betaFeedback

WorksPhysics and Astronomy

Fast closed-loop STEM navigation using vision-language models with typed decisions

Abstract

Closed-loop scanning transmission electron microscopy (STEM), in which each acquired image determines the next field of view and magnification, offers a route to autonomous atomic-scale materials analysis. Such a loop would automate navigation to defects and grain boundaries, a task that has long relied on operator judgment, and would enable systematic analysis of diverse local structures at scale. Realizing it hinges on the decision step, which must translate each acquired image into the next operation both quickly and correctly; yet which decision method can deliver both remains unclear. Here, we isolate this decision step. Using simulated atomic-resolution STEM images of a substitutional W dopant and a grain boundary in monolayer MoS2_2, where the true target position is known and every method starts from identical conditions, we compared four vision-language models (VLMs) and three decision methods. Our results demonstrate the effectiveness of Typed Decision, which directly obtains the probabilities of candidate labels from a pretrained VLM, for closed-loop STEM navigation. Typed Decision showed the shortest decision latency for all models (0.13-0.21 s) and generally achieved the highest navigation success rate among the three methods: it was the best for all models in W dopant navigation (66-91%) and for the best-performing model, Cosmos-Reason2 32B, in grain boundary navigation (43%). Furthermore, navigation to the different targets, the W dopant and the grain boundary, was possible by natural-language specification alone, without task-specific training of the VLM. These results indicate that Typed Decision is an effective decision method for realizing fast, accurate, and flexible closed-loop STEM.

Paper details

vlnc:W_yx9d74pt
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
arXiv nonexclusive license
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Physics and Astronomy

Not recorded for this paper

View statement
Looking for related papers…