Thermal-Hydraulic Machine-Learning Surrogate Modeling for Structural Optimization of an Electric-Vehicle Battery Thermal Management System

Abstract Structural design of battery thermal management systems requires prediction of maximum temperature, maximum temperature difference, and pressure drop, whose scales and response forms differ. Direct multioutput regression overlooks response differences and the data support of optimized designs. This study develops a response-aligned machine-learning surrogate coupled with coverage-constrained multi-objective optimization and fuzzy comprehensive evaluation. Screening ten regressors identifies Extra Trees as the thermal baseline and a multilayer perceptron for pressure drop. Nested fivefold cross-validation repeated over five seeds selects direct Extra Trees for maximum temperature, log-response Extra Trees for maximum temperature difference, and velocity-normalized pressure-drop reconstruction with a multilayer perceptron. A label-independent sampling rule adds 40 designs to the 80-sample modeling set, while a disjoint 40-design holdout evaluates the deterministic distributed thermal network (DTN) mapping. On this holdout, the log-response and velocity-normalized branches reduce root mean squared error by 10.48% and 24.76%, respectively; the 95% paired-bootstrap interval confirms the pressure-drop improvement. Nominal 90% cross-validation-plus intervals achieve empirical coverages of 87.5%, 90.0%, and 100% for the three outputs. Across resamples, contact-length ratio remains the highest-ranked thermal feature, and local parameter sweeps reproduce the DTN trends. With 10,000 evaluations, coverage-constrained nondominated sorting genetic algorithm II produces 334 nondominated solutions and contributes 99.1% of the combined front. Coverage-aware fuzzy evaluation selects a candidate with a score of 0.648; removing coverage changes the first-ranked design but retains nine of the top ten candidates. Within the DTN-defined domain, the framework enables structural screening and prioritizes candidates for physical validation.