(435a) Successful "Closed Loop" Materials Discovery: the Power and Challenges of Machine Learning
AIChE Annual Meeting
2024
2024 AIChE Annual Meeting
Topical Conference: Applications of Data Science to Molecules and Materials
Applications of Data Science to Molecules and Materials
Tuesday, October 29, 2024 - 3:30pm to 4:00pm
The Materials Genome Initiative was started in 2011 with the prospect of using machine learning to make materials “faster, cheaper, better.†A decade or so later, we are finally starting to have the tools to exploit that prospect more effectively. A major factor contributing to the inefficiency of material discovery is the large combinatorial space of materials candidates as well as processing conditions, which are often sparsely observed for a given application. Searches of this space are often formed by expert knowledge and clustered
close to known materials. Exhaustive experimental characterization or first principles calculations are also expensive, invariably leading to small available data sets. As a result, there is a need to develop algorithms that can efficiently search this large parameter space capable of dealing with “tiny†data sets. In this talk, we will introduce our approach to mitigate some of these issues using our in-house PAL 2.0, Bayesian optimization code base, that features a chemistry-informed belief model. A key characteristic of PAL 2.0 is the creation of a physics-based hypothesis using XGBoost and Neural Networks for “feature engineering,†which provides a physics-based prior to the Gaussian process model used in Bayesian optimization. Our method picks out the physical descriptors that are most representative of the material domain, making the search independent from expert knowledge. Here, we demonstrate PAL 2.0's use in a “closed-loop†set up with our collaborators at Hopkins Advanced Physics Laboratory to discover high temperature shape memory alloys for space actuation applications. We discuss the use of constraints to help “zero in†to manufacturable solutions in the fewest number of optimization iterations. In the process, we discover new multi-principal element alloys (MPEAs) with dramatically improved targeted properties, in this case alloy toughness.
close to known materials. Exhaustive experimental characterization or first principles calculations are also expensive, invariably leading to small available data sets. As a result, there is a need to develop algorithms that can efficiently search this large parameter space capable of dealing with “tiny†data sets. In this talk, we will introduce our approach to mitigate some of these issues using our in-house PAL 2.0, Bayesian optimization code base, that features a chemistry-informed belief model. A key characteristic of PAL 2.0 is the creation of a physics-based hypothesis using XGBoost and Neural Networks for “feature engineering,†which provides a physics-based prior to the Gaussian process model used in Bayesian optimization. Our method picks out the physical descriptors that are most representative of the material domain, making the search independent from expert knowledge. Here, we demonstrate PAL 2.0's use in a “closed-loop†set up with our collaborators at Hopkins Advanced Physics Laboratory to discover high temperature shape memory alloys for space actuation applications. We discuss the use of constraints to help “zero in†to manufacturable solutions in the fewest number of optimization iterations. In the process, we discover new multi-principal element alloys (MPEAs) with dramatically improved targeted properties, in this case alloy toughness.