Surrogate Models for Vibrational Entropy Based on a Spatial Decomposition

Tina Torabi, Christoph Ortner, Yangshuai Wang

TL;DR

This paper introduces a novel machine learning approach to efficiently predict vibrational entropy in crystalline materials. By decomposing entropy into local "site entropies" and developing surrogate models using the Atomic Cluster Expansion (ACE) framework, they significantly reduce computational costs while maintaining accuracy, paving the way for more efficient materials simulations.

What Problem Does This Paper Address?

Calculating vibrational entropy, a crucial property for understanding material behavior like defect formation and diffusion, is computationally very expensive. Traditional methods involve calculating and diagonalizing the dynamical matrix, which scales poorly with system size ($O(N^3)$). This intractability limits the application of these methods to large-scale simulations. While machine learning has been used to predict other material properties, a rigorous framework for predicting vibrational entropy, especially by leveraging its inherent locality, has been missing. This paper aims to bridge this gap by developing a theoretically grounded and computationally efficient surrogate model for vibrational entropy.

Methods

The core of the paper lies in two key methodological contributions:

  1. Theoretical Framework for Site Entropy: The authors build upon previous work to rigorously decompose the total vibrational entropy into contributions from individual atomic sites, termed "site entropies." They establish a theoretical foundation for the locality of these site entropies, meaning that the contribution of a single site to the total entropy is primarily influenced by its immediate atomic environment. This locality is mathematically proven and numerically validated.
  2. Surrogate Modeling with ACE: Leveraging the proven locality of site entropies, the authors employ the Atomic Cluster Expansion (ACE) framework. ACE is a powerful machine learning method for constructing interatomic potentials and, in this work, is adapted to learn surrogate models for site entropies. They train ACE models to predict site entropies directly from atomic configurations, significantly reducing the computational cost from $O(N^3)$ to approximately $O(N)$.

Datasets

The authors use simulated datasets for training and testing their models. Specifically:

  • Silicon Supercells: They generate configurations of silicon supercells with varying numbers of atoms (e.g., 32 atoms) and introduce defects like vacancies. These configurations are "rattled" by perturbing atomic positions to create diverse structural environments.
  • Copper System: For predicting attempt frequencies, they use a copper system with a vacancy.

The paper does not introduce new benchmark datasets but rather utilizes generated data based on established simulation practices.

Key Insights and Ideas

  • Entropy is Local: The most significant insight is the rigorous demonstration that vibrational entropy, when decomposed into site contributions, exhibits strong locality. This means that the entropy contribution of an atom is predominantly determined by its local atomic environment, not the entire crystal. This is a crucial theoretical justification for using localized machine learning models.
  • Site Entropy as a Building Block: The paper establishes that site entropies can serve as fundamental building blocks for predicting total entropy. This decomposition simplifies the problem and allows for more efficient modeling.
  • ACE for Entropy Prediction: The successful application of the Atomic Cluster Expansion (ACE) framework to predict site entropies demonstrates its versatility beyond interatomic potentials. ACE's ability to capture local atomic environments makes it well-suited for this task.
  • Computational Efficiency Gain: By moving from $O(N^3)$ calculations to $O(N)$ surrogate models, the paper highlights a substantial computational speedup, making entropy calculations feasible for much larger and more complex systems.

Key Results

  • Locality Proof and Validation: Theorems 3.1 and 3.2 mathematically prove and numerical simulations (Figures 2, 3, and 4) validate the algebraic decay of site entropy sensitivity with distance.
  • Accurate Surrogate Models: The developed ACE surrogate models achieve high accuracy in predicting both total and site entropies. For instance, RMSEs for total entropy on a test set were on the order of $10^{-4}$ to $10^{-3}$ kB.
  • Prediction of Attempt Frequency: The models were successfully applied to predict the entropic term in harmonic transition state theory (attempt frequency) for vacancy migration in Copper, demonstrating practical applicability. The RMSE for this prediction was around 2.341 Hz.
  • Comparison of Training Strategies: The study shows that training directly on site entropies can be more complex but potentially more accurate than training on total entropy, especially when considering the richness of derivative data available for site entropies.

Limitations and Future Work

  • Focus on Simple Systems: The current study primarily focuses on point defects in relatively simple crystalline structures. The authors acknowledge that extending the methodology to more complex defect geometries and materials is a natural next step.
  • Harmonic Approximation: The entire framework relies on the harmonic approximation of vibrational entropy. While this is a standard and often reliable approximation, future work could explore incorporating anharmonic effects.
  • Computational Cost of Derivatives: While the surrogate models are efficient, calculating the reference data (e.g., Hessians for training) can still be computationally demanding. The paper mentions efforts to optimize derivative calculations using techniques like sparse matrix methods.
  • Generalization to Other Properties: The success in predicting entropy suggests potential for applying similar surrogate modeling techniques to other entropy-dependent properties or even other material properties that exhibit locality.

Summary updated

BibTeX
@article{torabi2025surrogate,
  author = {Tina Torabi and Christoph Ortner and Yangshuai Wang},
  title = {{Surrogate Models for Vibrational Entropy Based on a Spatial Decomposition}},
  journal = {Multiscale Model. Simul.},
  year = {2025},
  doi = {10.1137/24m165168x},
}
Figure