Polarization Dynamics in Ferroelectrics: Insights Enabled by Machine Learning Molecular Dynamics

TL;DR

This paper reviews how machine learning-enhanced molecular dynamics simulations can provide atomic-level insights into ferroelectric polarization dynamics, overcoming traditional computational limitations for designing advanced ferroelectric devices.

Contribution

It systematically discusses recent advances and methodological challenges in applying machine learning molecular dynamics to ferroelectric materials, highlighting future directions for predictive modeling.

Findings

MLMD captures polarization switching and domain behaviors with near first-principles accuracy.

Advances in force fields and pretraining improve simulation predictive power.

Identifies key challenges like long-range electrostatics and multiferroic coupling.

Abstract

Ferroelectric materials with switchable spontaneous polarization underpin non-volatile memories, transistors, sensors, and emerging neuromorphic chips. Their performance and stability are governed by polarization dynamics and domain kinetics, making a microscopic understanding of these processes and precise atomic level control of polarization domains key challenges for next-generation ferroelectric electronics. Due to the limitations of the characterization technology with atomic level in experiment, high precision atomic simulations become important. First principles calculations are inherently limited in accessible length and time scales, making it difficult to capture the complex features of dynamic processes. Machine learning molecular dynamics (MLMD) offers a compelling solution by encoding quantum-mechanical accuracy into force fields, thereby enabling large scale dynamic simulations with near first-principles fidelity. This Perspective highlights the advantages of MLMD for simulating polarization switching, domain nucleation and migration, topological polar textures and curvature-driven ferroelectric phenomena, while providing a systematic overview of recent progress in these areas. We further discuss methodological challenges that limit predictive capability, including long range electrostatics, coupled lattice-spin degrees of freedom in multiferroics, and data efficient pre-training of large atomistic models. Corresponding advances in long range aware force fields, spin dependent machine learning models, and large scale pretraining are expected to move MLMD toward a genuinely predictive framework for the design of ferroelectric and multiferroic materials.