(509cr) Neural Networks Learn Fundamental Adsorption Energy Scaling Relations

Quantum mechanical simulations have the potential to accelerate the discovery of novel catalysts. Techniques with high accuracy require substantial computational resources. For this reason, there is a need to develop lower cost methods. It is common to employ adsorption energy scaling relations, which are linear relationships between the adsorption energies of two adsorbates. This approach is very structure sensitive, so it is a challenge to develop relations for complex surfaces. Machine learning may allow quantum mechanical simulations to be approximated. Here we show that neural networks learn fundamental scaling relations. These networks have been trained to predict the relaxed energy of an adsorbate placed on a surface. Linear scaling relations are even learned for adsorbates that have been excluded from the training process. The architectures of Schnet, Dimenett++, and CGCNN were all shown to learn scaling relations to varying degrees of accuracy. The learning process, as a function of training examples seen, and the learned representations are explored.