Multi-objective optimisation of building envelopes: A surrogate-assisted three-tier verification framework
Building and Environment, cilt.306, 2026 (SCI-Expanded, Scopus)
- Yayın Türü: Makale / Tam Makale
- Cilt numarası: 306
- Basım Tarihi: 2026
- Doi Numarası: 10.1016/j.buildenv.2026.115340
- Dergi Adı: Building and Environment
- Derginin Tarandığı İndeksler: Science Citation Index Expanded (SCI-EXPANDED), Scopus, Compendex, Environment Index, Geobase, Greenfile, ICONDA Bibliographic, The International Construction Database (ICONDA)
- Anahtar Kelimeler: Building envelope, Building performance simulation, Evolutionary algorithms, Generative design, Multi-objective optimisation, Sustainable architecture
- Ankara Yıldırım Beyazıt Üniversitesi Adresli: Evet
Özet
Building envelope design mediates four competing performance domains (thermal transmittance, structural load-bearing capacity, daylight provision, and material cost), yet conventional sequential, single-discipline practice systematically yields designs strictly inside the achievable Pareto frontier and, in some climates, failing strength-limit-state checks under code-specified loads. We introduce MOBO-Envelope, a three-tier physics-informed multi-objective optimisation framework. Path-A searches the trade-off space via NSGA-III accelerated by an ARD Matérn-5/2 Gaussian-process surrogate. Path-B verifies the full Pareto front with standards-grade Python simulators (openseespy ASCE 7–22 FEA, ISO 13790 thermal, IES LM-83-12 sDA, pvlib). Path-C cross-checks the complete front against external simulator binaries (EnergyPlus 26.1, Radiance 6.0) on canonical building models, as a high-fidelity check rather than a per-project certification. The Path-A thermal objective is the envelope-level effective U-value; whole-building annual energy use intensity (EUI) is retained as the high-fidelity verification quantity at Path-C. Across three climate-zone case studies (Houston 2A, NYC 4A, Minneapolis 6A), the surrogate achieves 5-fold cross-validated R2∈[0.93,0.996] and the bootstrap 95% hypervolume confidence interval spans ≤7% of the mean across five seeds. Path-A vs. Path-C Spearman correlations of ρtherm∈[0.62,0.73] (rising to 0.84 under a realistic autosizing HVAC model), ρday∈[0.76,0.87], ρstruct≥0.999 indicate that the surrogate ranks structural and thermal candidates reliably, while the daylight proxy is best treated as an early-stage screening indicator. All 789 Pareto designs across the three cases are re-evaluated at Path-C, and front fidelity is assessed against an independently optimised high-fidelity reference: the framework front retains 86% of the reference hypervolume, reaches the same best annual EUI and minimum structural utilisation at a 4.7% cost penalty, and exceeds direct high-fidelity search until that search is granted ≈1204 simulator evaluations. Against a code-compliant minimum-section baseline, best-compromise designs raise structural reserve by 0.70−0.81 and sDA by 9.3−38.6 percentage points at a 15.6−32.5% cost premium and a −2.7% to +87% effective-U-value regression, an observed Pareto trade-off. A direct-sun ASE1000,250 check on the verification subsets shows that the daylight proxy tracks over-exposure more faithfully than daylight itself (ρ=0.69–0.84 against ASE vs. ρ=0.38–0.52 against saturated high-fidelity sDA), so daylight-driven selections require a glare-aware re-evaluation. An algorithmic ablation within the analytical objective space, with all methods scored against the same reference point, shows the surrogate-assisted pipeline attains 29−33% higher hypervolume than vanilla NSGA-II/III at a matched evaluation budget, and that those baselines do not reach its front quality within 10000 evaluations, a ∼30× larger budget (residual margin 2.8−3.7%); this search-efficiency advantage is separate from the fidelity comparison above and does not transfer to it. The framework links sketch-stage architectural design to standards-compliant multi-physics verification in a reproducible workflow of ∼17 s analytical search plus ≈12 min complete-front high-fidelity verification per case, with the fidelity cost of surrogate acceleration quantified against an independently optimised high-fidelity reference rather than assumed, and with full code and artefacts publicly released.