Fully First-Principles Surface Spectroscopy with Machine Learning

Abstract

Our current understanding of the structure and dynamics of aqueous interfaces at the molecular level has grown substantially due to the continuous development of surface-specific spectroscopies, such as vibrational sum-frequency generation (VSFG). As in other vibrational spectroscopies, we must turn to atomistic simulations to extract all of the information encoded in the VSFG spectra. The high computational cost associated with existing methods means that they have limitations in representing systems with complex electronic structure or in achieving statistical convergence. In this work, we combine high-dimensional neural network interatomic potentials and symmetry-adapted Gaussian process regression to overcome these constraints. We show that it is possible to model VSFG signals with fully ab initio accuracy using machine learning and illustrate the versatility of our approach on the water/air interface. Our strategy allows us to identify the main sources of theoretical inaccuracy and establish a clear pathway toward the modeling of surface-sensitive spectroscopy of complex interfaces.

First-principles spectroscopy of aqueous interfaces using machine-learned electronic and quantum nuclear effects

“…However, for SFG, which probes the water molecules at the interface, directly evaluating eqn (1) with the total first-principles polarization and polarizabilities of a water slab is problematic as it is a sum of two equal and opposite signals from its two interfaces. 51 While this problem may be alleviated using molecular dielectric responses 52 and by artificially mirroring the slab 35 so that the SFG signals no longer cancel, the molecular identity of atoms is invalid for reactive systems or exhibits artefacts in general. These challenges highlight the need for a general and uniquely-defined expression of SFG intensity.…”

“…Recent work in ref. 51 exploits that equivariant ML models predict polarization and polarizabilities as a sum of atomic contributions. This construct allows definitions of quasi molecular dielectric responses, which are sums of atomic components for the atoms in a molecule, as proxies for molecular polarizations and polarizabilities used to predict SFG intensities as has been done using forcefields.…”

“…Modelling SFG bands is extremely challenging as it requires long simulation statistics for molecules at the interface, atomic or molecular components of the polarization and the polarizability tensors with the environment dependence. It has, therefore, only been studied with force-elds like POLY2VS 35 and MB-pol 52 and more recently with MLPs, 18,51 with less accurate quantum dynamics methods.…”

Delta Machine Learning for Predicting Dielectric Properties and Raman Spectra

“…Hence, more recently, ML methods have also been applied to predict tensorial properties including polarizabilities, which currently is a very active research area . Both symmetry-adapted kernel-based methods − and neural-network approaches − have been used for this, as well as a physical based small parametric model . In addition, ML methods have also been applied to compute aspects of Raman spectra directly, without explicit consideration of polarizabilities. − Delta ML (Δ-ML) is a combined approach to predicting physical quantities: a computationally inexpensive approximation is used as a first step and ML methods are then applied to learn only the differences between first-step predictions and true values .…”

Neural Network-Based Sum-Frequency Generation Spectra of Pure and Acidified Water Interfaces with Air

“…Therefore, they cannot describe the change in molecular dipole and polarizability when a hydronium ion is converted into a water molecule upon proton transfer. Another set of methods based on novel machine-learning approaches have recently reproduced the IR, Raman, and SFG spectra of liquid water, − but they have so far only been developed on nonreactive systems.…”