Title: PFT: Phonon Fine-tuning for Machine Learned Interatomic Potentials

Abstract:

While numerous material characteristics hinge on higher-order derivatives of the potential energy surface, machine-learned interatomic potentials (MLIPs) that rely on conventional training losses—typically focused on energy, force, and stress residuals—often suffer from inaccuracies in curvature. These errors can significantly compromise the reliability of vibrational property predictions. To address this, we propose phonon fine-tuning (PFT), a method that directly oversees the second-order force constants of materials. This is achieved by aligning the energy Hessians generated by the MLIP with the force constants derived from finite-displacement phonon calculations performed via density functional theory (DFT).

To ensure scalability for large supercells, PFT employs stochastic sampling of Hessian columns, calculating the loss through a single Hessian-vector product. Additionally, we implement a straightforward co-training approach that integrates upstream data to prevent catastrophic forgetting. When evaluated on the MDR Phonon benchmark, PFT boosts the performance of Nequix MP by an average of 55% across various phonon thermodynamic properties, establishing a new state-of-the-art accuracy for models trained on Materials Project trajectories. Furthermore, PFT demonstrates robust generalization capabilities, enhancing predictions for properties dependent on derivatives beyond the second order, such as thermal conductivity, which relies on third-order derivatives of the potential energy.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC