Session: ASME Undergraduate Student Design Expo

Paper Number: 172601

Efficient Exploration of Shape Memory Alloy Process-Composition Space Using Multi-Objective Bayesian Optimization 

Designing Nickel-Titanium-Hafnium (NiTiHf) Shape Memory Alloys (SMAs) optimized for autonomous aerospace actuation requires precisely tuned thermal properties, minimizing both thermal hysteresis (ΔT) and mean transformation temperatures (T). Traditionally, material design in this field relies on costly, labor-intensive trial-and-error experiments. Efficiently maximizing insights from limited experimental data through computational methods is thus critical to accelerate discovery, reduce costs, and enable rapid iteration of SMA materials beyond the existing Pareto front.

NiTiHf alloys offer significant potential for morphing aerospace structures due to thermally reversible phase transformations near targeted temperatures at takeoff and cruising altitudes. NASA has characterized SMA compositions and thermal properties across small regions of the vast parameter space. Previous machine learning efforts established multi-objective Bayesian optimization (MOBO) with Gaussian Process Regression (GPR) surrogates as effective for exploring such multi-dimensional design spaces. GPRs model complex, nonlinear mappings from composition and processing to properties while providing uncertainty estimates to guide optimization—without costly physics-based simulations. However, standard MOBO approaches often struggle to scale due to the exceedingly large number of process parameters that can be considered and the computational cost of evaluating sizable discretized candidate sets. These challenges motivate more scalable strategies to explore the multi-objective Pareto front.

This research has two goals: (1) Develop computational methods to efficiently identify promising NiTiHf alloy candidates by minimizing both ΔT and T beyond the current Pareto front, and (2) experimentally validate these predictions to confirm the accuracy and reliability of optimized alloy designs. We developed a refined MOBO framework using the BoTorch Python library, integrating several features to enable efficient search over a vast, noisy design space. Our strategy incorporates noisy expected hypervolume improvement (logNEHVI) for robust acquisition under uncertainty, dynamic reference point adjustment to reflect evolving predictions, and UCB (upper confidence bound) filtering to eliminate an adjustable portion of more than seven hundred thousand candidates prior to optimization. Grid-based candidate sets are generated using experimentally informed bounds and step sizes, ensuring laboratory feasibility. To promote diversity, a larger pool of high-performing candidates is first selected, from which six distinct alloy-processing combinations will be down-selected using K-Means clustering for experimental validation. The optimization models are trained on a filtered dataset of NiTiHf alloys compiled by NASA and tailored to consist of the thermal properties considered here. The performance of the proposed MOBO framework is then quantified by R² values for ΔT and T as has been done in previous studies.

Novel candidate alloys predicted to push the pareto front will be fabricated in batches to demonstrate sequential learning. The samples will be created by arc melting pure elements and pre-alloyed stocks in a controlled argon environment using a GTAW-capable Centorr arc melting furnace. Each composition will undergo precise wire EDM cutting, surface preparation, and repeated melting cycles, followed by heat treatment in a calibrated box furnace according to specified thermal schedules. Differential Scanning Calorimetry samples will be prepared and tested to assess thermal hysteresis and transformation temperatures for Pareto front comparison.

Our preliminary findings suggest that meaningful property optimization can be achieved using a small number of strategically selected experiments. The ability to expand the known Pareto front with limited data would highlight the effectiveness of our MOBO framework in navigating complex, high-dimensional design spaces. We hope to enable data-efficient discovery of novel alloy-processing combinations through this methodology, offering a scalable path toward the design of next-generation SMAs for aerospace and beyond.

Presenting Author: Dylan Winer Georgia Institute of Technology

Presenting Author Biography: Dylan Winer is a third-year Mechanical Engineering student at the George W. Woodruff School of Mechanical Engineering at the Georgia Institute of Technology. He conducts research with the MINED Group under Professor Surya Kalidindi and the Stebner Lab under Professor Aaron Stebner, where his work focuses on data-driven design and optimization of shape memory alloys.

Authors:

Dylan Winer Georgia Institute of Technology
Jihye Hur Georgia Institute of Technology
Michael Buzzy Georgia Institute of Technology
Surya Kalidindi Georgia Institute of Technology
Aaron Stebner Georgia Institute of Technology