Scalable Gaussian Processes for Predicting the Properties of Inorganic\n Glasses with Large Datasets

Suresh Bishnoi, R. Ravinder, Hargun Singh, Hariprasad Kodamana, N. M. Anoop KrishnanarXiv 2020 · DOI · arXiv

Abstract

Gaussian process regression (GPR) is a useful technique to predict composition--property relationships in glasses as the method inherently provides the standard deviation of the predictions. However, the technique remains restricted to small datasets due to the substantial computational cost associated with it. Here, using a scalable GPR algorithm, namely, kernel interpolation for scalable structured Gaussian processes (KISS-GP) along with massively scalable GP (MSGP), we develop composition--property models for inorganic glasses based on a large dataset with more than 100,000 glass compositions, 37 components, and nine important properties, namely, density, Young's, shear, and bulk moduli, thermal expansion coefficient, Vickers' hardness, refractive index, glass transition temperature, and liquidus temperature. Finally, to accelerate glass design, the models developed here are shared publicly as part of a package, namely, Python for Glass Genomics (PyGGi).

Introduction

Designing novel glass compositions is a critical bottleneck in materials science. While over 200,000 glass compositions are known, the complex relationships between their chemical makeup and physical properties are only partially understood. Machine learning (ML) has emerged as a powerful tool to bridge this gap. However, traditional deterministic models like Deep Neural Networks (DNNs) or Random Forests provide point predictions without quantifying confidence.

Gaussian Process Regression (GPR) is considered the "gold standard" for materials property prediction because it naturally provides a standard deviation for every prediction, serving as a built-in reliability metric. Historically, GPR has been limited to small datasets due to its $O(n^3)$ computational complexity. This work overcomes these limitations by implementing a scalable GPR framework capable of handling over 100,000 glass compositions.

Methodology

To achieve scalability, the researchers utilized two advanced algorithms:

  1. KISS-GP (Kernel Interpolation for Scalable Structured Gaussian Processes): Reduces computational complexity to $O(n)$ through structured kernel interpolation.
  2. MSGP (Massively Scalable Gaussian Processes): Further optimizes test-time predictions to $O(1)$ complexity and mitigates the "curse of dimensionality."

The models were built using the GPyTorch library with Radial Basis Function (RBF) kernels. The dataset was sourced from literature and the INTERGLAD Ver. 7 database, covering 37 key chemical components (oxides and halides) selected via Least Angle Regression (LARS). Nine target properties were modeled: density, elastic moduli (Young’s, shear, bulk), thermal expansion coefficient (TEC), Vickers’ hardness, refractive index, glass transition temperature ($T_g$), and liquidus temperature.

Key Results and Performance

The scalable GPR models demonstrated state-of-the-art accuracy, outperforming linear regression, XGBoost, and DNNs across all nine properties.

As shown in the parity plots, the models achieved exceptional precision, particularly for density ($R^2 = 0.97$) and refractive index ($R^2 = 0.96$). Even for complex mechanical properties like Young's Modulus, the GPR ($R^2 = 0.90$) outperformed deep learning alternatives ($R^2 = 0.86$).

Uncertainty Quantification and Glass Design

A unique advantage of this GPR approach is the ability to map predictive uncertainty across the compositional space. By visualizing the standard deviation in ternary diagrams, researchers can identify "unexplored" chemical regimes where the model is less confident.

This "uncertainty map" serves as a guide for experimentalists, highlighting regions where new data would most effectively improve the model. This active learning potential is a significant step forward from deterministic ML models, which may provide inaccurate predictions outside the "convex hull" of their training data without warning.

Conclusion and Open Science

The developed models are publicly available through the Python for Glass Genomics (PyGGi) package. While the models face challenges in extreme extrapolation and data-sparse regions, they provide a scalable, reliable framework for glass discovery. Future work will extend this methodology to other material classes, leveraging the power of scalable GPR to explore vast, uncharted compositional spaces.

Summary updated

Abstract

Gaussian process regression (GPR) is a useful technique to predict\ncomposition--property relationships in glasses as the method inherently\nprovides the standard deviation of the predictions. However, the technique\nremains restricted to small datasets due to the substantial computational cost\nassociated with it. Here, using a scalable GPR algorithm, namely, kernel\ninterpolation for scalable structured Gaussian processes (KISS-GP) along with\nmassively scalable GP (MSGP), we develop composition--property models for\ninorganic glasses based on a large dataset with more than 100,000 glass\ncompositions, 37 components, and nine important properties, namely, density,\nYoung's, shear, and bulk moduli, thermal expansion coefficient, Vickers'\nhardness, refractive index, glass transition temperature, and liquidus\ntemperature. Finally, to accelerate glass design, the models developed here are\nshared publicly as part of a package, namely, Python for Glass Genomics\n(PyGGi).\n
BibTeX
@misc{bishnoi2020scalable,
  author = {Suresh Bishnoi and R. Ravinder and Hargun Singh and Hariprasad Kodamana and N. M. Anoop Krishnan},
  title = {{Scalable Gaussian Processes for Predicting the Properties of Inorganic\textbackslash\{\}n Glasses with Large Datasets}},
  year = {2020},
  eprint = {2007.02795},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2007.02795},
}
Figure