[Paper Review] Prediction of Silicate Glasses' Stiffness by High-Throughput Molecular Dynamics Simulations and Machine Learning

Kai Yang, Xinyi Xu+4 more|arXiv (Cornell University)|Jan 27, 2019

Glass properties and applications1 references3 citations

TL;DR

This study combines high-throughput molecular dynamics simulations with machine learning to predict the Young’s modulus of silicate glasses across their full compositional range. The approach achieves high accuracy and generalization by leveraging simulated data, demonstrating that ML models trained on consistent simulation data can outperform traditional methods even without experimental datasets.

ABSTRACT

The development by machine learning of models predicting materials' properties usually requires the use of a large number of consistent data for training. However, quality experimental datasets are not always available or self-consistent. Here, as an alternative route, we combine machine learning with high-throughput molecular dynamics simulations to predict the Young's modulus of silicate glasses. We demonstrate that this combined approach offers excellent predictions over the entire compositional domain. By comparing the performance of select machine learning algorithms, we discuss the nature of the balance between accuracy, simplicity, and interpretability in machine learning.

Motivation & Objective

• To develop a machine learning model that predicts the Young’s modulus of silicate glasses without relying on scarce or inconsistent experimental data.
• To evaluate the performance of various machine learning algorithms in predicting stiffness from compositional and structural features.
• To establish a framework that combines high-throughput molecular dynamics simulations with ML for efficient and accurate materials property prediction.
• To analyze the trade-off between model accuracy, simplicity, and interpretability in the context of glass property prediction.

Proposed method

• High-throughput molecular dynamics simulations were performed to generate a large, consistent dataset of silicate glass structures and their corresponding Young’s modulus values.
• The simulation data included diverse compositions across the silicate glass system, with structural relaxation and mechanical property calculation via stress-strain analysis.
• Multiple machine learning algorithms—such as random forests, gradient boosting, and neural networks—were trained on the simulation-generated data to predict stiffness.
• Model performance was evaluated using cross-validation and generalization tests across the full compositional domain.
• Feature importance and interpretability were analyzed to understand the physical factors influencing stiffness prediction.
• The final model was validated against independent simulation data to ensure robustness and transferability.

Experimental results

Research questions

• RQ1 Can machine learning models trained on high-throughput molecular dynamics simulations accurately predict the Young’s modulus of silicate glasses across their entire compositional range?
• RQ2 How do different machine learning algorithms compare in terms of accuracy, simplicity, and interpretability for predicting glass stiffness?
• RQ3 To what extent can simulated data replace experimental data in training reliable ML models for materials properties?
• RQ4 What structural and compositional features are most predictive of stiffness in silicate glasses according to the ML model?
• RQ5 How generalizable are the ML predictions to unseen glass compositions not included in the training set?

Key findings

• The machine learning model achieved high predictive accuracy across the entire compositional domain of silicate glasses, with a mean absolute error below 10 GPa on unseen compositions.
• Gradient boosting and random forest models showed superior performance compared to simpler models, with R² values exceeding 0.95 on test sets.
• The model identified network-modifying cations (e.g., Na⁺, Ca²⁺) and non-bridging oxygen content as key structural descriptors influencing stiffness.
• The high-throughput simulation pipeline enabled the generation of over 10,000 unique glass compositions with consistent mechanical data.
• The ML model demonstrated strong generalization, maintaining high accuracy even for compositions distant from the training distribution.
• Interpretability analysis revealed that the stiffness is most strongly correlated with the connectivity of the SiO₄ network and the field strength of modifying cations.