making things happen with electron microscopes🔬 she/her 🐘: https://xitter.cf/t.co/R0PzrAIen6 🦋: https://xitter.cf/t.co/q3MDRCDW1D
back focal plane (oregon)
Joined March 2022
- Tweets806
- Following1.9K
- Followers1.8K
- Likes15.9K
Isabel retweeted
Surface level bits of an Ion Implant ROM.
These are an absolute pain to decode optically, requiring delayering and a timed Dash etch (HF, Acetic, & HNO3) under halogen light. This procedure makes ones brown a little bit faster than it does zeroes.
Isabel retweeted
The CR-8 @crowd_supply page is live! Make sure to subscribe so you don't miss the campaign launch!
crowdsupply.com/dragon-labs/…
Isabel retweeted
Kirin 9030 has bad dishing at M2. Only layer I see with this from @SemiAnalysis_ breakdown. The CDU also looks somehow off for this layer.
Isabel retweeted
In 1911, Heike Kamerlingh Onnes, by virtue of being the first person to liquify helium, serendipitously discovered that the resistance of mercury dropped to zero below 4.2K–the first superconductor. For the next decades, some of the luminaries of 20th century physics tackled the superconductivity problem, without achieving the holy grail of a microscopic theory about why it happens. It was not until two smoking gun experiments in the 1950s pointed in the correct direction: isotope effect experiments in mercury showed that the atomic lattice, not just electrons, were implicated, and low-temperature specific heat experiments gave evidence for a small energy gap associated with the superconducting state. These key experiments were among thousands of valid works about superconductivity, and certainly no one knew what would be the smoking gun experiment in 1911 or even in 1941 (the author of the isotope effect paper even stated that their motivation was that they had different isotopes of mercury laying around).
With all the key observations in place, Bardeen, Cooper, and Shrieffer were able to formulate the BCS theory explaining what is now called the ‘conventional’ type of superconductivity.
This history illustrates the important role of serendipity and path-dependence in superconductivity research (as well as other areas dealing with complex many-body systems). If Onnes chose to measure aluminum instead of mercury, he would not have discovered superconductivity when he did; if someone had tried to FAFO superconductivity in different isotopes of a metal earlier, the mechanism of superconductivity could have been established much sooner.
Similarly if superconductivity was discovered in Ba1-xKxBiO3 (BKBO, Tc up to 34K) *before* it was discovered in La2-xBaxCuO4 (LBCO, Tc up to 35K, first cuprate discovered), humanity could have gone another generation without discovering cuprate superconductivity in excess of liquid nitrogen temperatures.
Isabel retweeted
From Scotch-tape exfoliation to wafer-scale electronics.
Our improved roll-to-roll mechanical exfoliation method yields homogeneous large-area films of 2D materials like MoS2, WSe2, hBN and graphite. Fully dry and scalable!
Link in the 1st comment 👇
#2Dmaterials #nanotech
Isabel retweeted
Autonomous atomic-scale fabrication: ML-guided electron beams sculpting 2D materials
Functional properties of 2D materials like MoS2 are exquisitely sensitive to atomic structure. A metallic Mo6S6 nanowire along a hole edge in semiconducting MoS2 yields a 1D-2D heterostructure with electronic behavior nothing like the pristine monolayer. Building this atom by atom means ejecting sulfur atoms at chosen sites, and earlier automated workflows leaned on lucky statistics, with success rates near 64%.
Zijie Wu and coauthors build a fully autonomous platform that closes the loop between imaging, ML, and beam control in a STEM. After each HAADF acquisition, three lightweight models decode the image: a U-net ensemble (ELIT) locates atoms, a second U-net segments collapsed MoS-nanowire regions, and a random forest classifies each site as Mo, S, or sulfur-vacancy-line. The classifier learns from 55 hand-labeled neighborhoods, each encoded by distances and intensity to the 5 nearest atoms. No huge simulated datasets, no GAN style transfer, no foundation model.
Selected sulfur sites go to an FPGA-controlled scan routine that places the beam in Archimedean spirals with radially symmetric dose. Three strategies are shown: targeted nanowire growth along a pore edge, freestanding MoS-NWs between two seed pores, and directional growth using DBSCAN and line fitting on the largest cluster. The directional case reaches 65% success over 42 runs, with decoding in about 1 second on a GPU.
What stands out from an ML-for-science angle is the deliberate small-data philosophy. For atomic labeling, big vision models trained on simulated images often transfer poorly to experiments. Pairing simple models with intuition-based feature engineering is faster to deploy and easier to tune.
For applied R&D in semiconductors, quantum devices, and 2D electronics, this reframes what atomic-scale design means. Teams can specify a target defect topology and let the platform iterate toward it without massive labeled datasets, and the framework should transfer to other TMDs, making it relevant for catalysis, energy materials, and quantum sensors.
Paper: Wu et al., npj Computational Materials (2026) — CC BY 4.0 | doi.org/10.1038/s41524-026-0…
