This article explores the effectiveness of uncertainty-based active learning (AL) in approximating black-box functions (BBFs) relevant to materials science. AL techniques, which prioritize data points with the highest uncertainty, are evaluated for their potential to reduce the data needed to achieve accurate BBF predictions. The study examines various datasets, including ternary alloy compositions, inorganic materials, small molecules, and polymers, to determine AL’s efficiency across different dimensional and structured feature spaces. Findings suggest that AL enhances prediction accuracy over random sampling in low-dimensional settings, such as ternary alloys. However, for high-dimensional datasets, especially those with unbalanced distributions like polymers and inorganic materials, AL often performs comparably to or worse than random sampling. This indicates that AL’s effectiveness is context-dependent, with limited advantage in high-dimensional feature spaces common in materials informatics.
For more details, you can view the corresponding PDF file under the following link:
https://www.nature.com/articles/s41598-024-76800-4
Please consult also the Quantum Server Marketplace platform for the outsourcing of computational science R&D projects to external expert consultants through remote collaborations:
#materials #materialsscience #materialsengineering #computationalchemistry #modelling #chemistry #researchanddevelopment #research #MaterialsSquare #ComputationalChemistry #Tutorial #DFT #simulationsoftware #simulation