Machine-learning assisted optimization of residential building envelopes across contrasting Italian climates: energy, thermal comfort, and carbon footprint

Energy efficiency, thermal comfort, and life-cycle carbon are central to residential design, yet practical automated workflows that scale across climates remain rare. This study aims to optimize residential envelopes in two contrasting Italian climates, identify cross-climate solutions that remain competitive in both contexts, and then re-rank the shortlist with life-cycle carbon (A1–A3, B6) to reveal energy–carbon interactions. This study develops and validates an automation pipeline around EnergyPlus that explores 14,400 envelope options and simulates 2,880 space-filling cases (Latin Hypercube with corner-point coverage) to model EUI, heating/cooling loads, and comfort (PMV/PPD). Eight algorithms are benchmarked as surrogates under a uniform tuning budget. CatBoost delivers the best fidelity for EUI and heating, while a tuned artificial neural network performs best for cooling and PMV; both achieve R 2 > 0.96 with low errors across targets. Re-simulation and Pareto-front identification yield five top-performing envelopes per city and reveal a robust overlapping subset that remains competitive in both climates. Although the most energy-efficient design (Wall 12) achieved top simulation scores, its high embodied carbon reversed its overall ranking; the Wall 14 configuration emerged as the most balanced low-carbon solution across climates, as confirmed by a component-level sensitivity analysis identifying walls as the dominant carbon contributors. The workflow demonstrates cross-climate portability by accurately predicting building performance in an unseen climate (Rome) using local weather descriptors and envelope U-values, incorporates boundary-aware sampling, identifies high-accuracy surrogate models for each key performance indicator (EUI, heating, cooling, PMV), and compares carbon-informed re-rankings with energy and comfort-based rankings of building envelopes.