Self-Supervised Learning for Glass Composition Screening

TL;DR

This work proposes a novel self-supervised learning framework designed specifically for screening glass compositions within pre-defined glass transition temperature ranges, and introduces an innovative data augmentation strategy grounded in asymptotic theory.

Abstract

Glass composition screening is essential for advancing new glass materials, yet the inherent complexity of multicomponent systems presents significant challenges. Current supervised learning methods for this task rely heavily on large amounts of high-quality data and are prone to overfitting on noisy samples, which limits their generalization ability. In this work, we propose a novel self-supervised learning framework designed specifically for screening glass compositions within pre-defined glass transition temperature (Tg) ranges. We reformulate the screening task as a classification problem, aiming to predict whether the glass transition temperature of a given composition falls within a target interval. To improve the model's robustness to noise, we introduce an innovative data augmentation strategy grounded in asymptotic theory. Additionally, we present DeepGlassNet, a dedicated network architecture developed to capture and analyze the complex interactions among constituent elements in glass compositions. Experimental results demonstrate that DeepGlassNet achieves superior screening accuracy compared to traditional methods and exhibits strong adaptability to other composition-related screening tasks. This study not only provides an efficient methodology for designing multicomponent glasses but also establishes a foundation for applying self-supervised learning in material discovery. Code and data are available at: https://github.com/liubin06/DeepGlassNet

Figures (12)

• Figure 1: Illustrative example of supervised learning and self-supervised learning. Supervised learning optimizes the consistency between model predictions and labels, while self-supervised learning maximizes representation consistency between augmented views of the same instance through pretext task.
• Figure 2: (a) Distributions of raw and cleaned dataset, (b) TSNE visualization of cleaned dataset: The x- and y-axes display two-dimensional embeddings obtained through TSNE dimensionality reduction of 18 glass compositional features, while the z-axis corresponds to the target property label T$_g$.
• Figure 3: Machine learning model architecture. The dimensions of each module are denoted by the following symbolic conventions: bs represents batch size, n denotes the number of compositional elements, emb_dim signifies embedding dimension, and feature_dim indicates output feature dimension.
• Figure 4: Illustrative Explanation of Graph Convolution(Demonstrated on four glass components labeled 1–4). (a) Component Graph. Nodes represent glass components. Edge weight $a_{ij}$ quantifies the interaction strength between components $i$ and $j$, $i,j \in \{1,2,3,4\}$. (b) Adjacency Matrix. Matrix representation of the component graph. Each element $a_{ij}$ in $\mathbf{A}$ explicitly encodes the pairwise interaction strength between component $i$ and component $j$. (c) Graph Convolution Mechanism for node 1: Neighboring nodes (2, 3, 4) propagate their input features $\mathbf{h}_2, \mathbf{h}_3, \mathbf{h}_4$ to node 1, weighted by interaction strengths $a_{21}, a_{31}$ and $a_{41}$. The output feature $\mathbf{z}_1$ is generated through weighted aggregation of these signals with node 1’s intrinsic feature $\mathbf{h}_1$.
• Figure 5: Self-supervised learning pipline.