Executive Summary : | Artificial Intelligence and Machine Learning (ML) methods have influenced almost all research areas of science over the past decade and theoretical chemistry, is not an exception. It is possible for an appropriate ML algorithm to identify an underlying pattern of a data set. Such advancement can ultimately lead to a powerful model, which is the broad objective of this proposal. In this project, we propose to develop a code set-up, which will be able to perform on-the-fly dynamics using ML method. In this new setup, the potential energies required for the dynamics will be obtained from ML based calculations. In order for an ML algorithm to identify the pattern of a data set, numerical representation of that data set is required. Such a representation is known as descriptor. In the literature, many ML descriptors are found, namely, invariant many-body interaction descriptor, smooth overlap of atomic positions, atom-centered symmetric functions, distance matrix, Coulomb matrix, and many more. In the proposed research, the potential energy and its gradient will be considered to be functions of interatomic distances which can be obtained if the Cartesian coordinates are available. An appropriate Machine learning algorithm will be used to train a set of pre-calculated data based on ab initio methods. A supervised learning technique will be followed in this case. Translational/Rotational invariance of the energy will be taken care of by using proper algorithms. Multi-Layer Perceptron (MLP) & Deep Neural Network (DNN), Support Vector Regression (SVR), and Convolutional Neural Network (CNN) algorithms of ML techniques will be tested for the current research. Ozonolysis of catechol is one of the very relevant reactions that take place in the environment. Calculations suggest that B3LYP/6-311+G (2df, 2p) level of theory represent the potential energy very well. On-the-fly chemical dynamics simulations are being performed with PM7 semiempirical level which works very similar to B3LYP/6-311+G (2df, 2p) level. ML algorithm will be used to train in a supervised way on B3LYP/6-311+G (2df, 2p) potential energy points and the predicted energies will be first checked with some pre-calculated points. After obtaining satisfactory test results, this ML-PES will be directly used in the on-the-fly dynamics. Calculations will be performed in both Gas and Condensed phase systems and the results will be compared. Another possibility would be a typical bimolecular nucleophilic substitution (SN2) reaction, namely, X– + CH3Y → XCH3 + Y–, with X = Cl, OH. and Y = I, F. The ab initio dynamics studied earlier suggests that these reactions are highly sensitive and without proper PESs, the results can be entirely wrong. In one case a rebound mechanism led to a direct reaction and for other, a potential energy minimum was observed in the exit channel due to the formation of a hydrogen bond. It will be interesting to see if such observation can be reproduced with ML-based PESs. |