A multi-stage supervised learning optimisation approach on an aerogel glazing system with stochastic uncertainty

Climate-adaptive aerogel materials and resilient condition-dependent thermophysical properties with stochastic uncertainty can enhance the reliability and robustness of aerogel glazings, whereas multidimensional optimal design is highly dependent on stochastic uncertainty magnitude and sampling size, leading to ineffectiveness or inefficiency of traditional physics-based models. Furthermore, given the time-variant meteorological parameters with high-level uncertainties, climate-adaptive design on aerogel materials in building glazing systems can resist heat flux and reduce heat gain, so as to reduce the cooling energy consumption in subtropical climates. In this study, uncertainty optimisation was conducted in a subtropical climate region with sensitivity analysis, using a two-stage learning approach. Results indicate that, with the increase of stochastic sampling size from 18 to 72, the training epoch required to learn accurate optimisation function increases from 5000 to 20,000. Compared to the deterministic scenario, a gradual decrease in total heat gain can be noticed for uncertainty-based optimal scenarios. Furthermore, dynamic thermal performance is highly dependent on uncertainty magnitudes, but insensitive to stochastic sampling size. This study quantifies the impact of sample size and uncertainty magnitude on dynamic thermal performance with frontier guidelines, providing climate-adaptive aerogel glazings under stochastic scenario uncertainties.