Bayesian AI from ORNL Accelerates Autonomous Scanning Tunneling Microscopy to Map Atomic Structure-Property Links

Bayesian Deep Learning Accelerates Atomic-Scale Discovery at ORNL

A team at Oak Ridge National Laboratory has built a faster way to map how atomic structures drive electronic behavior. By combining Bayesian deep learning with scanning tunneling microscopy (STM), the system searches large sample areas, flags regions of interest, and runs spectroscopy with minimal human input.

The model's uncertainty estimates guide where the microscope probes next. Result: broader coverage, higher precision on hotspots, and far less time than conventional raster-and-scan routines.

"This method makes it possible to study a material's properties with much greater efficiency," said Ganesh Narasimha from ORNL. "Usually, we would need to scan a large region, and then several small regions, and perform spectroscopy, which is very time-consuming. Here, the AI algorithm takes control and does this process automatically and intelligently."

Why this matters for scientists

• Active learning loops cut idle acquisition and focus effort on informative sites.
• Bayesian models provide uncertainty-aware decisions, reducing redundant measurements.
• Structure-property links surface faster, improving hypotheses and follow-up experiments.
• Approach is material-agnostic and portable across STM, related SPM modes, and other probes.

How it works

• Bayesian deep learning: Neural nets produce both predictions and uncertainty, enabling exploration-exploitation decisions.
• Autonomous STM control: The algorithm selects new scan points and spectroscopy targets, updating its model as data arrive.
• Active learning: Each step prioritizes the next most informative measurement rather than uniform grids.
• Multiscale correlation: Atomic features (e.g., defects) are linked to electronic responses, revealing which local structures drive measured properties.

Case study: Europium zinc arsenide (EuZn2As2)

The team validated the system on europium zinc arsenide, a magnetic semimetal with distinctive electronic behavior. Using STM, the model identified atomic-scale features, including a vacancy defect, and mapped their electronic signatures.

The workflow connected local structure to response while minimizing scan time. Although demonstrated on EuZn2As2, the method extends to other materials where defects, interfaces, and heterogeneity matter.

What you can apply in your lab

• Instrument loop: Expose STM control APIs to a Python agent that proposes next-point measurements based on uncertainty.
• Modeling stack: Start with a Bayesian CNN or ensemble for image patches; add a lightweight Gaussian process for quick uncertainty calibration.
• Acquisition policy: Use uncertainty sampling or expected information gain for point selection; throttle with safety checks for tip integrity and drift.
• Spectroscopy triage: Trigger I-V or dI/dV only on high-value coordinates; cache raw traces with metadata for later refinement.
• Data plumbing: Maintain a streaming store (HDF5/Zarr) with synchronized position, bias, current, temperature, and timestamp.
• Compute: Run inference on a local GPU; pin critical routines to real-time threads to keep microscope latency low.
• Validation: Periodically compare AI-selected spectra against a small uniform baseline to monitor bias and drift.

Metrics to track

• Informative measurements per hour vs. uniform scanning.
• Reduction in total spectroscopy events at fixed confidence.
• Agreement between AI-predicted and measured electronic features (e.g., dI/dV peaks).
• Stability metrics: tip lifetime, drift rate, and re-approach counts.

Where to read more

Paper: Uncovering multiscale structure-property correlations via active learning in scanning tunneling microscopy (npj Computational Materials, 2025). DOI: 10.1038/s41524-025-01642-1

Lab: Oak Ridge National Laboratory. https://www.ornl.gov/

Skills and training

If you're building an active-learning pipeline for microscopy or spectroscopy, a focused upskilling track helps. See curated AI course paths by role: Complete AI Training - Courses by Job