Introduction
As electric vehicles (EVs) demand longer ranges and faster charging, battery thermal management systems (BTMS) have become a critical bottleneck. High temperatures increase the risk of thermal runaway and shorten battery life. While "cell-to-pack" (CTP) designs utilizing direct liquid immersion cooling offer superior heat transfer, determining the optimal spatial arrangement of cells within these packs is a complex combinatorial challenge. Traditional design methods rely on slow, iterative trial-and-error using Finite Element (FE) simulations, which are inefficient for exploring the vast landscape of possible layouts.
To address this, researchers have developed a "cooling-guided" diffusion model that automatically optimizes battery cell arrangements. By integrating physics-based surrogate models and feasibility classifiers into the generative process, the system produces layouts that are significantly more efficient at dissipating heat than standard generative AI methods.
Methodology: The Cooling-Guided DDPM
The proposed architecture, a Cooling-Guided Denoising Diffusion Probabilistic Model (DDPM), consists of three primary components:
- The Generative Engine (TabDDPM): A specialized diffusion model for tabular data (X and Y coordinates). It learns the distribution of valid layouts and generates new ones by progressively removing noise from a random state.
- Classifier Guidance: To ensure physical feasibility (e.g., preventing cell overlap), a binary classifier was trained. During generation, the classifier's gradient "nudges" the model toward feasible designs.
- Cooling Guidance: This core innovation uses a surrogate neural network to predict the maximum temperature of a layout. Gradients from this model "steer" the diffusion process toward regions of the design space that minimize temperature.
To facilitate training, the team developed a simplified simulation model that approximates high-fidelity electrothermal behavior while running 90% faster than traditional methods.
Data Strategy and Training
The study utilized a multi-stage data pipeline to train the various components:
- Initial Layouts: 100,000 configurations generated via a "repulsive force" algorithm.
- Expanded Training Set: Augmented to 700,000 layouts (via mirroring and rotation) for DDPM training.
- Feasibility Dataset: 1.4 million layouts used to train the classifier.
- Simulation Subset: 10,000 configurations processed through simplified FE simulations. The Synthetic Minority Over-sampling Technique for Regression (SMOGN) was applied to balance this set, ensuring the surrogate model could accurately predict rare, high-performance layouts.
A key finding was the importance of "canonical" ordering; sorting cell coordinates consistently (e.g., bottom-left to top-right) helped the DDPM recognize spatial patterns more effectively.
Key Results
The cooling-guided DDPM was evaluated against TabDDPM (unconditioned) and CTGAN (a generative adversarial network) using a Composite Quality Index (CQI) that factors in feasibility, diversity, and cooling efficiency.
- Efficiency: The proposed model achieved a CQI 5 times higher than TabDDPM and 66 times higher than CTGAN.
- Feasibility: The model reached a feasibility rate of 71.9%, vastly outperforming TabDDPM (6.8%) and CTGAN (0.7%).
- Cooling Performance: The guided model consistently generated layouts with lower maximum temperatures, successfully identifying high-efficiency cooling paths that baseline models missed.
- Classifier Reliability: The feasibility classifier achieved an F1 score of 0.91.
Limitations and Future Work
While promising, the current study is limited to a specific 20-cell square pack design and a 2C discharge rate. The model's performance is also capped by the accuracy of the surrogate model; any "blind spots" in the surrogate could be exploited by the generator. Future research will focus on generalizing the model to different pack shapes, discharge cycles, and additional objectives such as structural integrity and weight.
Summary updated