Machine-learning surrogates are often used to speed up thermal-system optimisation, but their aggregate performance metrics can conceal local modelling inaccuracies. A new preprint on arXiv proposes a two-stage residual learning approach that explicitly accounts for the local strength of the model during optimisation.
The study assesses local modelling strength through validation-based reliability, rather than relying solely on global error measures. This allows the optimisation process to weight predictions according to where the surrogate is most trustworthy. The abstract is truncated, so the full method and results are not yet available, but the motivation is clear: improving the reliability of surrogate-based optimisation in thermal systems.