CFD-informed machine learning for surface-temperature mapping of side-cooled 4680 lithium-ion cells under thermal uncertainty

Abstract Large-format 4680 lithium-ion cells present a thermal-monitoring challenge because surface measurements do not fully describe the temperature field and key heat-transfer properties remain uncertain. This study combines a validated three-dimensional steady thermal model with machine learning to reconstruct the external can-surface temperature field of a side-cooled 4680 cell. Seventy-five CFD cases were generated by varying radial jelly-roll conductivity from 0.90 to 2.00 W m − 1 K − 1 and equivalent interface thickness from 0.40 to 0.60 mm at a fixed 6.0 W heat input and 25°C cooling boundary. Temperatures from 25 surface locations produced 1,875 spatial samples. The CFD model reproduced three published validation temperatures within 0.196°C, and the two finest meshes differed by less than 0.01%. CNN-LSTM and SVM yielded similar held-out point-level performance, with R² values of 0.981 and 0.980 and RMSE values of 0.238 and 0.245°C, respectively; MLP was less accurate. Feature perturbation identified axial position as the dominant predictor, followed by x-position and radial conductivity. The resulting surrogate reproduces the simulated can-surface temperature field within the investigated steady-state parameter range and provides a computationally inexpensive basis for thermal-sensitivity analysis.