Phonons & anharmonicity from MLIPs

Once a machine-learned potential reproduces DFT forces over the relevant region of configuration space, it becomes a near-free force engine for lattice dynamics. Two things follow that DFT cannot cheaply give: harmonic force constants at large supercell size, and the finite-temperature anharmonic sampling that bare DFT phonons skip entirely. The catch is a DC-DFT poisoning of the force constants that no energy metric will reveal.

Harmonic force constants from an MLIP

The MLIP supplies forces ${\mathbf{F}}_i=-{\nabla }_{i}E$ analytically, so the second-order force constants

$${\Phi }_{i\alpha ,j\beta }=\frac{{\partial }^{2}E}{\partial u_{i\alpha }\partial u_{j\beta }}=-\frac{\partial F_{i\alpha }}{\partial u_{j\beta }}$$

come either by the same finite-displacement construction run on the potential (now cheap enough that the supercell can be made large enough for the force constants to decay, the constraint that bites in DFT), or by automatic differentiation through the model where the energy is analytic. Either way the dynamical matrix and dispersion follow as usual, and the ASR must still be re-imposed on the assembled $\Phi$. The MLIP’s advantage is exactly the regime where frozen-phonon DFT is expensive: low-symmetry cells, defects, large $\mathbf{q}$-meshes, and especially the displacement amplitudes needed to probe anharmonicity.

Anharmonic sampling

The harmonic approximation truncates $E$ at second order in displacement; real crystals, and any soft-mode system, need the cubic and quartic terms. With an MLIP the higher-order force constants can be fit from forces on displaced/rattled configurations, or the anharmonic free energy can be obtained by direct statistical sampling — thermostatted MD or, better for soft modes, a stochastic self-consistent renormalization that dresses the phonons with their own anharmonic interaction at temperature $T$. The MLIP is what makes the thousands-to-millions of force evaluations such sampling needs affordable. This is the bridge to genuinely anharmonic, temperature-dependent phonons.

The DC-DFT caveat — force constants poisoned invisibly

This is the load-bearing warning. The MLIP is trained on reference DFT labels; if the reference functional is density-sensitive for this system, its self-consistent density is wrong, and so are the reference forces. The potential then learns a curvature — the second derivative of the surface — that is systematically off, and the error lands squarely in ${\Phi }_{i\alpha ,j\beta }$, i.e. in the phonon frequencies and in whether a mode is soft. Two features make this nasty:

1. It is invisible to energy-only metrics. Energy RMSE can be excellent while the gradient field that sets the force constants is corrupted; curvature is a second derivative and need not track the energy fit at all.
2. It masquerades as physics. A density-driven curvature error can push a marginal mode imaginary (or stabilize one that should be soft), reproducing exactly the signature one would attribute to a real instability — so it pollutes the imaginary-mode diagnostic.

The mitigation is upstream, at the data: diagnose density sensitivity and, where present, label the reference set with a density-corrected functional (HF-DFT) before fitting, per density-driven error (DC-DFT). An MLIP cannot repair a defect baked into its training forces; force-constant validation against directly-computed DFPT or frozen-phonon references on a few configurations is the only honest check.

Prerequisites

machine-learned potentials · finite-displacement phonons · density-driven error (DC-DFT)

Builds toward: anharmonic phonons & SSCHA · imaginary modes: artifact vs. instability

Key references

• MLIP phonon workflows — A. Togo & I. Tanaka, Scr. Mater. 108, 1 (2015) (phonopy); J. George, G. Hautier, A. P. Bartók, G. Csányi & V. L. Deringer, J. Chem. Phys. 153, 044104 (2020).
• Anharmonicity & self-consistent renormalization — I. Errea, M. Calandra & F. Mauri, PRB 89, 064302 (2014); review: L. Monacelli et al., J. Phys.: Condens. Matter 33, 363001 (2021) (SSCHA); O. Hellman, I. A. Abrikosov & S. I. Simak, PRB 84, 180301 (2011) (TDEP).
• Density-driven error in the reference forces — M.-C. Kim, E. Sim & K. Burke, PRL 111, 073003 (2013).