Quantum chemistry predictions with neural networks encoding pairwise interactions

Abstract

Accurate ab-initio prediction of electronic energies is very expensive for macromolecules by explicitly solving post-Hartree-Fock equations. After a introductory chapter on quantum chemistry and artificial neural networks, in Chapter 2, we illustrate an expressive and transferable deep neural network (T-dNN) model for the predictions of electron correlation energies at the MP2 and CCSD levels of theory, trained with the large amount of pairwise descriptors and energies hidden in a small amount of molecular data. The model is data efficient and makes highly transferable predictions across alkanes of various lengths, organic molecules with non-covalent and biomolecular interactions, as well as water clusters of different sizes and morphologies. Existing machine learning models, including T-dNN, attempt to predict the energies of large molecules by training small molecules, but eventually fail to retain high accuracy as the errors increase with system size. In chapter 3, we move on to remove the intrinsic failure of predictions on the large systems, by fine-tuning the pretrained representation on small systems with only few molecular or crystal data with ResT-dNN. Our model introduces a residual connection to explicitly learn the pairwise energy corrections, and employs various low-rank retraining techniques to modestly adjust the learned network parameters. We demonstrate that with as few as only one larger molecule retraining the base model originally trained on only small molecules of (H2O)6, the MP2 correlation energy of the large liquid water (H2O)64 in a periodic supercell can be predicted at chemical accuracy. Similar performance is observed for large protonated clusters and periodic poly-glycine chains. A demonstrative application is presented to predict the energy ordering of symmetrically inequivalent sublattices for distinct hydrogen orientations in the ice XV phase. We believe that underpinning the success of the T-dNN and ResT-dNN models, is the huge amount of pairwise data decomposed from few molecular training data, in addition to the engineered electronic features. The encouraging results motivate us to generalize it into existing graph neural networks to circumvent their need for large amount of molecular training data in the future.