Active learning is increasingly adopted in AI-driven materials design to accelerate the discovery of high-performance materials under limited experimental budgets. Gaussian Process Regression (GPR) is particularly attractive for composition–process ...
Active learning is increasingly adopted in AI-driven materials design to accelerate the discovery of high-performance materials under limited experimental budgets. Gaussian Process Regression (GPR) is particularly attractive for composition–process exploration because it provides both predictions and uncertainty estimates. However, as the dimensionality of design variables (e.g., composition, heat-treatment temperature, and time) increases, data sparsity often leads to unreliable uncertainty estimates, which can degrade exploration efficiency.
To address this challenge, we propose an active learning framework based on Manifold-Constrained Semi-Supervised Representation Learning (MC-SSAE) to enable sample-efficient exploration in high-dimensional spaces. The framework learns a low-dimensional latent representation via an autoencoder while preserving local geometry through manifold regularization. In addition, a teacher–student scheme with confidence-aware pseudo-labeling stabilizes representation learning even with scarce labeled data.
When coupled with GPR for uncertainty-driven acquisition, the learned latent space resulted in improved predictive performance and faster convergence compared to conventional unsupervised autoencoders and baseline semi-supervised autoencoders. These results indicate that MC-SSAE serves as an effective preprocessing strategy for active learning-based experimental design in high-dimensional materials problems, including alloy design and process optimization.