Speaker
Description
Semiempirical quantum-mechanical (SQM) methods combine proper quantum-mechanical description of molecular systems with approximations that offer enormous increase of computational efficiency compared to more complex ab initio quantum mechanical or DFT methods. In the past decade, we have worked on improving the description of non-covalent interactions in SQM methods, and our PM6-D3H4X method is one of the top contenders in the field. Despite these successes, our extensive work with SQM methods also revealed their limitations. The most important unsolved issues are non-covalent interactions at short range, and the description of conformation energies.
To overcome these issues, we built a machine learning (ML) model serving as a correction for SQM method PM6 and trained it on an extensive database computed with quality DFT. The resulting PM6-ML method is able to correct all the remaining problems noted above, and it clearly outperforms both all previous SQM methods and standalone ML models in an extensive set of validation data sets covering different phenomena [ChemRxiv 2024]. In comparison to the previous Δ-ML approaches based on SQM methods, PM6-ML covers larger chemical space what makes it applicable to e.g. computer-aided drug design, and takes an advantage of a linear-scaling implementation of the SQM calculation what allows working with systems with thousands of atoms. An implementation of the method is available at [GitHub].
