Science-Driven Automated Experiments

This invention presents technologies for characterizing physical properties of a sample's surface by combining image processing with machine learning techniques. It enables the automation of experiments by scanning the sample's surface, dividing the image into patches, and associating these patches with physical measurements. Through training a model to understand the relationship between the surface's local features and the measurements, the method predicts outcomes for unmeasured areas, selects the most informative next measurement based on these predictions, and iteratively refines its understanding. This process continues until specific criteria are met, significantly optimizing experimental efficiency, and reducing sample exposure to potentially damaging conditions.

Description

The disclosed technology involves scanning a sample surface, creating image patches, and linking these to physical measurements at certain locations. A machine learning model is trained to correlate the image patches' features with the physical measurements. This model then predicts the physical characteristics for unmeasured patches, identifies which patches should be measured next, and instructs an instrument to perform these measurements. This iterative process, which includes retraining the model with new data, continues until a predetermined threshold is reached. The approach allows for joint determination of model parameters, prediction of representations for unmeasured patches, and reduction of prediction uncertainties. This method is adaptable to various instruments and measurement types, making it broadly applicable across scientific and engineering disciplines.

Benefits

• Efficient exploration of sample properties with minimal data collection.
• Reduced risk of sample damage from excessive measurements.
• Ability to discover complex relationships between physical properties and structural features.

Applications and Industries

• Material science, for exploring properties of new materials.
• Pharmaceutical and biomedical research, for studying biological samples.
• Semiconductor industry, for surface analysis and quality control.

Related ORNL Technologies

• Correlative Multimodal Chemical Imaging via Machine Learning
• Machine Learning on Multimodal Chemical and Whole Slide Imaging Data for Predicting Cancer Labels Directly from Tissue Images
• Pan-Sharpening for Microscopy
• Hypothesis Learning in Automated Experiments
• Human-in-the-Loop Machine Learning for Automated Experiments

Contact

To learn more about this technology, email partnerships@ornl.gov or call 865-574-1051.