Machine Learning of Parameters for Accurate Semiempirical Quantum Chemical Calculations

摘要: We investigate possible improvements in the accuracy of semiempirical quantum chemistry (SQC) methods through use machine learning (ML) models for parameters. For a given class compounds, ML techniques require sufficiently large training sets to develop that can be used adapting SQC parameters reflect changes molecular composition and geometry. The ML-SQC approach allows automatic tuning individual molecules, thereby improving without deteriorating transferability molecules with descriptors very different from those set. performance this is demonstrated OM2 method using set 6095 constitutional isomers C7H10O2, which accurate ab initio atomization enthalpies are available. ML-OM2 results show improved average much reduced error range compared standard results, mean absolute errors dropping 6.3 1.7 kcal/mol. They also found superior specific reparameterizations (rOM2) same isomers. thus holds promise fast reasonably high-throughput screening materials molecules.