Learning interpretable surface elasticity properties from bulk properties via neural network equation learners
Published in International Journal of Mechanical Sciences, 111218, 2026
Abstract:
This work introduces a neural network-based equation learner designed to discover closed-form expressions for surface elasticity properties, combining customized activation functions and connection-based pruning to discover parsimonious, closed-form equations. When applied to seven face-centered cubic metals, the approach reveals interpretable relationships describing both low and high Miller index surfaces. The researchers found that surface elasticity properties decompose into a universal, geometry-driven orientation function, and material-specific baseline coefficients. A key finding is the differential behavior across property hierarchies: lower-order properties like surface tension demonstrate primarily geometric dependence, while higher-order properties such as surface stress exhibit more intricate geometry-material interactions. The work demonstrates how neurosymbolic machine learning can effectively bridge atomistic simulations with physical laws, enabling discovery of generalizable structure-property relationships in materials science.