(246h) Graph Convolutional Machine Learning Methods for the Predictions of Adsorption and Thermochemistry and Surface Stability

Prediction of surface phenomena such as surface stability or adsorption thermochemistry in catalysis have traditionally focused on simple coordination-based models or correlated these properties with electronic descriptors such as d-band properties. In recent work we showed that flexible coordination based fingerprints with sufficient data (thousands of calculations) were sufficient to achieve ~0.2 eV MAE accuracy for CO/H adsorption energies across a wide range of catalyst compositions. In this work we show that graph convolutional methods that learn important features and local activity from local atomic environments outperform our previous approach and are capable of achieving ~0.14 eV accuracy across thousands of surfaces with >30 unique elements while simultaneously predicting the contribution of each surface atom to the overall property. We also demonstrate this approach for a new dataset of intermetallic surface energies and show that these methods are capable of qualitatively predicting dominant nanoparticle facets. This works for sufficiently large datasets and demonstrates the transition from small-data (featurization and machine learning) to large data (deep learning) approaches for the prediction of surface phenomena in catalysis.