A deep learning strategy to calibrate heteroatomic interactions in metal alloys

A machine-learning-assisted strategy is proposed to calibrate the heteronuclear parameters of the Tight-Binding Second-Moment Approximation (TB–SMA) potential using finite-temperature experimental data. The method involves the use of neural-network surrogate models trained on a large dataset of fictitious binary alloys, generated by randomly sampling TB–SMA parameter sets within physically meaningful intervals. Each surrogate model learns to predict thermodynamic observables — mixing enthalpy and lattice parameter — directly from the potential parameters. Once trained, the networks provide instantaneous predictions, eliminating the need for costly simulations during the optimization loop. The surrogate models are then embedded in a minimization scheme that adjusts the mixed interaction parameters to reproduce experimental thermodynamic data at selected compositions and at given temperatures. This workflow is applied to ten binary alloys formed by Cu, Ni, Pt, Pd, and Rh, obtaining parametrizations that accurately match experimental trends. The approach is general and well adapted to complex multi-element systems as high-entropy alloys. It can be extended to other potential forms and target properties.