Green-Kubo simulation of transport properties: from ab initio to neural-network molecular dynamics

Transport coefficients are an essential element for the understanding of many physical phenomena and technological applications, such as planetary evolution, energy saving, thermal dissipation, heat management in devices, the study of biological processes and chemical reactions. In the thesis, we compute the thermal conductivity and the viscosity of water within linear response theory from equilibrium molecular dynamics simulations. For both coefficients, we use two approaches: in one, the potential energy surface (PES) is derived on the fly from the electronic ground state of density functional theory (DFT); in the other, the PES is represented by a deep neural-network (DNN) trained on DFT data. Our machine learning approach proved to predict the transport properties of water with ab initio accuracy at a computational cost only slightly higher than the one of an empirical force field. With the neural-network potential, we could study the transport properties of water over a wide range of temperatures, adopting two popular DFT approximations: the PBE generalised gradient approximation (GGA) and strongly constrained and appropriately normed (SCAN) meta-GGA.