Multi-objective optimization (MOO) coupled with building performance simulation has become a standard approach for exploring trade-offs among energy use, comfort, cost, and environmental impacts. Yet the same characteristics that make simulation-based MOO reliable, high-fidelity physics, detailed schedules, and explicit systems modeling, also make it expensive, data-intensive, and difficult to generalize across buildings and climates. In parallel, AI-driven acceleration (surrogate modeling, meta-model-assisted search, and hybrid simulation-learning workflows) has enabled orders-of-magnitude speedups, opening the door to larger design spaces and richer objective sets. However, many reported surrogate-assisted MOO pipelines remain narrowly scoped: models are often trained for a single building geometry under a single climate file and then optimized within that same context, limiting transferability to other climates, morphologies, operations, and system configurations. This paper synthesizes the state of simulation-based and AI-driven MOO for envelope-centric building design. It highlights methodological strengths (transparent physics, explicit constraint handling, and multi-criteria decision support), diagnoses recurring limitations (computational burden, discrete design spaces, workflow fragility, and evaluation inconsistencies), and emphasizes the generalizability challenge as a central barrier to practical deployment. The review concludes with research directions on benchmark-driven validation, uncertainty-aware and robustness-based optimization, interoperable BIM-BEM-LCA data pipelines, and climate- and geometry-spanning surrogate models that can support credible, scalable decision-making.
Kamazani et al. (2026) studied this question.