Comprehensive modeling and experimental validation of polymer desiccant wheels across global operating and design ranges: A comparative study of effectiveness and multilayer perceptron models

Abstract

Accurately predicting the behavior of polymer desiccant wheels across coupled operating and geometric domains is essential for building-level simulations. This study establishes three models—constant effectiveness (CE), adaptive multi-grid effectiveness (AMGE), and a hyper-parameter-tuned multilayer perceptron regressor (MLP_R)—that map ten inputs (inlet air states, face velocity, wheel speed, and key geometric parameters such as wheel-area ratio, diameter, and depth) to the outlet process-air conditions. The models are trained on an extensive manufacturer dataset and validated against independent experiments that satisfy mass- and energy-balance criteria within ±5 %. Global sensitivity analysis reveals moderate pairwise correlations but strong nonlinear coupling among the inputs, underscoring the need to treat operation and design as an integrated space. The AMGE model reproduces every training sample exactly and limits experimental relative errors to −5.3 % for temperature and 9.8 % for humidity. The MLP_R achieves coefficients of determination of 0.5820 for temperature and 0.7900 for humidity, whereas the CE model attains 0.4863 and 0.7901, respectively. These results show that the AMGE approach provides a fast, geometry-aware surrogate. Its superior performance therefore holds promise for robust design, adaptive control, and energy optimization of next-generation desiccant air-conditioning systems.