Machine learning surrogates for Al2O3–TiO2/water hybrid nanofluid battery thermal management under variable heat generation

Abstract Hybrid Al 2 O 3 –TiO 3 /water nanofluids are promising coolants for lithium-ion battery thermal management systems, but evaluating their flowrate, heat-generation, and composition effects using CFD alone remains computationally demanding. This study develops machine-learning surrogate models for a liquid-cooled LG Chem 15 Ah pouch cell using a 72-case validated ANSYS Fluent dataset. The dataset combines a 28-case baseline matrix with a 44-case heat-generation extension covering 55–277.5 W. Three surrogate models multilayer perceptron (MLP), support vector regression (SVR), and convolutional neural network–long short-term memory (CNN-LSTM) were trained and evaluated using Reynolds number, heat generation rate, Al 2 O 3 volume fraction, and TiO 2 volume fraction as inputs. The predicted outputs were the heat transfer coefficient and maximum hotspot temperature. The results show that the heat transfer coefficient is governed mainly by Reynolds number, while the maximum hotspot temperature is controlled by both Reynolds number and heat generation rate. Nanofluid composition provides a consistent but secondary correction to the thermal response, with the hybrid nanofluid giving the lowest hotspot temperature within the sampled range. SHAP attribution and bootstrap confidence intervals were used to interpret the MLP response and assess its sensitivity to dataset resampling, while constrained random sampling was used to map trends within the investigated design space. The resulting framework provides rapid, multi-output screening of hybrid-nanofluid BTMS operating conditions without requiring a new CFD calculation for each query, although predictions remain limited to the validated Reynolds-number, heat-load, and composition ranges.