Scaling Up MLIP Models: A Comparative Study of TrIP2's Accuracy and Transferability Across Chemical Systems

The prohibitive computational cost of high-level quantum mechanical calculations has driven the development of machine learning interatomic potentials (MLIPs) that can efficiently approximate energies and forces with much lower computational expense than traditional quantum mechanical methods. However, many existing MLIPs suffer from limited generalizability, particularly with novel systems beyond their training sets. Here we introduce TrIP2, an advanced version of the Transformer Interatomic Potential (TrIP) trained on the expanded ANI-2x dataset, including more diverse organic molecular configurations with sulfur, fluorine, and chlorine. It leverages the equivariant SE(3)-Transformer architecture, incorporating physical biases and continuous atomic representations. Benchmarking on COMP6 energy and force calculations, structure minimization tasks, QC torsion energy profiles, and molecules with unexpected conformational energy minima demonstrates TrIP2's high accuracy and transferability, outperforming or matching ANI-2x, TrIP, AIMNet2, and MACE-OFF23. Without requiring any architectural modifications, TrIP2 successfully capitalizes on additional training data to deliver enhanced generalizability and precision, establishing itself as a robust and scalable framework capable of accommodating future expansions with minimal retraining. These attributes position TrIP2 as a highly promising tool for the development of transferable interatomic potentials.