[Paper Review] Machine Learning Prediction of Accurate Atomization Energies of Organic Molecules from Low-Fidelity Quantum Chemical Calculations

Logan Ward, Ben Blaiszik+4 more|arXiv (Cornell University)|Jun 7, 2019

Machine Learning in Materials ScienceMaterials Science43 references57 citations

TL;DR

This paper presents machine learning models that predict G4MP2-level atomization energies of organic molecules with near-quantum accuracy at a fraction of the computational cost by learning the difference between B3LYP (low-fidelity) and G4MP2 (high-fidelity) energies. The best models achieve a mean absolute error of 0.005 eV for molecules with up to 9 heavy atoms and 0.012 eV for larger molecules, using SchNet and FCHL-based Δ-learning approaches, and are publicly accessible via a web API on DLHub.

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.

Motivation & Objective

• To overcome the accuracy-cost tradeoff in quantum chemical calculations for organic molecules.
• To develop machine learning models that predict high-accuracy G4MP2 atomization energies using low-cost B3LYP data as input.
• To evaluate and compare the performance of deep learning (SchNet) and traditional ML (FCHL) approaches in predicting accurate molecular energies.
• To enable accurate, fast predictions for molecules larger than those in the training set, including those with 10–14 heavy atoms.
• To make state-of-the-art, high-accuracy models freely accessible via a web interface for broad community use.

Proposed method

• Trained models on the QM9-G4MP2 dataset containing 117,232 molecules with 1–9 heavy atoms, using B3LYP and G4MP2 energies.
• Employed a Δ-learning strategy to predict the difference between B3LYP and G4MP2 atomization energies, improving accuracy over direct prediction.
• Used SchNet, a deep convolutional neural network, to learn molecular representations from atomic coordinates and element types.
• Applied FCHL, a kernel-ridge regression method based on many-body atomic environment descriptors, for traditional ML-based prediction.
• Retrained models on perturbed molecular conformers to reduce dependence on exact B3LYP-optimized geometries.
• Published the best-performing models on DLHub with a simple Python API for cloud-based inference, eliminating local installation requirements.

Experimental results

Research questions

• RQ1 Can machine learning models trained on low-fidelity B3LYP data accurately predict high-fidelity G4MP2 atomization energies?
• RQ2 How does Δ-learning—learning the difference between low- and high-fidelity energies—compare to direct learning or transfer learning in accuracy and robustness?
• RQ3 To what extent can models generalize to molecules larger than those in the training set, especially with 10–14 heavy atoms?
• RQ4 How sensitive are model predictions to the quality of molecular geometry inputs, such as B3LYP-optimized vs. force-field-generated coordinates?
• RQ5 Can a web-based interface enable widespread, accessible use of high-accuracy ML models without requiring expertise in ML frameworks?

Key findings

• The FCHL Δ-learning model achieved a mean absolute error (MAE) of 0.005 eV for molecules with 9 or fewer heavy atoms, outperforming B3LYP by a factor of 10.
• The SchNet Δ-learning model achieved an MAE of 0.012 eV for molecules with 10–14 heavy atoms, demonstrating strong extrapolation capability.
• Models trained with perturbed conformers reduced dependency on exact B3LYP-optimized geometries, though performance degraded for larger molecules.
• The FCHL-based model achieved three times better accuracy than the SchNet model but required at least 100 times more computational time per prediction.
• Both models significantly outperformed B3LYP, reducing error by up to a factor of 10 across all molecular sizes in the test sets.
• The best models are publicly available via a web API on DLHub, enabling easy integration into research workflows without local ML infrastructure.