(662d) Machine Learning Electron Density for Chemical Property Predictions in Catalysis

Recent advances in machine learning have demonstrated the ability to predict many properties of molecules and materials including formation energy, atomic forces, dipole moments, and polarizability. However, many of these properties can be easily computed from the ground-state electron density obtained via a density functional theory calculation. Thus, significant effort has also been spent in developing machine learning models that can predict electron density. Because it extends in three-dimensional space, electron density predictions require specialized model architectures which are more complex than models of scalar quantities such as energy. Models must also enforce desirable properties such as electron conservation and long-range decay. We present a graph neural network model that strictly retains these properties. We demonstrate the use of physics-based methods to calculate properties of heterogeneous catalysts and adsorbates directly from the model’s electron density predictions. Namely, we use Bader charge partitioning to predict partial charges and density functional theory to predict exchange-correlation and electrostatic energies. We briefly discuss challenges that remain in predicting other properties of interest such as atomic forces and total energy from electron densities. These results represent a new paradigm of machine learning in molecular science: training a model on electron density alone enables predictions of a wide array of other properties.