Automated reaction path searches using the Artificial Force Induced Reaction (AFIR) method are typically based on expensive Density Functional Theory (DFT) computations. Researchers at ICReDD have explored the viability of reducing the computation time and cost of AFIR-based kinetic studies by replacing the expensive DFT computations with fast Neural Network Potential (NNP) predictions. They found that a general purpose NNP model struggles to provide accurate results on its own, but combining the NNP with predictions from an external model enabled it to reproduce the results of the more time-consuming DFT calculations.
As a test case reaction, the team considered the hydrogenation of ethylene via a simplified Wilkinson’s catalyst. Initially, DFT calculations were performed to provide training data. When the NNP model was trained using a portion of the DFT results, it could accurately predict the remainder of the DFT results. However, using that trained NNP model to perform a new AFIR search on its own produced inaccurate results.
One major reason for the poor performance is that the NNP’s mathematical model is not physics-based, so there was nothing stopping the model from exploring unrealistic molecular geometries. The artificial force of the AFIR method sometimes leads to exploration of such unrealistic geometries, but since such states were not in the training data the NNP did not know how to handle them.
To address this challenge, researchers included physics-based principles by utilizing delta learning to combine the NNP model with results from the GFN2-xTB level of theory. This additional theory acts as a guardrail by identifying unrealistic geometries and preventing the NNP from exploring them further. The NNP-xTB combination was able to reproduce reaction yields predicted by DFT with a training rate of 50% or greater. This rate is consistent with Generative Topographic Mapping results that showed the chemical space of the entire reaction network was already sampled at 50% search completion. Additionally, the combined NNP-xTB model performed better than either model alone, showing a synergistic benefit and highlighting the importance of designing adapted applications.