Machine learning interatomic potential can infer electrical response (2504.05169v1)

Abstract: Modeling the response of material and chemical systems to electric fields remains a longstanding challenge. Machine learning interatomic potentials (MLIPs) offer an efficient and scalable alternative to quantum mechanical methods but do not by themselves incorporate electrical response. Here, we show that polarization and Born effective charge (BEC) tensors can be directly extracted from long-range MLIPs within the Latent Ewald Summation (LES) framework, solely by learning from energy and force data. Using this approach, we predict the infrared spectra of bulk water under zero or finite external electric fields, ionic conductivities of high-pressure superionic ice, and the phase transition and hysteresis in ferroelectric PbTiO$_3$ perovskite. This work thus extends the capability of MLIPs to predict electrical response--without training on charges or polarization or BECs--and enables accurate modeling of electric-field-driven processes in diverse systems at scale.

Overview of Machine Learning Interatomic Potentials for Electrical Response Prediction

The paper presents a novel methodology to infer electrical response properties from machine learning interatomic potentials (MLIPs) using the Latent Ewald Summation (LES) framework. This research addresses a longstanding challenge in computational material science: modeling the interactions of material and chemical systems with electric fields accurately and efficiently. Traditional methods like density functional theory (DFT) are computationally expensive, especially for large systems or over extended timescales. This research introduces a MLIP approach that predicts electrical responses from energy and force data, bypassing the direct inclusion of charges and polarizations in the model.

Key Contributions and Findings

1. Latent Ewald Summation: The LES framework enables the extraction of polarization and Born effective charge (BEC) tensors directly from MLIPs. This allows for accurate prediction of electrical response properties such as infrared (IR) spectra and ionic conductivity. The model is trained only on energy and force data, inferring long-range interactions without explicitly considering electrostatics during training.
2. Application to Diverse Systems: The methodology was tested on various systems including bulk water, high-pressure superionic ice, and ferroelectric PbTiO$_3$ perovskite. It successfully predicted the electrical responses without being trained on polarization or BEC directly, showcasing its versatility across different materials and conditions.
3. Validation and Performance: The predicted BEC tensors were compared against DFT results, demonstrating strong agreement with established theoretical values. Specifically, for bulk water, the predicted IR spectra closely matched experimental data, covering both intermolecular and intramolecular vibrational modes. In superionic water, BEC predictions captured phase transitions and ionic conductivity accurately, validated against DFT molecular dynamics results.
4. Practical Implications: The inferred charges and the resulting BEC tensors allow for the application of an external electric field to MLIPs, enabling simulations of electric-field-driven phenomena. This advancement provides a scalable and efficient method to explore charge transfer and polarization processes in diverse material systems.

Theoretical and Practical Implications

• Theory and Computation:

This research bridges the gap between quantum mechanical predictions and machine learning models by embedding long-range electrostatics into MLIPs. This integration enhances the theoretical understanding of charge interactions and polarization in complex systems. The framework's robustness across different phases of matter and thermodynamic conditions underscores its potential to generalize beyond the training data, hinting at broader applicability for other material classes.

• Future Developments:

The success in extracting BECs and predicting electrical responses opens avenues for enhanced MLIP development. Future work could integrate more complex electrostatic interactions and non-linear responses, further extending the predictive capabilities of MLIPs. Additionally, the framework could be refined to include explicit training on polarization data, potentially increasing the model's accuracy and applicability.

• Application in Material Science:

The methodology enables new strategies for simulating and designing materials with optimized electrical properties, relevant to industries involving semiconductors, piezoelectrics, and electrolytes. By making large-scale simulations of electric-field-effects feasible, it could transform material discovery and design processes, providing insights that were previously unattainable with traditional computational methods.

In conclusion, this paper introduces a transformative approach to modeling material responses to electric fields using MLIPs. The ability to predict and simulate electrical phenomena at scale through the LES framework stands to significantly impact computational material science and associated fields.