A team of ORNL researchers has achieved a critical enhanced understanding of the processes used by neural networks in selecting certain alloys for use in particular applications. This discovery has enabled a novel “learn from AI” paradigm for finding new investigational paths for better alloy design.
A cross-cutting team of ORNL researchers has achieved an enhanced understanding of why AI networks select certain alloy candidates for particular applications. Specifically, by observing and comparing the activated paths of thousands of neural networks trained in-house, the researchers were able to determine what properties the networks sought out when selecting an alloy candidate. The team captured this knowledge in a software package that covers the entire process of explainability, allowing researchers to more confidently and efficiently harness the power of AI in the creation of novel alloys to advance applications in transportation and power production.
Because the alloy development data sets generated by decades of experimentation and simulation are simply too large for human analysis, and alloy testing is inherently expensive, researchers have turned to artificial intelligence. While AI is indeed good at predicting alloy properties, the reasons why AI networks select certain alloys as promising candidates for particular applications has plagued researchers. The team’s approach to understanding AI explainability is applicable to a wide range of AI-focused materials challenges and the members plan to open source the software soon.
Research team: Matt Lee, Dongwon Shin, J. Travis Johnston, Junghoon Chae, Yukinori Yamamoto, and Mike P. Brady
Funding: This research was supported by the Laboratory Directed Research and Development Program of ORNL, managed by UT-Battelle, LLC, for the U.S. DOE.
Contact: Matt Lee, lees4@ornl.gov
