Abstract
Modeling of chemical reactions is essential for understanding kinetic mechanisms and predicting possible outcomes of reacting systems. Quantum mechanical calculations are accurate but often prohibitively expensive. Deep learning has emerged as a faster alternative, but progress is slowed by a fragmented software ecosystem that hinders reuse, fair comparison, and reproducibility. We present ChemTorch, an open-source framework that streamlines model development, experimentation, hyperparameter tuning, and benchmarking through modular pipelines, standardized configuration, and built-in data splitters for in- and out-of-distribution evaluation. We envision ChemTorch as a foundation for community-driven method development and reproducible benchmarking in chemical reaction modeling. As a first step toward unified benchmarks, we compare four representative modalities for barrier-height prediction on the RDB7 data set, including fingerprint-, sequence-, graph-, and 3D-based approaches. Our results highlight clear advantages of structurally informed models and sharp performance drops under out-of-distribution conditions, highlighting the importance of rigorous benchmarking.