Two-stage cGAN-ridge surrogate for battery cold plate topology design

Topology optimisation of liquid-cooled battery cold plates requires repeated computational-fluid-dynamics (CFD) solutions. This cost limits design-space exploration. This study develops a two-stage surrogate for candidate generation and screening. Stage 1 uses a FiLM-conditioned conditional generative adversarial network to generate topology, temperature, and pressure fields from operating conditions. It produces 256 × 256 fields in about 35 ms. On 71 holdout cases, it gives a topology IoU of 0.704 and a temperature SSIM of 0.692. The outlet mean temperature exceeds the inlet mean temperature in all cases. Connectivity holds in 57 of 71 cases and mean pressure monotonicity is 66.9%. Stage 2 uses Ridge regression with geometry and operating-condition features to rank candidates. The main evaluation is geometry-grouped five-fold cross-validation, repeated over three seeds. Ridge gives 𝑅 2 = 0 . 3 1 1 –0.607 and Spearman 𝜌 = 0 . 7 6 1 –0.830 for the three temperature targets. It does not predict pressure drop reliably ( 𝑅 2 = − 0 . 0 0 9 , 𝜌 = 0 . 0 6 9 ). Pressure drop is therefore excluded from the automated score and is checked by CFD. A nested grouped comparison shows that tree-based models are promising offline alternatives, but they have not been deployed in the workflow. The method is a constraint-penalised surrogate, not a PDE-residual model. The workflow is designed to send only 2–3 of 20 candidates to CFD verification, which would remove most CFD runs from a full screening sweep. This reduction target has not yet been validated against a full-CFD ranking of all 20 candidates. Full CFD verification remains required for final design selection.