A machine learning method has been developed to create reduced-order models (ROMs) with tens to hundreds of dimensions, derived from thermo-fluid analysis data. Due to their low computational cost, ROMs are well-suited for data assimilation, where observational data are used to calibrate the ROM parameters. This calibration enables the ROM to provide accurate predictions that support efficient equipment design, operation, and maintenance (O&M). In this study, we propose a data assimilation method that applies the multi-objective optimization algorithm COMO-CMA-ES, which is based on CMA-ES. By incorporating domain-specific insights into the design of the search space and objective functions, the optimized ROM achieves both high predictive accuracy and interpretability. The proposed method was applied to a data assimilation problem involving a battery system, where a 41-dimensional ROM was refined using observational data from six locations distant from heat-generating components. As a result, the optimized ROM predicted temperatures at unobserved locations with a root mean square error (RMSE) of less than 1 K.
Suzuki et al. (Wed,) studied this question.