Unpacking the 'Black Box': Making Neural Network Potentials Understandable

Machine learning has emerged as a powerful technique for modeling complex systems, particularly through the use of neural network potentials to define effective energy models. These potentials are invaluable for incorporating electronic structure effects at the atomistic resolution or for coarse-graining atomic degrees of freedom. However, a significant drawback has been their inherent 'black box' nature; unlike traditional, more transparent functional forms, the energy inferred by neural network potentials has been notoriously difficult to interpret, leading to skepticism within the scientific community.

Addressing this critical challenge, new research extends tools from the burgeoning field of explainable artificial intelligence (XAI) to graph neural networks used for coarse-grained potentials. This innovative approach enables the practical decomposition of neural network potentials into discrete n-body interactions. As lead author Bonneau notes in the paper, "With these tools, neural network potentials can be practically decomposed into n-body interactions, providing a human understandable interpretation without compromising predictive power." This breakthrough effectively opens up the 'black box' of these models, offering clarity without sacrificing the predictive accuracy that makes them so valuable.

The efficacy of this method was demonstrated across a range of coarse-grained systems, including two fluids (methane and water) and the protein NTL9. The obtained interpretations suggest that well-trained neural network potentials learn physical interactions that are consistent with fundamental principles. This work not only enhances the transparency and trustworthiness of machine-learned potentials but also solidifies their role as a powerful and interpretable tool in scientific discovery.