At BMW Group’s research and innovation center, I developed AI-based surrogate models to replace traditional physical simulations of auxiliary spring behavior in vehicle dynamics.

What I built

• End-to-end ML pipeline in Python/PyTorch: data generation, model training, evaluation, and inference
• Evaluated 4 architectures (Transformer, FCNN, cGAN, Encoder-Decoder) for characteristic curve prediction
• Designed custom loss functions (Curvature-Informed, Physics-Informed) tailored to domain constraints
• Deployed the production model into BMW’s existing C++ simulation toolchain

Impact

• Encoder-Decoder achieved best real-world performance (MSE 0.0011); custom loss improved Transformer test loss by 61.6%
• Enabled faster simulation workflows during early vehicle development stages
• Collaborated across 3 engineering teams to ensure legacy system compatibility and reproducibility