Cornell University researchers have built an AI system called EMSeek that can analyze an electron microscopy image and deliver a full materials-science report in 2 to 5 minutes, a task that traditionally takes human experts roughly 50 times longer. Published on April 1, 2026, in Science Advances, the work addresses one of the most persistent bottlenecks in materials research: the slow, manual process of turning microscope images into usable scientific insight.
Five Agents, One Continuous Pipeline
EMSeek is not a single algorithm. It is a modular, provenance-tracked multiagent platform that chains together five distinct units, each responsible for a different stage of analysis. The system begins with feature identification, where it segments and labels structures visible in an electron microscopy image. It then moves to crystal-structure determination, matching those features to known atomic arrangements. Property prediction follows, estimating physical characteristics such as formation energy. A literature comparison step cross-references results against published findings. Finally, a report generation unit compiles the entire chain of evidence into a structured document.
What separates this from a simple automation script is the provenance tracking built into every handoff between units. According to the Science Advances article, the system enforces unit consistency at each transition, while an internal module called Scribe compiles image masks, crystallographic information files, and property tables into a single auditable package. That design choice means a researcher can trace any output back to the raw image and verify each intermediate step, a feature that matters when results feed into decisions about battery chemistries or semiconductor designs.
Because each agent passes along both its results and a record of how it obtained them, EMSeek effectively builds a machine-readable lab notebook for every sample. If the crystal-structure unit proposes an unexpected phase, for example, a human can inspect the segmentation maps and property predictions that led to that conclusion. This structured transparency is intended to make the system suitable not just for exploratory analysis but also for work that must withstand peer review.
Speed That Changes the Research Cycle
The headline number, 2 to 5 minutes per analysis, deserves context. Conventional expert workflows for the same task involve loading images into separate software tools, manually identifying features, consulting crystallographic databases, running property calculations, and writing up results. According to Cornell’s news office, that process typically takes about 50 times longer than EMSeek’s automated pipeline.
A 50-fold speedup does not just save time on individual samples. It changes the economics of exploration. When analysis is slow and expensive, researchers tend to study only the most promising candidates. When it takes minutes, they can afford to screen dozens of samples in a single session, catching unexpected structures that might otherwise go unexamined. For fields like clean energy materials, where the search space for viable compounds is enormous, that shift from selective to systematic screening could compress discovery timelines by months or years.
Faster turnaround also feeds back into experimental design. Instead of waiting days for interpretation, teams can adjust synthesis parameters or imaging conditions in near real time, guided by EMSeek’s reports. That kind of tight loop between experiment and analysis has been a long-standing goal in automated laboratories and could make high-end microscopes more productive across an institution.
Who Built It and Why It Matters
The project was led by Fengqi You, a faculty member at Cornell’s Duffield College of Engineering and co-director of the Cornell AI for Science Institute. You’s lab has focused on applying optimization and machine learning to complex engineering problems, and EMSeek represents a direct application of that expertise to a domain where human analysts have long been the rate-limiting step.
The broader Cornell research ecosystem has been investing in AI-driven scientific tools for several years, and EMSeek fits into that institutional push. The main university hub at cornell.edu highlights cross-disciplinary initiatives that pair physical sciences with computation, while the technology-focused campus maintains an archive of AI-related work that underscores how quickly these methods are moving from theory into laboratory practice. Within that context, an end-to-end microscopy assistant is less a standalone curiosity than a natural extension of a campus-wide strategy.
More broadly, EMSeek reflects a growing recognition across materials science that the bottleneck is no longer data collection. Modern electron microscopes generate images faster than humans can interpret them. The real constraint is analysis, and that is precisely where agentic AI systems like EMSeek apply pressure. By formalizing what experts do into a series of coordinated agents, the Cornell team is testing whether that expertise can be scaled across many more samples and instruments.
How EMSeek Compares to Parallel Efforts
Cornell’s platform is not the only project trying to automate the path from microscopy to materials insight. A related effort called AutoMat, described in a preprint on arXiv, tackles a similar problem with a different architecture. AutoMat is an agentic pipeline that converts scanning transmission electron microscopy (STEM) images into atomic crystal structures and predicts physical properties from those structures. The team behind it introduced a benchmark called STEM2Mat-Bench and reports quantitative metrics including lattice RMSD, formation-energy MAE, and structure-matching success rate.
The key difference is scope. AutoMat focuses on crystal structure reconstruction and property prediction from STEM images, while EMSeek wraps those capabilities into a broader workflow that includes literature comparison and automated report generation. EMSeek also emphasizes provenance tracking across every step, a design priority that AutoMat’s published description does not highlight to the same degree. For a working scientist, that distinction matters: a tool that produces a number is useful, but a tool that produces a number with a full audit trail is publishable.
There is also a difference in how the two projects position their benchmarks. AutoMat leans on standardized metrics to compare against alternative reconstruction pipelines, whereas EMSeek is framed as a practical assistant embedded in existing laboratory routines. In practice, researchers may end up combining ideas from both systems, using benchmarked reconstruction modules inside a provenance-aware reporting shell.
The Broader AI-in-Microscopy Movement
Both EMSeek and AutoMat belong to a wider wave of machine learning integration in microscopy. Researchers at MIT, for example, have explored what they call “smart microscopy,” where AI systems optimize beam usage by directing attention to regions the algorithm deems most scientifically interesting. That approach operates at the image acquisition stage, deciding what to look at, while EMSeek operates at the interpretation stage, deciding what the image means.
Together, these efforts suggest a future where the entire microscopy workflow, from sample loading to published result, involves AI at every step. An experiment might begin with an acquisition system that dynamically steers the beam, proceed through an analysis pipeline like EMSeek that reconstructs structure and properties, and end with automated drafting of figures and text for a manuscript. Each component is still in development, but the direction of travel is clear.
The risk, of course, is that automation introduces systematic blind spots. If the training data behind EMSeek’s feature identification module underrepresents certain crystal families, the system could consistently miss or misclassify those structures. The provenance tracking built into EMSeek offers some protection here, since a human reviewer can inspect each intermediate output, but it does not eliminate the problem. No amount of audit trail helps if the reviewer trusts the machine and skips the check.
What Remains Unproven
For all its promise, EMSeek’s public record still has gaps. The Science Advances paper reports case studies and performance comparisons, but widespread adoption will hinge on how the system behaves on messy, real-world data outside curated test sets. Electron microscopy images vary widely by instrument, operator, and sample preparation technique, and it remains to be seen how robust the five-agent pipeline is across that diversity.
Another open question is integration. Many laboratories already maintain custom workflows built from commercial and open-source tools. Plugging EMSeek into those environments will require interfaces that respect existing data formats, security policies, and collaboration patterns. The Cornell team’s emphasis on provenance is a step in that direction, but turning a research prototype into a dependable workhorse is a separate engineering challenge.
There is also the matter of trust. Automated literature comparison and report generation blur the line between analysis and authorship, raising questions about how to credit AI systems in scientific publications and how to guard against subtle errors propagating into the record. Institutions such as Cornell University are already debating guidelines for responsible AI use in research, and tools like EMSeek will likely accelerate those conversations.
Still, the trajectory is hard to ignore. From multiagent pipelines that reconstruct atomic lattices to acquisition systems that steer microscope beams, AI is moving steadily closer to the core of experimental science. If EMSeek and its peers can deliver on their promise of speed, transparency, and robustness, the slowest part of many microscopy experiments may soon be preparing the sample, not interpreting the image.
More from Morning Overview
*This article was researched with the help of AI, with human editors creating the final content.
