Fine‐tuning and Application for Materials Large Language Models

View at publisher Summarise with VIRAT
Authors: Xue Jiang, Yanjing Su
Cited by: 0
References: 90

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 pivotal role of language models, particularly large language models (LLMs), in materials science, focusing on their evolution, capabilities, and fine-tuning strategies. The chapter traces the progression from statistical language models to neural-based approaches, culminating in the advent of LLMs, which are characterized by their vast parameters and training data scales. Fine-tuning emerges as a critical technology for customizing LLMs to specific domains, such as materials research, enhancing their ability to predict material properties, suggesting synthesis routes, and analyzing experimental results. The chapter discusses the shift from transfer learning paradigms to models like GPT-3, which exhibit zero-shot generalization. It highlights the importance of fine-tuning on task-instruction datasets and aligning with human preferences to improve generalization and mitigate misalignment issues. The emergent abilities of LLMs, including reasoning and in-context learning, are attributed to their enormous scale. The chapter also addresses the tradeoffs and challenges associated with LLMs, such as slow training speeds and high operational costs, and discusses optimization strategies like parameter-efficient tuning and model quantization. Data preprocessing, including filtering and deduplication, is emphasized as crucial for maintaining data quality.

Fields of Study & Subjects

Machine Learning in Materials Science Topic Modeling Multimodal Machine Learning Applications

Publication details

Journal / source
Wiley eBooks
Pages
419–448
Publication date
17 March 2026
Language
EN
Document type
Book chapter