(681e) Machine Learning for Predicting Accurate Quantum Chemical Energies

Reliable in silicodrug and materials design requires accurate yet fast methods for high-throughput screening. The most accurate high-level quantum chemical methods are generally too slow for that purpose. We demonstrated that accurate quantum chemical energies can be predicted by combining fast machine learning (ML) and low-level quantum mechanics (QM) methods given sufficiently large training sets.

One of such hybrid QM/ML techniques directly corrects energies calculated with low-level QM methods using ML (Î?-ML approach). Fast density functional theory (DFT) or semiempirical quantum chemical (SQC) methods can be used as low-level QM methods. Î?-ML approach readily achieves chemical accuracy (error ca. 1 kcal/mol).[1]

Another approach is to improve semiempirical Hamiltonian itself by using ML to predict on-the-fly SQC parameters for individual molecules (automatic parametrization technique, APT). APT stands in stark contrast to the traditional special-purpose reparametrization (SPR), where parameters are optimized for specific type of molecules and used unchanged for every other target molecule. Thus APT has several advantages over SPR: its accuracy can be further improved by increasing training set size, it is essentially black boxed, molecules far outside the training set are calculated with accuracy of SQC method with standard parameters.[2]

Finally, improved ML-based techniques can be used for very fast and accurate modeling of molecular potential energy surfaces, which are used to calculate highly accurate rovibrational spectra of small molecules (error ca. 1 cm^â??1).[3]

[1] R. Ramakrishnan, P. O. Dral, M. Rupp, O. A. von Lilienfeld, J. Chem. Theory Comput. 2015, 11, 2087.

[2] P. O. Dral, O. A. von Lilienfeld, W. Thiel, J. Chem. Theory Comput. 2015, 11, 2120.

[3] P. O. Dral, A. Owens, S. N. Yurchenko, W. Thiel, in preparation.