Machine learning prediction of accurate atomization energies of organic molecules from low-fidelity quantum chemical calculations

Abstract

Recent studies illustrate how machine learning (ML) can be used to bypass a core challenge of molecular modeling: the tradeoff between accuracy and computational cost. Here, we assess multiple ML approaches for predicting the atomization energy of organic molecules. Our resulting models learn the difference between low-fidelity, B3LYP, and high-accuracy, G4MP2, atomization energies, and predict the G4MP2 atomization energy to 0.005 eV (mean absolute error) for molecules with less than 9 heavy atoms and 0.012 eV for a small set of molecules with between 10 and 14 heavy atoms. Our two best models, which have different accuracy/speed tradeoffs, enable the efficient prediction of G4MP2-level energies for large molecules and are available through a simple web interface.

A data ecosystem to support machine learning in materials science

“…However, while ab initio methods exist that can predict the energy of molecules with accuracies comparable to the uncertainty in corresponding experimental data (e.g., G4MP2 [31]), the computational expense of these high-accuracy methods limits their widescale use. To address this issue, Ward et al [32] built machine learning models that use a recently-published MDF dataset to predict high-accuracy, G4MP2 energies from the outputs of faster, but inaccurate calculations (B3LYP).…”

“…Time to predict the G4MP2-level atomization energy of 100 molecules given input B3LYP energies and relaxed structure using a machine learning model produced by Ward et al[32] as a function of molecule size. Timings were measured over 64 identical runs with one servable container running in the DLHub service.…”

“…The authors would also like to acknowledge and thank the researchers who made their datasets and/or models and codes openly available [26,30,32].…”

3D Dense Convolutional Neural Networks Utilizing Molecular Topological Features for Accurate Atomization Energy Predictions

“…Therefore, in the present work, we attempt to leverage and adapt some of the latest developments in fields such as computer vision for the task of predicting atomization energies at high levels of accuracy. In this context, we note that Ward et al 28 have achieved a highly impressive out-of-sample mean absolute error (MAE) of the order of 0.1 kcal mol −1 (versus the G4(MP2) 29 level of theory) on the QM9-G4MP2 dataset 27,30 using the SchNet and FCHL 31 models in conjunction with the Δ-Machine Learning approach. 32 As the name suggests, the Δ-ML strategy targets learning the energy difference between an expensive target level of theory and a cheaper baseline level of theory, thus exploiting the systematic nature of the error between the two theoretical methods.…”

“…Thus, given the energy at the baseline theory, energy at the expensive level of theory could be obtained using the ML-learned additive correction term. Indeed, Δ-ML procedures have been shown to provide significantly better accuracy than models attempting to learn absolute energies directly, 28 thus allowing to reach chemical accuracy (±1.0 kcal mol −1 ) and within striking distance of the elusive benchmark accuracy (±1.0 kJ mol −1 ) with respect to the experimental value (or a high level of theory) through machine learning means. Therefore, we have also incorporated the Δ-ML model in our proposed machine learning protocol.…”