Case Studies of Materials Predictive Modeling

View at publisher Summarise with VIRAT
Authors: Xue Jiang, Yanjing Su
Cited by: 0
References: 66

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

This chapter explores the application of language models in predictive materials modeling, focusing on polymer property prediction. The polyBERT framework treats polymer SMILES (PSMILES) representations as a formal chemical language, leveraging a transformer model pretrained on a corpus of 100 million PSMILES strings. This approach generates dense numerical embeddings, or transformer fingerprints, which encode structural features without manual feature engineering. These embeddings serve as inputs for supervised models predicting a wide range of polymer properties, forming a scalable pipeline for ML-driven polymer informatics. The chapter details dataset construction, model architecture, fingerprint analysis, and predictive performance, highlighting the broader implications of language-augmented strategies in materials science. The polyBERT framework centers on a sophisticated chemical language model, based on the DeBERTa architecture, trained on 100 million PSMILES strings to interpret polymer structures. Canonicalization of input strings ensures consistent representation, addressing variability in PSMILES notation. The model undergoes self-supervised learning via masked language modeling, achieving high accuracy in reconstructing masked tokens. The resulting polyBERT fingerprints are dense, 600-dimensional representations, compared favorably with sparse, high-dimensional PG fingerprints. UMAP visualizations demonstrate that polyBERT fingerprints cluster chemically similar polymers effectively.

Fields of Study & Subjects

Machine Learning in Materials Science Computational Drug Discovery Methods History and advancements in chemistry

Publication details

Journal / source
Wiley eBooks
Pages
339–386
Publication date
17 March 2026
Language
EN
Document type
Book chapter