Multi-objective optimization of desiccant wheel performance using adaptive sampling and SHAP analysis

Design optimization of desiccant wheel-based dehumidification systems requires performance data that sufficiently cover a high-dimensional design space. However, obtaining such data through conventional design of experiments is burdensome because experiments and repeated high-fidelity simulations are both time-consuming and costly. In this study, an FDM model was developed from coupled heat and mass transfer equations for a solid desiccant wheel and validated against published experimental data. The validated model was then used as a physics-based reference solver to generate training data for an artificial neural network (ANN) surrogate model. To improve data efficiency under a limited simulation budget, an error-guided adaptive sampling strategy was proposed and compared with one-shot Latin hypercube sampling (LHS). Under the same total sample budget, the adaptive strategy reduced the mean squared error by 12.86% for moisture removal rate (MRR) prediction and by 19.07% for regeneration heat load prediction. The selected ANN surrogate was subsequently coupled with the non-dominated sorting genetic algorithm II (NSGA-II) to perform multi-objective optimization, where the objectives were to maximize MRR and minimize regeneration heat load. The resulting Pareto-optimal solutions were further analyzed using SHapley Additive exPlanations (SHAP) to quantify the contributions of six design variables and to interpret their directional effects along the Pareto front. In addition, SHAP clarified how key variables change systematically from energy-prioritized solutions to performance-prioritized solutions. The proposed framework improves surrogate-learning efficiency, enables energy-oriented design optimization, and provides an interpretable basis for selecting practical operating and design conditions in desiccant wheel systems.