"Assessing the reliability of machine learning models in predicting properties of solid-state materials "

January 27, -
Speaker(s): Christopher A. Sutton, PhD
Advances in machine learning (ML) are making a large impact in many disciplines, including materials and computational chemistry. A particularly exciting application of ML is the prediction of quantum mechanical (QM) properties (e.g., formation energy, bandgap, etc.) using only the composition or structure as input. Assuming sufficient accuracies in the ML models, these methods enable screening of a considerably large chemical space at orders of magnitude lower computational cost than QM methods. Despite the promise of ML in material screening, several key challenges remain in both applying and interpreting the results of ML algorithms. Here, we will discuss our efforts in addressing these issues, including our recent work on predictive screening of new solid-state materials and using interpretable descriptors and our work opening the black box of ML methods by identifying the domain of applicability, i.e., where a given model is reliable.
Sponsor

Pratt School of Engineering

Co-Sponsor(s)

Chemistry; Duke Materials Initiative; Energy Initiative; Mechanical Engineering and Materials Science (MEMS); Physics; University Program in Materials Science and Engineering (MatSci)

Contact

Tyson, Quiana
660-5263