Overlap-Aware Clean-Region Prediction for Autonomous STEM

In autonomous atomic-resolution electron microscopy, selecting where to place the electron probe on a nanoparticle is as critical as selecting which particle to examine. When neighboring particles overlap in projection, diffraction acquired from the overlap region contains superimposed signals from multiple crystals, which can cause zone-axis alignment to fail and prevent successful two dimensional atomic resolution imaging of samples. Conventional instance-segmentation methods can separate particle identities, but they are not explicitly designed to identify the non-overlapping regions required for diffraction-safe probe placement. We therefore formulate clean-region prediction as an explicit perception task and introduce a semantics-conditioned flow model for overlapping nanoparticles. A semantic segmentation network is first trained to classify pixels as background, non-overlap, or overlap and is subsequently frozen as a fixed feature provider. Its predicted probability maps condition a Cellpose-style flow network for instance separation. Rather than constructing flow targets from complete amodal particle masks, we construct them directly from each particle's non-overlap region and restrict flow supervision to these uniquely owned pixels, avoiding ambiguous supervision in projected overlap regions. On a held-out test set of 30 regions of interest containing 253 nanoparticles, the resulting framework raises $F_1$ from 0.595 for a fine-tuned Cellpose baseline sharing the same backbone to 0.748, and performs comparably to a fine-tuned Cellpose-SAM with a larger transformer encoder (0.748 vs. 0.702) while achieving higher beam purity (beam purity 0.953 vs. 0.927). This framework provides an on-the-fly perception module for selecting diffraction-safe probe positions for autonomous atomic resolved STEM experiments.