QUBO-Compatible Active Learning for Inverse Design of High-Entropy Alloys
Machine-learned forward models can rapidly predict alloy properties, but their use for inverse design remains challenging when the search should also retain compatibility with quadratic unconstrained binary optimization (QUBO). Here, we develop a QUBO-compatible active-learning framework for inverse design of high-entropy alloys using a pretrained graph-neural-network predictor as a fixed property oracle. A property-guided binary variational autoencoder provides a binary latent representation, while an ensemble of quadratic factorization machines guides candidate selection. We systematically benchmark the framework through controlled latent-space ablations and comparison with direct composition-space optimization. The results show that candidate generation is a major determinant of search performance: local perturbations around previously high-performing latent codes provide the largest workflow-specific improvement, while surrogate-based selection further prioritizes candidates within the enriched search pool. The resulting QUBO-compatible workflow remains competitive with strong classical optimization strategies, although a composition-space genetic algorithm achieves the highest mean score. Finally, the learned quadratic surrogate can be exported directly as a QUBO. These results show that effective data acquisition can be separated from the final QUBO optimization endpoint, providing a benchmarked route for QUBO-compatible data-driven materials inverse design.