(513ef) An Active Learning Framework for Accelerating Saddle Point Searches Applied to Propylene Epoxidation

The production of propylene oxide through direct epoxidation with molecular oxygen catalyzed by metal-based catalysts such as Ag and Cu has been exciting chemists for years.¹ Of particular note here is the problem of studying the transition states involved in this reaction through analysis of saddle points. Previous work on accelerating saddle point searches have focussed on machine learning based models applied to nudged elastic bands (NEBs).² Here we aim to apply a novel active learning framework to decrease the computational times used in calculating these NEBs. The building of these NEBs often involves hundreds or thousands of ab-initio force calls which are computationally expensive. The active learning framework developed as a part of the open source AMPTorch allows for significant reduction of these force calls. AMPTorch is a modification of the software Atomistic Machine Learning Potentials (AMP).³ The framework couples physics such as the Morse potential, and the associated parameters to improve the model efficiency and convergence. The goal here is to apply this framework to the reactions involved in production of propylene oxide and improve the computational efficiency of finding their transition states. Applications to simpler systems show a reduction in the number of force calls by a factor of nearly 3, the aim would be to achieve at least such a reduction or better with this reaction. The final goal is to present an effective active learning framework that can accelerate these saddle point searches and be widely adopted.

References:

1. Yimeng Dai et al. ‘Significant enhancement of the selectivity of propylene epoxidation for propylene oxide: a molecular oxygen mechanism’. (2017) Phys.Chem.Chem.Phys. 19, 25129.
2. Andrew Peterson. ‘Acceleration of saddle-point searches with machine learning’. (2016) J. Chem. Phys. 145, 074106.
3. Alireza Khorshidi, Andrew Peterson. ‘Amp: A modular approach to machine learning in atomistic simulations’. (2016) Computer Physics Communications 207:310-324.