Efficient Nudged Elastic Band Method using Neural Network Bayesian Algorithm Execution

Minimum energy pathways (MEPs) provide critical insights about transition states and energy barriers for chemical systems. A popular method for MEP dis- covery is the nudged elastic band (NEB) algorithm, which involves an expensive optimization using hundreds to tens of thousands of potentially expensive simulations. AI methods can reduce this cost, but have historically focused on either directly running NEB on static, pre-trained models or actively updating simple (Gaussian process) simulation surrogates. To our knowledge, we are the first to unite these two regimes by using Bayesian Algorithm Execution (BAX), a technique from Bayesian experimental design, to fine-tune EquiformerV2, a foundation model. We demonstrate the resulting neural network BAX (NN-BAX) method on Lennard-Jones transitions and observe that NN-BAX requires one to two orders of magnitude fewer energy/force function evaluations compared to classical NEB, with negligible loss in accuracy. We highlight that this targeted fine-tuning procedure retains the simulation efficiency of previous active approaches, while allowing scalability to systems of higher complexity and dimensionality. Finally we introduce Foundation-BAX, a technique to achieve an even greater speedup on multi-step transitions.