Abstract
Designing new oxide glasses for applications ranging from smartphone screens to nuclear waste storage is a complex task due to the nearly infinite combinations of chemical compositions. Traditional physics-based models often fail to capture the non-linear relationships between composition and mechanical properties like Young’s modulus. While machine learning (ML) offers a potential solution, standard techniques like Neural Networks (NNs) frequently overfit on the small, "sparse" datasets typical in materials science. This study demonstrates that Gaussian Process Regression (GPR), particularly when paired with an Automatic Relevance Determination (ARD) kernel, provides a robust alternative, achieving high predictive accuracy ($R^2 > 0.9$) and essential uncertainty estimates for sparse glass datasets.
The Challenge of Sparse Data in Glass Science
The mechanical stiffness of glass, quantified by Young’s modulus, is driven by complex atomic arrangements that do not follow simple linear rules. While big data has made Neural Networks the "gold standard" for many ML tasks, materials science often operates in a data-poor regime. Experimental data is expensive and time-consuming to acquire, resulting in datasets with fewer than 100 samples.
In these "sparse" environments, NNs tend to overfit—memorizing the noise in the training data rather than learning the underlying physical trends. This leads to models that perform well on known data but fail catastrophically when presented with new, unseen glass compositions.
Methodology
Researchers compared three distinct modeling approaches across four silicate glass families (CAS, NCS, NGS, and NBS) sourced from the INTERGLAD database:
- Makishima-Mackenzie (MM) Model: A traditional physics-based analytical model assuming linear relationships.
- Neural Networks (NN): A parametric ML model where the number of hidden units was optimized to mitigate overfitting.
- Gaussian Process Regression (GPR): A non-parametric, probabilistic approach that assumes data follows an underlying Gaussian distribution.
A critical component of the GPR approach was the Automatic Relevance Determination (ARD) kernel. This allowed the model to automatically weight the importance of different oxide concentrations, effectively filtering out irrelevant features and focusing on the primary drivers of glass stiffness.
The Overfitting Trap
The study highlighted the inherent risks of using NNs for small datasets. For the Calcium Aluminosilicate (CAS) dataset, the NN's performance on the test set was highly sensitive to the model architecture. As shown in the figure below, while training accuracy continues to improve with more hidden units, the test accuracy peaks and then plummets, a classic sign of overfitting.
Results and Discussion
GPR significantly outperformed both the physics-based MM model and the optimized NNs. While the MM model struggled with non-linearities (achieving $R^2$ as low as 0.23 for borosilicates), GPR maintained high accuracy across all datasets.
For the CAS dataset, where the NN reached an $R^2$ of 0.59, the GPR model achieved an $R^2$ of 0.88. In the case of the Sodium Germanium Silicate (NGS) dataset, GPR reached a near-perfect $R^2$ of 0.999.
Beyond raw accuracy, GPR offers a unique advantage: quantifiable uncertainty. Unlike NNs, which provide a single point prediction, GPR provides a mean prediction along with a standard deviation. This "confidence interval" allows researchers to gauge the reliability of a prediction for a novel composition, which is invaluable for guiding experimental synthesis.
Conclusion
For the sparse datasets common in materials science, Gaussian Process Regression offers a superior balance of accuracy and reliability compared to traditional Neural Networks. By avoiding the overfitting trap and providing built-in uncertainty estimates, GPR serves as a powerful tool for the accelerated discovery of high-performance oxide glasses. Future work may extend this framework to other properties like chemical durability and glass transition temperature, utilizing "active learning" to strategically identify the most promising new compositions for laboratory testing.
Summary updated