Efficient Materials Informatics between Rockets and Electrons

TL;DR

This dissertation discusses the design of multi-alloy functionally graded materials (FGMs) incorporating ultra-high temperature refractory high entropy alloys (RHEAs) towards gas turbine and jet engine efficiency increase reducing CO2 emissions, as well as hypersonic vehicles.

Abstract

The true power of computational research typically can lay in either what it accomplishes or what it enables others to accomplish. In this work, both avenues are simultaneously embraced across several distinct efforts existing at three general scales of abstractions of what a material is - atomistic, physical, and design. At each, an efficient materials informatics infrastructure is being built from the ground up based on (1) the fundamental understanding of the underlying prior knowledge, including the data, (2) deployment routes that take advantage of it, and (3) pathways to extend it in an autonomous or semi-autonomous fashion, while heavily relying on artificial intelligence (AI) to guide well-established DFT-based ab initio and CALPHAD-based thermodynamic methods. The resulting multi-level discovery infrastructure is highly generalizable as it focuses on encoding problems to solve them easily rather than looking for an existing solution. To showcase it, this dissertation discusses the design of multi-alloy functionally graded materials (FGMs) incorporating ultra-high temperature refractory high entropy alloys (RHEAs) towards gas turbine and jet engine efficiency increase reducing CO2 emissions, as well as hypersonic vehicles. It leverages a new graph representation of underlying mathematical space using a newly developed algorithm based on combinatorics, not subject to many problems troubling the community. Underneath, property models and phase relations are learned from optimized samplings of the largest and highest quality dataset of HEA in the world, called ULTERA. At the atomistic level, a data ecosystem optimized for machine learning (ML) from over 4.5 million relaxed structures, called MPDD, is used to inform experimental observations and improve thermodynamic models by providing stability data enabled by a new efficient featurization framework.

Figures (117)

• Figure 1: Intermediate material modeling scales bridging together quantum physics and aerospace engineering to enable high-technology solutions through excellence of underlying ensembles of materials. In this work, all of the scales are brought together to take advantage of data and knowledge from all relevant sources. Top render of hypersonic vehicle reproduced from DARPA under public domain and gray nozzle renders from Hofmann2014 under CC BY-NC-ND 4.0 License. Several images occur later in the manuscript in Figures \ref{['pathplan:fig:lowgradientsquared']}, \ref{['ultera:fig:dashboard']}, \ref{['ultera:fig:dataloops']}, \ref{['inverse:fig:cgandemo']}, \ref{['infeasibilitygliding:fig:glide']}, \ref{['crystall:fig:ndbi2clusters']}, \ref{['pysipfenn:fig:ks2022']}, \ref{['sipfenn:fig:oqmdperformance']}, and \ref{['mpdd:fig:dataset']}.
• Figure 2: Schematic outline of this dissertation flowing through 3 overarching types of materials science research. It starts from atomistic treatment (blue) allowing modeling of physical materials (blue) and leading to design (green). For each category, three most significant advancements done in this work have been selected to showcase computational infrastructures and methods to extend our understanding or capabilities.
• Figure 3: The model design process schematic.
• Figure 4: Three selected architectures designed within the present work. Optimized for: (Left) OQMD performance, (Middle) predicting new materials, (Right) small size at good performance. Internally in the code, they are designated as NN9, NN20, and NN24.
• Figure 5: Training Loss to Validation Loss in a model that does without (NN9) and with overfitting mitigation (NN20), plotted versus training progress.