Towards Reliable Machine Learning for Computational Chemistry and Perspectives for ML Excited-State Molecular Dynamics
Dr. Pavlo O. Dral
Max-Planck-Institut für Kohlenforschung
Venue: Institut de Chimie Radicalaire – Aix Marseille University
Campus Saint Jerome, ICR, Porte BJ5, 4th floor, D42, Library
Date: Wednesday, October 25, 2017
Part I: 10:00-11:00
Part II: 14:00-15:00
Machine learning (ML) becomes increasingly popular in computational chemistry as it allows for significant reduction of the required computational time for many applications. Since ML is not based on any physically motivated model, it should be used with caution as ML can lead to huge unphysical outliers. I will discuss several ways suggested by us for making ML a reliable tool for computational chemistry. One approach is to use low-level, fast QM methods, e.g. semiempirical QM or DFT, as fail-safe and apply ML to correct predictions made by these QM methods. This technique has now become a standard in the field and is called Δ-learning. Another approach is to use ML to improve semiempirical Hamiltonian itself. It is also possible to reduce the number of large outliers by making ML work in the interpolation rather than extrapolation regime with structure-based sampling of training points. This sampling combined with self-correction has allowed us to reduce the number of QM calculations by up to 90% to simulate rovibrational spectra with errors of less than 1 cm−1. Finally, I discuss the perspectives and challenges of using ML for performing excited-state molecular dynamics.
 R. Ramakrishnan, P. O. Dral, M. Rupp, O. A. von Lilienfeld, J. Chem. Theory Comput. 2015, 11, 2087–2096.
 P. O. Dral, O. A. von Lilienfeld, W. Thiel, J. Chem. Theory Comput. 2015, 11, 2120–2125.
 P. O. Dral, A. Owens, S. N. Yurchenko, W. Thiel, J. Chem. Phys. 2017, 146, 244108.