3D Computer Vision Models Predict DFT-Level HOMO-LUMO Gap Energies from Force-Field-Optimized Geometries



We investigate 3D deep learning methods for predicting quantum mechanical energies at high-theory-level accuracy from inexpensive, rapidly computed molecular geometries. Using space-filled volumetric representations (voxels), we explore the effects of radial decay from atom centers and rotational data augmentation on learnability. We test several published computer vision models for 3D shape learning, and construct our own architecture based on 3D inception networks with physically meaningful kernels. We provide a framework for further studies and propose a modeling challenge for the computer vision and molecular machine learning communities.


Supplementary material

Computational details; model architectures; voxel parameter screen results; full-scale modeling results.

Supplementary weblinks

Example data processing and voxelization code, model classes, and training and evaluation code. Demonstrated on a small subset (10k) of the full dataset.