Capturing the potential energy landscape of large size molecular clusters from atomic interactions up to a 4-body system using deep learning

Exploring the structure and properties of molecular clusters with accuracy using the ab initio methods is a resource intensive task due to the increasing cost of the ab initio methods and the number of distinct conformers as the size increases. The energy landscape of methanol clusters has been previously explored using computationally efficient empirical models to collect a database of structurally distinct minima, followed by re-optimization using ab initio methods. In this work, we propose a new method that utilizes the database of stable conformers and borrow the fragmentation concept of many-body-expansion (MBE) methods in ab initio methods to train a deep-learning machine learning (ML) model using SchNet. Picking 684 local minima of (CH3OH)(5) to (CH3OH)(8) from the existing database, we can generate similar to 51 000 data points of one-body, two-body, three-body and four-body molecular systems to train an ML model to reach a mean absolute error (MAE) of 3.19 kJ mol(-1) (in energy) and 2.48 kJ mol(-1) angstrom(-1) (in forces) tested against ab initio calculations up to (CH3OH)(14). This ML model is then used to create a database of low energy isomers of (CH3OH)(n) (n = 15-20). The proposed scheme can be applied to other hydrogen bonded molecular clusters with an accuracy of first-principles methods and computational speed of empirical force-fields.

Automated summary

This study proposes a new method that utilizes a database of stable conformers and borrows the fragmentation concept of many-body expansion (MBE) methods in ab initio methods to train a deep-learning machine learning (ML) model using SchNet. This method allows for accurate exploration of the structure and properties of molecular clusters with the accuracy of first-principles methods and the computational speed of empirical force-fields.