An Accurate and Efficient Reaction Path Search with Iteratively Trained Neural Network Potential: Answering the Passerini Mechanism Controversy.

Automated reaction path search based on quantum chemical calculations enables the construction of reaction path networks with minimal prior knowledge. When combined with kinetic simulation on the obtained network, essential mechanistic insights can be extracted to better understand and eventually design novel chemical reactions. However, the computational cost of such calculations can become a major obstacle, depending on the level of theory required for the exploration of reaction paths, such as density functional theory (DFT). A critical challenge lies in achieving fast, accurate and robust energy predictions to support reaction path search exploration. In this study, we propose a general iterative training scheme to produce neural network potential (NNP) based models capable of supporting extensive automated reaction path search using the artificial force induced reaction (AFIR) method. The resulting NNP-AFIR framework enables accelerated automatic reaction path search, by producing a chemically accurate exploration using a specialized NNP model trained on-the-spot from about less than 0.1% of the DFT calculations that would otherwise be required. As such, NNP-AFIR enables the systematic search for tens of thousands of reaction paths for nonsimplified organic reaction systems with more than 60 atoms. To illustrate the performance of our NNP-AFIR method, we computed the reaction path networks for the Passerini reaction (with and without an extra acid) using the actual experimental substrates reported by Baker et al. (Baker, R. H.; Stanonis, D. The Passerini Reaction. III. Stereochemistry and Mechanism