Machine Learning for Complex Materials Science
Date of Award
Spring 2024
Document Type
Dissertation
Degree Name
Doctor of Philosophy (PhD)
Department
Mechanical Engineering & Materials Science (ENAS)
First Advisor
Schroers, Jan
Abstract
Complex materials science problems, such as metallic glass formation, originate from the collective behavior of many atoms and are by many orders of magnitude too large to be solved by first-principles calculations. The successful application of machine learning (ML) in various fields outside of materials science suggests its potential in addressing such complex problems. To test its efficacy, we attempt to predict bulk metallic glass formation using ML. Surprisingly, we find that a general-material ML model with 201 alloy features, constructed through simple statistical functions from 31 elemental features, is indistinguishable from models that are unphysical or do not consider any features when the prediction accuracy is tested in an interpolation manner. Only when significant separation of training and testing data is carried out does the general-material model perform better in this extrapolation mode than the unphysical or composition-only models, yet significantly worse than a human learning based 3-(alloy)feature model. We attribute the limited performance of the general-material model to its general inability to accurately represent alloy features through elemental features. As the potential data space is too large to determine a representative fraction that would allow ML models to work well even with poorly representative features, complex material science problems like metallic glass formation require physical insights about mixing behavior to develop effective ML models.Another challenge of applying ML to complex materials science problems is that the underlying mechanisms may vary within the considered problem space. To quantify this, we divide alloy data into subgroups and construct ML models to predict metallic glass formation. We discover that subgrouping guided by physical insights into the problem leads to significantly higher prediction accuracy. Specifically, when applying Inoue’s subgrouping which is based on the constituent elements of alloys, models specific to subgroups outperform those trained on the entire dataset. Furthermore, our results highlight the potential of ML strategies to quantitatively assess guiding principles utilized in materials science, allowing for a rigorous evaluation of empirical rules like Inoue's. Additionally, our analysis uncovers distinct mechanisms governing the glass-forming ability (GFA) within different subgroups, shedding light on the diverse nature of this phenomenon. Statistical methods for subgrouping prove less effective compared to physics-informed subgrouping. Lastly, recognizing the limitations such as data representation bias and reliance on binary labeling, we curate a dataset of binary alloy libraries fabricated using combinatorial co-sputtering and characterized through high-throughput strategies to achieve a better representation of the GFA in binary alloys. By employing a continuous labeling of critical cooling rates and physics-informed alloy features, the model displays robust extrapolation capabilities, successfully predicting the GFA of new alloy systems unseen by the model. This thesis demonstrates the significance of physics-informed insight into the problem to be addressed to build effective ML models. Such an integrated approach has the potential to unlock new insights and accelerate materials discovery in a wide range of applications beyond metallic glass formation.
Recommended Citation
Liu, Guannan, "Machine Learning for Complex Materials Science" (2024). Yale Graduate School of Arts and Sciences Dissertations. 1439.
https://elischolar.library.yale.edu/gsas_dissertations/1439
