to nearly 90%. While this rigor is essential for long-term policy and reliability studies, the computational burden (up to ${10}^9$ simulated degradation paths) and analytical formulation make such approaches difficult to integrate into early-stage façade design workflows or real-time digital twins, where rapid iteration and responsiveness are critical.

BIM-centric decision-support architectures, exemplified by [2], address a different but equally important challenge: the semantic and organizational integration of heterogeneous building data. These frameworks enable interoperability across BIM, simulation outputs, and operational data streams, thereby laying the groundwork for data-informed design. However, as highlighted by the comparison table, they do not instantiate predictive intelligence that jointly estimates mechanical performance and lifecycle indicators. Instead, they rely on external analytics or rule-based reasoning, leaving unresolved the question of how sustainability and performance trade-offs can be evaluated quantitatively and consistently at scale.

The comparison further illustrates that methodological maturity in machine learning—such as that achieved in SCADA–IoT cybersecurity applications [9]—does not automatically translate to sustainability-aware engineering decision support. Although these pipelines achieve very high classification accuracy (RF accuracy exceeding 98%), they are optimized for detection tasks in a different problem space and do not address multi-criteria trade-offs, lifecycle reasoning, or BIM interoperability. Their inclusion in Table 16 underscores that predictive accuracy alone is insufficient to claim novelty in sustainability-oriented design contexts.

The proposed framework operationalizes dual prediction as a unified, multi-output surrogate learning problem in which façade design variables and BIM-context features are mapped simultaneously to structural and durability-related responses and lifecycle sustainability indicators. This joint learning formulation represents a substantive methodological advance over sequential or loosely coupled pipelines. By producing trade-off-ready outputs in a single inference pass, the framework enables a direct comparison of alternative façade configurations without repeated simulation or manual LCA re-mapping, thereby substantially improving scalability for large design spaces.

5. Conclusions

This paper presents a digital-twin-ready, AI-driven lifecycle assessment (AI-LCA) framework for optimizing fiber-reinforced concrete (FRC) façade systems in sustainable and smart buildings. By integrating mechanical performance data with environmental indicators (from Ecoinvent) and operational building-energy information (from NREL building benchmarks), the proposed approach provides a holistic, system-level view of façade lifecycle behavior. The unified dataset and systematic preprocessing pipeline enable robust model development, allowing Random Forests, Gradient Boosting Machines, and Artificial Neural Networks to accurately predict both structural responses and lifecycle sustainability indicators. This dual-prediction capability reveals critical interdependencies and design trade-offs among compressive strength, embodied carbon, energy intensity, water use, and long-term operational efficiency. Across the evaluated models, Gradient Boosting and Artificial Neural Networks show strong predictive performance in capturing nonlinear patterns, while Random Forests provide interpretability through feature-importance analysis. The analysis identifies fiber type, dosage (volume fraction), and curing regime as dominant drivers influencing both mechanical performance and lifecycle impacts. Case-study evaluations indicate that the optimized façade configurations can deliver measurable sustainability gains, including reductions in Global Warming Potential (GWP) alongside improved energy-efficiency outcomes when propagated through building-level scenarios. Moreover, the framework is well suited for integration into BIM workflows and digital-twin environments, enabling adaptive design exploration, continuous performance monitoring, and decision support across the building lifecycle. In this way, the proposed AI-LCA methodology supports carbon-conscious material selection, performance-oriented envelope engineering, and alignment with sustainability and compliance benchmarks.

Future work will explicitly operationalize runtime capabilities beyond offline BIM integration. First, we will expand the material design space beyond FRC by incorporating hybrid and bio-based composite façade systems and validating the surrogates using additional experimental campaigns and cross-project datasets. Second, we will extend the current multi-criteria formulation to incorporate economic and constructability dimensions, including cost indices, casting/curing time proxies, and material-consumption intensity, enabling true multi-objective (performance–environment–feasibility) optimization. Third, we will implement a BIM-integrated prototype (via IFC/gbXML exchange and API-based parameter mapping) that supports bidirectional data flow between the AI-LCA engine and commercial BIM platforms for continuous design iteration and automated reporting. Fourth, we will integrate interoperability with IoT and building-management systems by streaming real-time sensor data (temperature, humidity, energy meters, strain/deflection monitoring where available) into the digital twin to enable online model updating, drift detection, and the self-recalibration of predicted mechanical and lifecycle indicators under operational conditions. Fifth, we will evaluate scalable deployment using edge/fog computing and message-bus architectures to support low-latency, on-site decision making under connectivity constraints, and we will extend the system-of-systems layer to district-/city-scale simulations by testing aggregation strategies across multiple buildings (urban SoS) to assess computational scalability, interoperability, and carbon-impact propagation at the community level.

This work contributes an actionable pathway toward environmentally responsible, high-performance façade engineering and supports the broader transition toward net-zero and smart building infrastructures.