Machine Learning Forcefield for Silicate Glasses

Han Liu, Zipeng Fu, Yipeng Li, Nazreen Farina Ahmad Sabri, Mathieu BauchyarXiv 2019 · DOI · arXiv

Abstract

Developing accurate, transferable, and computationally-efficient interatomic forcefields is key to facilitate the modeling of silicate glasses. However, the high number of forcefield parameters that need to be optimized render traditional parameterization methods poorly efficient or potentially subject to bias. Here, we present a new forcefield parameterization methodology based on ab initio molecular dynamics simulations, Gaussian process regression, and Bayesian optimization. By taking the example of glassy silica, we show that our methodology yields a new interatomic forcefield that offers an unprecedented description of the atomic structure of silica. This methodology offers a new route to efficiently parameterize new empirical interatomic forcefields for silicate glasses with very limited need for intuition.

Introduction

Simulating the atomic structure of glass presents a significant challenge in computational materials science. Researchers often face a choice between high-accuracy but computationally expensive ab initio molecular dynamics (AIMD) and fast but less accurate classical molecular dynamics (MD). Classical MD relies on empirical "forcefields"—mathematical models of atomic interactions—which require the optimization of numerous parameters, such as atomic charges and repulsion constants.

Traditionally, these parameters were tuned manually or via simple gradient-based methods. However, the parameter space for silicate glasses is notoriously complex and "rough," characterized by many local minima that trap traditional optimization algorithms. This paper introduces an automated machine learning pipeline designed to navigate this landscape and produce high-fidelity forcefields without the need for human intuition or biased starting points.

Methodology: An Automated ML Pipeline

The proposed framework utilizes a hybrid approach to optimize the parameters of a Buckingham-type potential. The process begins with high-fidelity AIMD reference data of liquid silica, which serves as the "ground truth" for atomic structure. Success is measured by a cost function ($R_\chi$) representing the difference between the structural fingerprints of the classical MD and the AIMD reference.

The core of the methodology is an iterative loop that combines Gaussian Process Regression (GPR) and Bayesian Optimization (BO). GPR acts as an interpolator, predicting the cost function across the entire parameter space while quantifying uncertainty. BO then uses these predictions to select the next most promising set of parameters to simulate, balancing the exploration of unknown regions with the exploitation of known low-error areas.

Methodological flowchart of the machine learning-based forcefield parameterization process.

Once the ML loop identifies the global minimum basin, a final Conjugate Gradient (CG) refinement step is applied to pinpoint the exact optimal parameter set.

The effectiveness of the ML loop is demonstrated by its ability to handle the "roughness" of the cost function. Unlike traditional methods that might get stuck in local ruts, the Bayesian approach systematically explores the landscape to find the true global minimum.

Illustration of (a) Gaussian Process Regression predicting the cost function with uncertainty and (b) the corresponding Expected Improvement (EI) used for Bayesian Optimization.

As shown in the figure above, the GPR model (a) fits known data points and provides a gray uncertainty band. The acquisition function (b) identifies the "best-tradeoff" point where the expected improvement is maximized, guiding the next simulation step in the cycle.

Key Results and Insights

The resulting ML-optimized forcefield demonstrates superior structural fidelity compared to industry standards like the BKS potential.

  • Superior Accuracy: The ML forcefield achieved a global cost function error ($R_\chi$) of 8.77%, a massive improvement over the standard BKS potential, which sits at 17.01%.
  • Structural Fidelity: The model matches AIMD reference peaks almost perfectly for Si–O and O–O partial distribution functions. It also accurately predicts 3-body correlations (bond angles), despite only being trained on 2-body data.
  • "Soft" Potentials: The optimization favored lower partial charges for Silicon (+1.955 vs. the traditional +2.4), suggesting that "softer" interactions better capture the disordered nature of glassy structures.

Conclusion

This research demonstrates that combining GPR and BO can effectively automate the design of forcefields for complex materials. The methodology is highly efficient, converging to the global minimum basin in just five iterations. While the current model is limited by the underlying Buckingham potential form, the framework itself is highly scalable and offers a new route for modeling multi-component glasses where human intuition often fails.

Summary updated

Abstract

Developing accurate, transferable, and computationally-efficient interatomic forcefields is key to facilitate the modeling of silicate glasses. However, the high number of forcefield parameters that need to be optimized render traditional parameterization methods poorly efficient or potentially subject to bias. Here, we present a new forcefield parameterization methodology based on ab initio molecular dynamics simulations, Gaussian process regression, and Bayesian optimization. By taking the example of glassy silica, we show that our methodology yields a new interatomic forcefield that offers an unprecedented description of the atomic structure of silica. This methodology offers a new route to efficiently parameterize new empirical interatomic forcefields for silicate glasses with very limited need for intuition.
BibTeX
@misc{liu2019machine,
  author = {Han Liu and Zipeng Fu and Yipeng Li and Nazreen Farina Ahmad Sabri and Mathieu Bauchy},
  title = {{Machine Learning Forcefield for Silicate Glasses}},
  year = {2019},
  eprint = {1902.03486},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1902.03486},
}
Figure