Computational chemistry traditionally faced a speed-accuracy trade-off in molecular simulations. Machine-learned force fields (MLFFs) improved accuracy but still struggled with the massive datasets needed for modern drug discovery. To address this, Schrödinger partnered with Google Cloud and used AlphaEvolve, an AI coding agent, to optimize their MLFF training pipeline. They identified two key algorithms, neighbor list computation and Ewald summation, as performance bottlenecks. The primary goal was to accelerate AI model training, specifically targeting the computationally demanding Ewald summation in their PyTorch code. AlphaEvolve was tasked with generating a more efficient implementation of this algorithm. The system successfully evolved the PyTorch code by replacing slow for-loops with parallel batch matrix multiplication for the Ewald summation. Rigorous evaluation confirmed the evolved code's functional correctness and performance improvements, with a success rate increasing significantly. This optimization resulted in a four-fold speedup in MLFF training and inference, accelerating drug discovery, catalyst design, and materials development. Schrödinger plans to further explore this evolutionary approach for optimizing custom GPU kernels.