Quantum Dynamics of Water from Møller-Plesset Perturbation Theory via a Neural Network Potential

Abstract

We report the static and dynamical properties of liquid water at the level of second-order Møller-Plesset per- perturbation theory (MP2) with classical and quantum nuclear dynamics using a neural network potential. We examined the temperature-dependent radial distribution functions, diffusion, and vibrational dynamics. MP2 theory predicts over-structured liquid water as well as a lower diffusion coefficient at ambient conditions compared to experiments, which may be attributed to the incomplete basis set. A better agreement with experimental structural properties and the diffusion constant are observed at an elevated temperature of 340 K from our simulations. Although the high-level electronic structure calculations are expensive, training a neural network potential requires only a few thousand frames. The approach is promising as it involves modest human effort and is straightforwardly extensible to other simple liquids.

Toward High-level Machine Learning Potential for Water Based on Quantum Fragmentation and Neural Networks

“…113 This direct comparison indicates that the NQEs would adjust the local chemical environment of water molecules, resulting in a less ordered structure. Lan et al 94 reported q values of 0.617 from classical simulation and 0.615 from quantum simulation, respectively, at 300 K using an NNP trained at the MP2 level, which slightly deviate from our present DP-MP2 potential results. Such deviations may arise from their small-sized training system (64 water molecules), significantly fewer number of reference structures (∼2000), and their choice of basis set (triple-zeta quality correlationconsistent basis sets).…”

“…This was also demonstrated in Lan et al's study based on the NNP trained from MP2, which predicted substantially smaller diffusion constants (0.0693 and 0.106 Å 2 /ps for classical and quantum simulations at 300 K, respectively). 94 The interaction PESs for a water dimer calculated at the MP2 and DFT levels of theory are shown in Figure S8 of the Supporting Information with reference to the CCSD(T) result. Despite the relatively higher accuracy of some particular density functionals, like ωB97M-V 120 and SCAN, 121 most of the currently available functionals are prone to self-interaction and delocalization errors, which has inspired Paesani and coworkers to use the density-corrected SCAN functional for water simulations.…”

“…These observations are in agreement with the previous studies. 31,94 Figure 4. Temperature dependence of the diffusion constant from the MD and CMD simulations under the NVE ensemble using the DP-MP2 potential in comparison to the experiment.…”

“…Lan and co-workers previously trained Behler−Parrinello NNPs for water at the MP2 level through the conventional QM calculations on relatively limited reference structures. 93,94 Herein, the EE-GMF method allows us to cover the representative bulk-size water structures over a broad range of phase space, allowing for realistic simulations of water under diverse thermodynamic conditions. The DP-MP2 potential also allows efficient MD simulation with the nuclear quantum effects (NQEs), which have been demonstrated to play an important role in characterizing the hydrogen-bonded structures.…”

Data-Efficient Machine Learning Potentials from Transfer Learning of Periodic Correlated Electronic Structure Methods: Liquid Water at AFQMC, CCSD, and CCSD(T) Accuracy

“…Machine-learned potentials (MLPs) have emerged as an extremely promising approach to accurately model ab initio potential energy surfaces of condensed-phase systems while being orders of magnitude more computationally efficient to evaluate. For liquid water, MLPs have been successfully developed at various levels of electronic structure ranging from different levels of DFT − to, more recently, using the random phase approximation (RPA) and MP2. , The modeling of liquid water and other molecular systems with more accurate electronic structure methods, such as coupled-cluster theory or quantum Monte Carlo, has been limited, so far, to training on finite clusters of molecules. − When training on small clusters, higher-order many-body interactions must be included by other means such as by using the TTM4-F potential, as is done for the MB-Pol water model. − Other cluster-based models for water have gone on to explicitly include 4-body terms − and also train on larger water clusters . MLPs fit to periodic electronic structure offer the opportunity to readily capture many-body electronic structure effects, since these are naturally included in the electronic structure calculation.…”