Computational chemistry researchers have traditionally faced a frustrating trade-off when simulating molecular interactions: use fast classical force fields that sacrifice precision or rely on accurate quantum-mechanical methods that run too slowly on large jobs.
Machine-learned force fields (MLFFs) close that gap by training neural networks on high-fidelity quantum data. When it comes to modern drug discovery and materials design, though, there’s demand for even faster processing speeds to handle massive chemical libraries involved. To overcome such performance constraints, Schrödinger partnered with Google Cloud to deploy AlphaEvolve, an evolutionary AI coding agent developed by Google DeepMind that iteratively generates and refines algorithms to find the most efficient code path overcoming the algorithmic bottleneck.
Schrödinger — a leader in developing scientific software for over three decades — identified two critical algorithms within their MLFF training pipeline that limited performance: neighbor list computation and Ewald summation. These algorithms aggregate data from atomic neighbors and calculate long-range potentials, but both became limiting factors in training and inference speed.
Schrödinger's primary technical goal was speeding up AI model training for energy and force calculations. Specifically, they targeted the Ewald summation, a critical but computationally demanding function used in molecular mechanics. The Ewald sum was the main performance constraint in Schrödinger's PyTorch code. It had no established vectorized algorithm and often relied on simple for-loops that ran slowly on large simulations.
By incorporating AlphaEvolve into their models, the system could generate a batched implementation of the Ewald summation using parallel batch matrix multiplication. This would evolve the PyTorch code to outperform existing custom kernels.
Schrödinger used a rigorous multi-layered evaluation framework to confirm the evolved code was both performant and scientifically accurate:
Inverse time (primary metric): The core objective was to maximize throughput by reducing calculation time, from a baseline score of 7.9.
Functional correctness: All evolved programs had to pass a full test suite, including regression tests on complex systems such as disordered water models.
Success rate: This was measured by the share of programs that were both functionally correct and faster than the baseline.
“AlphaEvolve allows us to explore larger chemical spaces faster and more efficiently than ever before. Faster MLFF inference carries real business impact, shortening R&D cycles in drug discovery, catalyst design, and materials development, and enabling companies to screen molecular candidates in days rather than months.” — Gabriel Marques, technical lead of machine learning, Schrödinger
By applying AlphaEvolve, Schrödinger replaced simple for-loops in the Ewald summation code with parallel batch matrix multiplication. This optimization raised the program success rate from less than 1% (40 out of 5,000 evaluations) to more than 60%, while improving the performance metric from the baseline of 7.9 to nearly 30.
Optimizing these foundational algorithms delivered a 4x speedup in both MLFF training and inference. This acceleration lets researchers compress molecular screening timelines and directly benefits several key research areas:
Drug discovery: Identifying viable therapeutic candidates quickly to address urgent medical needs.
Catalyst design: Developing efficient chemical processes for industrial applications.
Materials development: Designing next-generation materials with custom properties for electronics and energy storage.
Schrödinger plans to apply this evolutionary approach to custom GPU kernels to test whether AI-generated code can outperform human-engineered implementations.
Read the full technical paper on AlphaEvolve to learn how evolutionary AI agents optimize scientific codebases, or contact the Google Cloud AI team to discuss accelerating your research workflows.