Heat transport in water from Deep Neural Network potentials

The computation of heat transport coefficients from first principles has been made feasible by recent theoretical advances [1,2], but is still computationally extremely expensive, thus limiting the time and length scales of the simulations that one can afford. Neural-network (NN) inter-atomic potentials hold the promise of combining the accuracy and transferability of ab initio simulations with the affordability of classical potentials. We develop a NN implementation of the energy flux based on the Deep Potential--Smooth Edition (DeepPot-SE) NN model [3], which takes full advantage of the crucial role of gauge [1] and convective [2] invariances. Our methodology is first validated against DFT-PBE ab initio results for liquid water [1], and then used to compute the thermal conductivity from a potential trained on accurate DFT data obtained from the SCAN functional over a broad range of physical conditions.
[1] A. Marcolongo, P. Umari, S. Baroni, Nat. Phys. 12 (1), 80-84 (2016).
[2] R. Bertossa, F. Grasselli, L. Ercole and S. Baroni, Phys. Rev. Lett., 122, 255901 (2019).
[3] L. Zhang, J. Han, H. Wang, W. A. Saidi, R. Car, Weinan E, Adv. Neural Inf. Process Syst. 31, 4436-4446 (2018).