(PHYMOL) Physics, Accuracy and Machine Learning: Towards the next-generation of Molecular Potentials

The fundamental interactions between (neutral) molecules are relatively weak, but they determine much of the complex phenomena in solids, liquids, and gases. These intermolecular interactions are of paramount importance at interfaces, in molecular crystals, in cells, and even in interstellar gas clouds. These interactions are not easy to compute from first principles because the small size of these energies places extreme demands on the theoretical and numerical methods used. They are also often quite difficult to model (i.e. to construct an analytic, easily computable representation) accurately due to the subtle effects of anisotropy (atoms in a molecule are not spheres), many-body non-additivity (the whole is not the sum of its parts), and quantum tunneling (charge-transfer or delocalization). Simple models ignore, or average out many of these subtleties, but while these computationally simple models allow the study of large systems at long time-scales, they do so at the cost of accuracy and predictive power.

This is best exemplified in blind tests of organic molecular crystal structure prediction conducted by the Cambridge Crystallographic Data Centre, which have conclusively demonstrated that the most reliable predictions of the structures and free energy ranking of the molecular crystals is obtained by a combination of advanced theory-based models with ab initio methods such as density functional theory, and these combined approaches vastly outperformed the empirical models. There are tangible consequences for the increase in predictive power that arises from paying attention to physical details of the phenomena we aim to model: in the case of molecular crystals this leads to a better understanding of the stability of the crystalline material and its polymorphs. A profound understanding is the key to avoid the humanitarian and financial disasters that have arisen when one drug form transforms into another in an unexpected way, as it has happened with the antiretroviral medication Ritonavir.

Another application is transit transmission spectroscopy, where one observes absorption by an exoplanetary atmosphere as the planet transits between us and its sun. Here, collision-induced absorption (one of the highly sensitive ways in which we will assess reference data in PHYMOL) gives information on the atmospheric pressure. Of particular interest is the measurement of collision-induced absorption by O2, as evidence of an O2-rich atmosphere could be explained by photosynthesis, and therefore life, on the exoplanet.

Central to these applications is the PES - the potential energy surface - that needs to be constructed as an accurate mathematical model that is capable of accurately describing the intermolecular interactions as well as those within the molecular complexes, and also the couplings between these. Additionally, the PES must be computationally cheap to evaluate so as to allow us to simulate large systems for long time-scales. This is where we see the strong interlinking of physical models and machine learning: by combining the best of these two - in a sense, by combining human learning with machine learning - we will see the biggest advances in intermolecular model building for targeted applications.

In this proposal we seek to make the development of the PES easy in Human terms by mapping key parameters of the PES onto properties of the electronic density. The latter is relatively easily computable to a high accuracy using standard methods. Such a mapping could make PES development fast enough that they could be updated using the course of a molecular simulation. This would herald a new era in ab initio methods for computer simulations.