AstroBind: Machine learning prediction of binding energy distributions on interstellar water ice from geometric surface descriptors

Abstract

Binding energies on amorphous solid water regulate molecular desorption in molecular clouds and protoplanetary disks, but computing them from first principles across the heterogeneous binding sites of disordered ice surfaces is computationally prohibitive. We aim to develop a rapid, interpretable method for predicting binding energies on amorphous solid water that captures surface heterogeneity without requiring explicit electronic-structure calculations. We trained a machine-learning model using 27 geometric descriptors of the local ice-adsorbate environment. The model was trained on binding energies for 13 adsorbates on amorphous solid water clusters and evaluated through pooled leave-one-cluster-out validation across 15 ice surfaces. Its transferability was assessed on radical species excluded entirely from training. The model achieves a pooled leave-one-cluster-out coefficient of determination $R^2=0.90$ (the fraction of variance in the binding energies captured by the model, where 1 is a perfect fit), and a mean absolute error of 378 K across 15 ice surfaces. It also transfers to radical species withheld from training, with a pooled $R^2=0.79$, although its accuracy decreases when the unpaired electron contributes directly to the surface interaction. Our results demonstrate a route towards rapid, distribution-aware parameterisation of binding energies and desorption/diffusion rates in gas-grain astrochemical models, while preserving physical interpretability on heterogeneous ice surfaces.

Submitted to The Astrophysical Journal. Expressions of interest from potential reviewers with expertise at the intersection of machine learning and astrochemistry are welcome. All code and data will be publicly released upon acceptance