Modeling Molecular Spectra with Interpretable Atomistic Neural Networks

Deep neural networks are emerging as a powerful tool in quantum chemistry, combining the benefits of high-level electronic structure methods with excellent computational efficiency. The recently developed SchNet model provides an accurate description of molecules and materials across chemical compound space, as well as easy access to energy conserving force fields [1]. Here, we demonstrate that the modular nature of deep models can also be exploited to enhance their versatility and offer insights beyond the basic relationships learned by the network. First, we adapt existing architectures to model different spectroscopic quantities, such as molecular infrared spectra [2]. Going beyond the simple prediction of properties, we then explore modifications of SchNet in the form of latent features. Although these variables are inferred, they correspond to readily interpretable physical concepts, such as molecular charge distributions [3].
[1] K. T. Schütt et al., J. Chem. Phys. (2018).
[2] M. Gastegger et al., Chem. Sci. (2017).
[3] K. T. Schütt et al., arXiv:1806.10349 (2018).