PEMFC Performance Prediction & Design Optimization
NUS summer research — dual MLP models for PEMFC thermal and pressure prediction, genetic design search validated in COMSOL, and Optuna hyperparameter tuning.
Context
Summer research at National University of Singapore, Prof. Birgersson’s Lab (May – Aug. 2024).
What I built
- Trained dual MLP neural networks (43,961 parameters) with PyTorch and MATLAB for PEMFC thermal and pressure modeling, achieving 3% MAE.
- Implemented a genetic search algorithm using the dual networks as objective functions, achieving 2% error in optimal fuel cell design validated through COMSOL simulations across 50+ test scenarios.
- Bayesian-tuned hyperparameters with Optuna across 30+ configurations, reducing MAE by 15% and cutting training time by 40% vs. grid search.
Technologies
PyTorch, MATLAB, Optuna, COMSOL, Python