Since ChatGPT’s release in 2022, general-purpose large language models (LLMs) have been increasingly adopted as AI assistants, though their effectiveness in specialized scientific domains remains limited. While domain-specific LLMs, such as BloombergGPT for finance or Med-PaLM 2 for healthcare, address these limitations, their training costs and data requirements remain substantial. This project proposes developing a compact, domain-specific language model tailored to porous acoustic metamaterials. Our method involves continual pre-training using targeted acoustics textbooks and research papers, followed by supervised fine-tuning with domain-specific question-answer pairs. To ensure accuracy and relevance, the model will utilize retrieval-augmented generation (RAG), dynamically accessing an external knowledge base of recent literature and technical resources. By reducing model size and training complexity, we aim to enable local deployment, enhancing privacy and allowing researchers to customize the model for specific subfields, such as foam-based acoustic absorbers or resonant structure optimization. This work demonstrates a practical approach to creating lightweight, domain-focused language assistants that support research and innovation in acoustic metamaterials.
Celeste et al. (2025) studied this question.