Physical and Analytical Chemistry Seminar
Lecturer: Dr. Pavlo Dral
Location: Faculty Seminar Room
Machine learning (ML) calculations are much faster than typical quantum mechanical (QM) calculations. ML is however not based explicitly on any physically motivated model and therefore predictions made with ML can be very wrong. We proposed several strategies for more reliable use of machine learning (ML) techniques in computational chemistry. In one of them, called Δ-ML or Δ-learning, we use low-level, fast QM calculations as a fail-safe. With this approach the accuracy of low-level QM method can be improved significantly, e.g. performance of semiempirical QM methods corrected by ML nears that of more accurate DFT methods. In another approach we use ML to improve semiempirical Hamiltonian, which leads to more accurate electronic structure calculations. We have also demonstrated that many unphysical outliers of ML calculations can be eliminated by structure-based sampling. This sampling combined with self-correction can be used to create very accurate ML potentials able to simulate rovibrational spectra with spectroscopic accuracy and only 10% of computational cost required for pure QM approach.
We also develop general-purpose semiempirical quantum chemical methods with improved accuracy. Our new orthogonalization- and dispersion-corrected methods (ODMx) are generally more accurate than older orthogonalization-corrected methods (OMx) for both ground- and excited-state properties and are as good as OMx with dispersion corrections for noncovalent interactions.
 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.
 P. O. Dral, X. Wu, L. Spörkel, A. Koslowski, W. Weber, R. Steiger, M. Scholten, W. Thiel, J. Chem. Theory Comput. 2016, 12, 1082–1096.
 P. O. Dral, X. Wu, W. Thiel, work in progress.