Learning an equation for copper

Varughese and colleagues combine symbolic regression with reinforcement-learning search to derive equations for interatomic interactions. Their January 2026 paper trains against density-functional-theory data and evaluates copper. The resulting models improve on the fixed-form Sutton–Chen potentials used in the comparison, including tests of material properties and melting behaviour. ¹

The comparison is specific: it does not establish superiority over every modern neural potential or over all elements. The evidence discussed here comes from computational modelling.

Our interest is in the opportunity to inspect the learned equation and challenge particular terms. A readable model still needs demanding tests outside its fitting examples. For follow-up coverage we would ask which omitted configuration changes its prediction most, and whether a simpler equation retains the performance that matters for the intended application.