Abstract
The computational prediction of atomistic structure is a long-standing problem in physics, chemistry, materials, and biology. Conventionally, force-fields or ab initio methods determine structure through energy minimization, which is either approximate or computationally demanding. This accuracy/cost trade-off prohibits the generation of synthetic big data sets accounting for chemical space with atomistic detail. Exploiting implicit correlations among relaxed structures in training data sets, our machine learning model Graph-To-Structure (G2S) generalizes across compound space in order to infer interatomic distances for out-of-sample compounds, effectively enabling the direct reconstruction of coordinates, and thereby bypassing the conventional energy optimization task. The numerical evidence collected includes 3D coordinate predictions for organic molecules, transition states, and crystalline solids. G2S improves systematically with training set size, reaching mean absolute interatomic distance prediction errors of less than 0.2 Å for less than eight thousand training structures — on par or better than conventional structure generators. Applicability tests of G2S include successful predictions for systems which typically require manual intervention, improved initial guesses for subsequent conventional ab initio based relaxation, and input generation for subsequent use of structure based quantum machine learning models.
Highlights
After training on sufficiently many examples, we find that G2S generated structures for out-ofsample graphs have a lower root-mean-square deviation (RMSD) than structures from ETKDG6 and Gen3D7 and exhibit high geometric similarity to the reference quantum chemical structure
We have presented G2S, a machine learning model capable of reconstructing 3D atomic coordinates from predicted interatomic distances using bond-network and stoichiometry as input
The applicability of G2S has been demonstrated for predicting structures of a variety of system classes including closed-shell organic molecules, transition state geometries, singlet carbene geometries, and crystal structures
Summary
For atomization energy prediction of C7O2H10 and C7NOH11 isomers, G2S and FCHL19 still reaches an accuracy of 5 kcal/mol mean absolute error (MAE) at 1024 training points, slowly approaching the coveted chemical accuracy of 1 kcal/mol, and almost matching the accuracy of a DFT structure-based BoB model. The advantage is most substantial for the small training set, in the limit of larger data sets, the performance curves of predictions based on G2S input level off, presumably due to the noise levels introduced by aforementioned error type B, i.e., inherent noise and conformational effects of the predicted structures. Possible further strategies to improve on G2S could include Δ-machine learning[28] where deviations from tabulated (or universal force-field based) estimates are modeled
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have