Open-source framework improves autonomous experiments

The invention of new materials, essential in the advancement of fields like battery technology and drug formulation design, has been catapulted by machine learning and artificial intelligence. But the autonomous experiments that can discover new materials need physics-based modeling and decision making processes.

Sutherland et al. developed AutoSAS, an open-source framework for automating data classification designed to aid in autonomous small angle scattering experiments. The program creates classification results and quantitative structural descriptors with human-defined candidate models, high-throughput combinatorial fitting, and information-theoretic model selection.

“We set out to develop a more robust, broadly applicable framework rooted in established scattering theory, one that could provide meaningful parameters while operating autonomously,” said author Peter Beaucage. “This advance was critical to enable true human-machine collaboration, in which an experimental platform not only collects data but also reasons about it in real time.”

The software tool was tested on an experiment for X-ray and neutron scattering-based optimization n of multicomponent liquid formulations. The program worked well and even discovered a structural transition boundary that hadn’t been previously identified.

“AutoSAS revealed phase boundaries and structural transitions in real time — phenomena that typically require hours of offline analysis,” Beaucage said. “By using information-theoretic model selection, it not only fits data reliably but also identifies which parameters carry genuine scientific meaning.”

AutoSAS is now being used by the researchers in an experiment on polymer self-assembly and inorganic nanostructure synthesis. They plan continue improving the program by integrating multimodal measurements and scaling up for higher throughput workflows.

Source: “AutoSAS: A new human-aside-the-loop paradigm for automated SAS fitting for high throughput and autonomous experimentation,” by Duncan R. Sutherland, Rachel Ford, Yun Liu, Tyler B. Martin, and Peter A. Beaucage, APL Machine Learning (2025). The article can be accessed at https://doi.org/10.1063/5.0271073 .

This paper is part of the Machine Learning for Self-Driving Laboratories Collection, learn more here .