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

Deep learning methods provide a novel way to establish a correlation between two quantities. In this context, computer vision techniques like 3D-Convolutional Neural Networks (3D-CNN) become a natural choice to associate a molecular property with its structure due to the inherent three-dimensional nature of a molecule. However, traditional 3D input data structures are intrinsically sparse in nature, which tend to induce instabilities during the learning process, which in turn may lead to under-fitted results. To address this deficiency, in this project, we propose to use quantum-chemically derived molecular topological features, namely, Localized Orbital Locator (LOL) and Electron Localization Function (ELF), as molecular descriptors, which provide a relatively denser input representation in three-dimensional space. Such topological features provide a detailed picture of the atomic configuration and inter-atomic interactions in the molecule and are thus ideal for predicting properties that are highly dependent on molecular geometry. Herein, we demonstrate the efficacy of our proposed model by applying it to the task of predicting atomization energies for the QM9-G4MP2 dataset, which contains ~134-k molecules. Furthermore, we incorporated the Δ-ML approach into our model, allowing us to reach beyond benchmark accuracy levels (~1.0 kJ mol⁻¹). We consistently obtain impressive MAEs of the order 0.1 kcal mol⁻¹ (~ 0.42 kJ mol⁻¹) versus G4(MP2) theory using relatively modest models, which could potentially be improved further using additional compute resources.