Active Learning for Data-Efficient Process Mapping and Material Design

Abstract

The mapping and optimization of advanced materials and manufacturing processes is frequently challenged and constrained by the cost of acquiring data. Whether a label is obtained through building and characterizing additively manufactured parts, or from a complex computational chemistry calculation, each observation in high-dimensional or complex design/process spaces often carries a significant experimental or computational cost, making exhaustive exploration infeasible. This dissertation investigates adaptive sampling strategies for mapping such design and process spaces efficiently. We focus particularly on active learning, in which a surrogate machine learning model is used to iteratively select the most information-dense samples in the given system. These active learning methods balance predictive uncertainty against broad coverage of the parameter space, a balance that is known as the “exploration-exploitation trade-off.” In this work, we explore this trade-off, leveraging hybrid active learning strategies that allow fine-tuning of this trade-off.

In the second chapter, we investigate a pulsed-wave (PW) laser powder bed fusion process (L-PBF), with the goal of building high-strength parts with oxide-dispersion strengthened (ODS) steels. We use feature engineering methods and various machine learning techniques to ultimately construct a process map of the high-strength region of the parameter space with physically-interpretable equations. We then use the same identified features to develop simple models for both the part density and nano-hardness.

In the third chapter, we explore various active learning techniques utilizing engineered datasets and simulated phase diagrams. Specifically, we investigate various hybrid active learning methods to efficiently build process maps and phase diagrams. We determine optimal uncertainty sampling methods and hybrid method trade-offs for various design problems. Finally, we propose a “switching” acquisition method that incorporates the strengths of two uncertainty sampling methods – margin sampling and least confidence – and performs better than either method could separately.

In Chapter 4, we created a high-entropy alloy (HEA) dataset centered around H-adsorption in the hydrogen evolution reaction (HER). We then use various regression, classification, and level-set active learning methods to map the region corresponding with high HER activity. We find that a space-filling exploration performs better than other tested learning methods, and we propose future paths for further investigation of this dataset.