Abstract

Machine learning is emerging as a versatile tool in quantum chemistry. It offers access to force fields combining the accuracy of high-level electronic structure methods with excellent computational efficiency. In this talk, I will present how machine learning can be used to construct models of molecular potential energy surfaces and properties. By including equivariant components it is furthermore possible to enhance accuracy and data-efficiency. I then explore how these models can be extended to capture a wide range of physical phenomena by introducing a dependence on external fields. This makes possible to simultaneously predict different types of molecular spectra such as infrared and Raman, as well as chemical shifts. This approach also offers a simple way to account for external influences in the form of explicit and implicit solvation environments, e.g. in the context of QM/MM simulations. Finally, I show how the fully analytic nature of the model can be leveraged to design specific chemical environments.