Predicting Young's modulus of oxide glasses with sparse datasets using machine learning

Abstract

Machine learning (ML) methods are becoming popular tools for the prediction and design of novel materials. In particular, neural network (NN) is a promising ML method, which can be used to identify hidden trends in the data. However, these methods rely on large datasets and often exhibit overfitting when used with sparse dataset. Further, assessing the uncertainty in predictions for a new dataset or an extrapolation of the present dataset is challenging. Herein, using Gaussian process regression (GPR), we predict Young's modulus for silicate glasses having sparse dataset. We show that GPR significantly outperforms NN for sparse dataset, while ensuring no overfitting. Further, thanks to the nonparametric nature, GPR provides quantitative bounds for the reliability of predictions while extrapolating. Overall, GPR presents an advanced ML methodology for accelerating the development of novel functional materials such as glasses.

Scalable Gaussian processes for predicting the optical, physical, thermal, and mechanical properties of inorganic glasses with large datasets

“…Development of reliable composition-property maps for a large class of glass components is the bottleneck impeding the design of new glass compositions. Machine learning (ML) methods [3][4][5][6][7][8] have been used to predict properties such as Young's modulus 9,10 , liquidus temperature 11 , solubility 12 , glass transition temperature 4,13 , dissolution kinetics 5,[14][15][16] , and other properties 17,18 . Most of these works employ traditional glass compositions as descriptors, while some other works employ physics-based descriptors 14,19,20 .…”

“…GPR employs a probabilistic approach which makes the inference on new data by learning the underlying distribution (mean and covariance) of the available data 22 . Note that various problems in mechanics and materials science employ a probabilistic framework (including GPR and Bayesian inference) to estimate material parameters 9,[23][24][25][26][27] . It has been shown that for small datasets, GPR models are more suitable in comparison to NN models for providing accurate composition-property predictions along with its confidence intervals in oxide glasses 9 .…”

Predicting densities and elastic moduli of SiO2-based glasses by machine learning

“…In this context, chemical compositions are straightforwardly used as one type of the most common descriptors as they are usually unique for each modeled material, and many material properties are eventually compositional dependent. In fact, several recent works have shown that using chemical compositions only as descriptors can describe the glass properties through the artificial neural network based ML algorithm [20][21][22]52 . However, only using compositional descriptors could make the model have limited extrapolative ability 13,24,26 .…”

Elucidating the constitutive relationship of calcium–silicate–hydrate gel using high throughput reactive molecular simulations and machine learning

“…In this section, predictions based on GP regression (see "Methods" section) are the focus. Two kernels i.e., radial basis function (RBF) and Matern kernels which are commonly adopted in the literature and also have been shown to produce accurate results 38 are implemented here. Figure 5 shows the predicted elastic constants using GPR with rbf kernel against the measured values computed by MD simulation.…”