Abstract

Recent second-harmonic scattering (SHS) experiments on electrolyte solutions have shown evidence of long-ranged intermolecular correlations between solvent molecules, persisting on the ~10 nm length scale at sub-molar concentrations; SHS is extremely sensitive to these kinds of correlations, and an atomistic understanding of these complex experiments is desirable. I will discuss work in our group towards a computational framework for modelling SHS experiments of general condensed-phase systems. Firstly, I will show that the orientational correlations between molecules must be accounted for to fully model the SHS signal, at variance with common assumptions. Accurate computational modelling of SHS experiments depends also on an accurate calculation of molecular hyperpolarizabilities: this requires that we take into account the effects of environmental and (nuclear) quantum-mechanical fluctuations. The calculation of these response tensors is computationally very demanding, and I will describe a simple machine-learning model that can sidestep these calculations for systems in which a molecular axis system can be defined unambiguously. In general systems, this method will not work, and so I discuss finally a framework that we have recently developed allowing for the prediction of any kind of tensorial property by machine-learning, taking into account the covariance of these properties under a rigid rotation. This method, called symmetry-adapted Gaussian process regression (SA-GPR), has the potential, among many other applications, to allow the simulation of light-scattering and spectroscopic experiments on general condensed phase systems, without the cost of quantum chemical methods.