Presentation Information
[9a-E310-4]Autonomous Scanning Probe Microscopy via Integration of Domain-Specialized Small Language Models and AI Measurement Tools
Zhuo Diao1, Kouma Matsumoto1, Linfeng Hou1, Masahiro Ohara1, Hayato Yamashita2, 〇Masayuki Abe1 (1.Osaka Univ., 2.Shizuoka Univ.)
Keywords:
scanning probe microscopy,autonomous experiment,large language model
Autonomous operation of scanning probe microscopy (SPM) is key to improving throughput and reproducibility. Atomic-resolution SPM demands deterministic execution, low latency, and domain adaptation, which probabilistic cloud-based large language models cannot reliably provide. Here we present a framework in which small language models (SLMs), fine-tuned on domain-specific data and run locally on a consumer GPU, are integrated with AI measurement tools to autonomously plan and execute measurements. Three SLM roles (Router, Knowledge-base, Command) were realized within 16 GB of GPU memory by dynamically injecting LoRA adapters into a 4-bit quantized base model (Phi-4). Command-generation accuracy reached 99.3% for explicit instructions and 95.2% for goal-level planning, outperforming a cloud model. From a high-level instruction, the system autonomously performed tip conditioning and drift correction, achieving room-temperature atomic-resolution imaging of Si(111)-(7x7) in 35 minutes.