Abstract
Accurate simulations of molecules require high-level electronic-structure theory in combination with rigorous methods for approximating the quantum dynamics. Machine-learning approaches can significantly reduce the computational expense of this workflow without any loss of accuracy. We discuss various methods for constructing potential energy surfaces including transfer learning, which requires a minimal number of expensive training points. In this way, we can study chemical reactions at a high level but a low cost. In particular, as the potentials are smooth and differentiable, they enable the use of more advanced semiclassical approximations to quantum dynamics, such as perturbatively corrected instanton theory, which can capture both tunnelling and anharmonicity.