TEM-Scientist
Verified experimental protocols for agentic electron microscopy
TEM-Scientist is a compile–rehearse–execute–reflect architecture for autonomous transmission electron microscopy. Rather than placing a language model in the millisecond-scale instrument control loop — where its stochasticity, latency, and silent failures compromise experimental safety — TEM-Scientist moves the model above the loop: it compiles a natural-language experimental goal into a typed, verifiable protocol, rehearses that protocol against a physics-aware simulator under injected faults, and then executes it with a deterministic runtime that requires no further model involvement.
The architecture separates deliberation from protection, and detection from correction. The compiled protocol is a finite state machine over a closed set of instrument primitives, carrying declared dose and motion invariants and a table of edge-triggered reflexes; a two-layer verification gate (static lints plus behavioral rehearsal) admits only protocols that satisfy the frozen objective contract; and a runtime with frame-rate reflexes responds to safety-critical events five orders of magnitude faster than a per-step model round-trip. The compiled protocol is itself the reviewable artifact — reproducible, citable, and diffable.
Status: Active development · Preprint 2026
Role: First author and lead developer, in collaboration with Yuqing Huang (ANL/CNM), Rama Vasudevan (ORNL/CNMS), Yuzi Liu (ANL/CNM), and Subramanian Sankaranarayanan (UIC/ANL).
Key contributions
- Compile–rehearse–execute architecture — separates model deliberation from deterministic execution
- Objective contract — an immutable, verifiable specification of what the experiment must achieve
- TWINTEM behavioral simulator — physics-aware digital twin with Poisson-limited imaging, drift, defocus, cumulative beam damage, and fault injection
- Frame-rate reflexes — protective response with median 6.6 µs latency, against 2.1 s for per-step LLM control
- 114-task benchmark catalog with 25-task executable subset and a held-out locked evaluation suite
- Failure taxonomy and repair loop — legible, layer-localized failure classes with static and behavioral diagnostics
Links
- GitHub: sukritimanna/TEM-Scientist
Citation
@article{manna2026temscientist,
title = {Verified Experimental Protocols for Agentic Electron Microscopy},
author = {Manna, Sukriti and Huang, Yuqing and Vasudevan, Rama and
Liu, Yuzi and Sankaranarayanan, Subramanian K. R. S.},
journal = {arXiv preprint},
year = {2026},
note = {Under review}
}