Abstract
This dissertation develops and deploys a quantum-assisted machine learning (QaML) framework for the design and discovery of high-entropy alloys (HEAs) with targeted mechanical properties. The work progresses through four interconnected stages: classical machine learning model development, physics-guided alloy design, quantum-annealing-based optimization, and end-to-end experimental alloy discovery.
A physics-informed machine learning framework is first constructed for predicting yield strength and fracture strain in BCC and BCC+B2 HEAs. A curated experimental dataset is paired with 28 physically motivated descriptors, and Genetic Algorithm–based feature selection identifies Random Forest Regression as the best-performing model, achieving an RMSE of 182.28 MPa (R² = 0.86) for yield strength and 9.19% (R² = 0.66) for fracture strain. SHAP analysis reveals the dominant physical mechanisms governing each property.
This classical framework is then deployed within an Integrated Computational Materials Engineering (ICME) pipeline to design aluminum-enriched refractory HEAs with controlled B2 ordering. Machine learning screening narrows 84 candidate quaternary systems to the AlHfNbTi platform, and Monte Carlo simulations guide compositional tuning from a strongly ordered, brittle alloy (Al₂₅Hf₂₅Nb₂₅Ti₂₅, 1.7 GPa, 2% strain) to a weakly ordered alloy with exceptional strength–ductility synergy (Al₁₀Hf₂₀Nb₂₂Ti₃₃V₁₅, 1.0 GPa, 9% ductility).
Feature selection is then recast as a discrete optimization problem and solved using quantum annealing on D-Wave hardware. QBoost and Quantum Mutual Information are developed and paired with a recursive batching strategy to map a 255-feature descriptor library onto current quantum processors. Quantum annealing is further integrated into model training through a quantum-assisted support vector machine and a novel quantum pruning method for neural network compression. For fracture strain classification, quantum-annealing-based pruning raises test accuracy from 0.872 to 0.923 and F1 score from 0.815 to 0.889, an outcome consistent with quantum annealing preferentially sampling broad basins that favor generalization.
Finally, the complete QaML pipeline is deployed for alloy discovery in the Al–Cr–Fe–Mn–Ti system, identifying Al₈Cr₃₈Fe₅₀Mn₂Ti₂ as the optimal composition. Experimental synthesis confirms a single-phase BCC solid solution with a yield strength of 568 MPa, compressive strain exceeding 40%, and corrosion resistance superior to 304 stainless steel. Critically, the baseline unpruned neural network classified this composition as brittle; it was the quantum-annealed pruning that enabled its discovery.
The dissertation proceeds from foundational methods (Chapters 2–3: supervised learning and quantum annealing) through classical model development and alloy design (Chapters 4–5) to quantum-enhanced optimization and model training (Chapters 6–7), culminating in end-to-end experimental alloy discovery (Chapter 8). Chapter 9 summarizes the principal contributions and outlines future directions.
