Abstract
Predicting response of materials to external stimuli is a primary objective of computational materials science. However, current methods are limited to small-scale simulations due to the unfavorable scaling of computational costs. Here, we implement an equivariant machine-learning framework where response properties stem from exact differential relationships between a generalized potential function and applied external fields. Focusing on responses to electric fields, the method predicts electric enthalpy, forces, polarization, Born charges, and polarizability within a unified model enforcing the full set of exact physical constraints, symmetries and conservation laws. Through application to $α$-SiO$_2$, we demonstrate that our approach can be used for predicting vibrational and dielectric properties of materials, and for conducting large-scale dynamics under arbitrary electric fields at unprecedented accuracy and scale. We apply our method to ferroelectric BaTiO$_3$ and capture the temperature-dependence and time evolution of hysteresis, revealing the underlying microscopic mechanisms of nucleation and growth that govern ferroelectric domain switching.
Introduction
In computational materials science, understanding how materials react to external stimuli like electric fields is essential for developing next-generation sensors, memory devices, and actuators. Traditionally, these properties are calculated using Density Functional Theory (DFT). While highly accurate, DFT's computational cost scales poorly with system size, restricting simulations to small, idealized models.
Machine Learning (ML) force fields have emerged as a powerful alternative, but most existing models treat dielectric properties (e.g., polarization) and mechanical properties (e.g., forces) as independent tasks. This disjointed approach often violates fundamental physical symmetries and conservation laws, such as energy conservation during dynamic processes or the acoustic sum rule required for charge neutrality.
Methods: Differentiable Learning Framework
The researchers addressed these limitations by developing a unified equivariant machine-learning framework implemented within the Allegro architecture. Instead of learning individual properties separately, the model learns a single generalized potential function—the electric enthalpy ($U$)—as a function of atomic positions ($r_i$) and the external electric field ($E$).
Key features of this approach include:
- Automatic Differentiation: Response properties are derived as exact derivatives of $U$. Polarization is the first derivative with respect to the field ($E$), while Born charges are mixed second derivatives with respect to both positions and the field.
- Sobolev Training: The model is trained on a loss function that simultaneously incorporates enthalpy, forces, polarization, and Born charges, ensuring the entire differential manifold of the material's response is captured.
- Equivariance: By utilizing an $O(3)$ equivariant representation, the model inherently respects rotational symmetries; for example, rotating a crystal correctly rotates its predicted dipole moment.
Key Results and Applications
The framework was validated on two benchmark materials: $\alpha$-SiO$_2$ (Quartz) and BaTiO$_3$ (Barium Titanate).
Vibrational and Dielectric Properties of $\alpha$-SiO$_2$
For $\alpha$-SiO$_2$, the model's predictions for infrared spectra and frequency-dependent dielectric constants showed excellent agreement with high-level Density Functional Perturbation Theory (DFPT) calculations. This demonstrates the model's ability to capture complex vibrational modes and their interaction with electric fields.
Ferroelectric Hysteresis in BaTiO$_3$
In the case of BaTiO$_3$, a classic ferroelectric material, the model successfully captured temperature-dependent hysteresis loops. The training involved active learning to sample diverse atomic configurations, including domain walls and phase transitions.
The model correctly predicted the decrease in the "coercive field" (the field strength required to flip polarization) as temperature increases. Furthermore, the efficiency of the ML approach allowed for simulations of supercells containing over 24,000 atoms—scales entirely inaccessible to standard first-principles methods.
Microscopic Insights and Future Directions
A significant advantage of this unified approach is its ability to reveal microscopic mechanisms. Simulations of ferroelectric switching in BaTiO$_3$ showed that the process is not a uniform bulk flip. Instead, it begins with the reversal of a single 5-atom unit cell, which then propagates as a 1D "needle" before expanding into a full domain.
By construction, the model satisfies physical constraints like the acoustic sum rule and ensures that polarization remains a conservative vector field. While the current model uses a local representation of atomic environments, it provides a robust foundation for future extensions. Potential future work includes bridging the gap between simulation and experimental timescales and extending the framework to other stimuli, such as magnetic fields or mechanical stress.
Summary updated