A Systematic Benchmark of Multifidelity Training Strategies for Reactive Machine Learning Potentials

Training reactive machine learning potentials (MLPs) requires extensive quantum chemical labeling of nonequilibrium configurations along reaction pathways, motivating the use of data from lower levels of theory through pretraining and Δ-learning. However, whether combining these strategies consistently improves accuracy and downstream performance remains unclear. Here, we benchmark direct and residual training, each with and without pretraining, across DPA-2, DPA-3, MACE, and DPA4-Neo, using 17.6 million configurations labeled with GFN2-xTB and 198,622 configurations labeled at the ωB97X-D/def2-TZVP level. With the full DFT training set, pretrained Δ-learning achieves the lowest energy or force MAE averaged over three training seeds in 31 of 32 combinations of architecture, data set, and property. Learning curves for DPA4-Neo nevertheless show that direct finetuning reaches the MAE of a model trained from scratch on the full data set using no more DFT training configurations than pretrained Δ-learning in all eight combinations of data set and property, while avoiding repeated GFN2-xTB evaluations during inference. Pretrained Δ-learning also improves reaction energies and forward barriers across all four architectures on GDB-10-rxn and reduces median Hessian eigenvalue errors for the evaluated DPA4-Neo checkpoint. However, local force errors can remain incompletely corrected, and these improvements do not consistently increase the success rates of reaction search workflows initialized from specified reactant and product geometries, indicating that the preferred protocol depends on the data budget, target property, and downstream workflow.