Data-driven machine learning framework for thermal performance prediction and control of mini-channel and metal-foam assisted battery immersion cooling systems

Immersion-cooling battery thermal management systems are emerging as a critical technology; however, their advancement has been limited by analysis routes being misaligned with real-time operation. High-fidelity computational fluid dynamics are costly for repeated what-if evaluations, and recent machine-learning applications have tended to be geometry-centric and quasi-steady, offering limited support for sequence-aware prediction and constraint-aware optimization. Therefore, an operation-level data-driven machine-learning framework is introduced for a mini-channel and metal-foam immersion-cooling system. An LSTM with an additional fully connected head (LSTM-FC) is used to jointly predict sequence-wide thermal responses and a spatially resolved surface temperature field; it is coupled to a constraint-aware inverse-design loop, thereby shifting the role of machine learning from geometry tuning to sequence-aware, control-oriented operation in immersion cooling. A computational model validated by experiments was employed to generate a comprehensive transient dataset over a wide operating range, and the LSTM-FC was trained to estimate key temperature indicators with low inference latency, achieving high accuracy and strong generalization under previously unseen high charging rates, mixed C-rate profiles, and extreme thermal conditions. The inverse design procedure optimized the inlet temperature and mass flow rate under explicit limits on the maximum temperature, temperature difference, and hotspot suppression, reducing the time above safety limits from 79.2% to 5.9% for the maximum temperature and from 61.4% to 13.6% for the temperature difference relative to constant operation. Therefore, the proposed LSTM-FC framework enables fast and accurate unsteady prediction with real-time constraint-aware operating point selection and establishes a basis for future extensions to design-level optimization and online adaptation.