Domain Knowledge Fine-Tuned Large Language Model Driven Approach for Designing Novel Al-Based bulk metallic glasses with High Glass-Forming Ability
Abstract: Al-based bulk metallic glasses (BMGs) exhibit promising application prospects in lightweight structural components, yet their limited glass-forming ability (GFA), with critical casting diameters typically constrained to ≤Φ2.5 mm, remains a fundamental barrier to industrial adoption. To address this bottleneck, we curated a domain-specific dataset comprising 636 experimentally validated Al-based BMG compositions and associated GFA metrics extracted from 503 peer-reviewed publications, and generated 32,443 structured question–answer pairs spanning both fact-based and mechanism-informed queries. Using this dataset, we performed LoRA-based fine-tuning of the Qwen2.5-7B-Instruct large language model. The fine-tuned model achieves an R² of 0.938 in quantitative GFA regression, substantially outperforming general-purpose and academic LLM baselines including Bohr Academic AI, Qinyan Academic AI, Claude 3, and the unmodified Qwen2.5-7B-Instruct. Guided by the model's top-ranked predictions, six novel Al-based BMG compositions were proposed; five were experimentally verified to form fully amorphous Φ3 mm rods via vacuum suction casting, while one exhibited minor crystallization. The five validated alloys demonstrate improved thermal stability with supercooled liquid region widths (ΔTx) ranging from 55 to 80 K. This work establishes a domain-specific LLM fine-tuning framework as a reliable and efficient pathway for accelerating the compositional design of high-GFA Al-based BMGs.
Keywords: Large language model, Bulk Metallic Glasses, Glass-Forming Ability
