LLM-assisted microscopy: translating expert ideas into API workflows

Research applicability

A concrete research-coding pattern connecting human microscopy expertise to instrument APIs, with explicit limits on the model's domain reasoning. It does not establish a fully autonomous or safe microscope controller.

Getting started

Editorial, untested: read the full methods and relevant API documentation, examine generated code without connecting to hardware, and require instrument-specialist approval before any supervised non-hazardous test. No installation or hardware execution was attempted.

Maintenance

Sources accessed 2026-10-17; service availability, maintenance and compatibility were not tested.

License

Reuse permissions for paper, data and code require separate assessment; no blanket licence or redistribution assurance.

Workflow & human oversight

Resource-level workflow level intentionally unspecified. Conceptual references and reusable process patterns do not themselves establish an autonomous workflow; researchers retain task-specific interpretation and oversight.

Workflow level is not an evidence score; higher is not better evidence.

Approach

The authors report translating experimental workflow ideas into executable microscope-API code. They also report that GPT4 cannot move beyond basic analyses or supply deeper technical experiment design; domain-specific fine-tuning is proposed as a possibility, not demonstrated as a better model.

Limitations

ArXiv abstract-only review for the workflow description; official ORNL journal metadata additionally reviewed, not the journal methods or supplements. The official record identifies a peer-reviewed Machine Learning: Science and Technology article (2024), DOI 10.1088/2632-2153/ad52e9. Direct publisher retrieval hit a bot challenge; a web-extracted publisher excerpt was read but is not cited as complete methods evidence. No code, microscope, prompts, accuracy, runtime, safety outcome or fine-tuning result was reproduced. Generic image analysis must not be upgraded to expert diagnostic interpretation.

Reproducibility

AI-agent primary-source reading and provenance checks only. No software, experiment, hardware, outcome or installation reproduced. Source verification is not outcome verification.

Adoption requirements

The described workflow uses microscope-control APIs and human scientific expertise. Editorial, untested adoption also requires authorised instrument access, API documentation, code inspection, privacy safeguards and laboratory-specific safety controls; exact dependencies and costs were not established from the abstract.