A multi-fidelity active learning framework for accelerated alloy design

Machine learning (ML) driven methodologies are more efficient than traditional trial-and-error-based materials design; however, extensive training datasets are required for their development. To overcome this challenge, we developed a multifidelity active learning (MFAL) framework that significantly improves optimization efficiency compared to typical evolutionary methods. This framework strategically balances the computational effort between high- and low-fidelity evaluations, thereby reducing experimental burden while effectively guiding the search towards optimal compositions. MFAL was applied to high-entropy alloy (HEA) design, enabling the discovery of compositions that approach the theoretical minimum coefficient of thermal expansion (CTE). The optimized composition was efficiently identified, validating the developed framework as robust and scalable. The quantitative study shows that the MFAL framework achieved near-optimal convergence in around 75 iterations, using only 55–65% costly high-fidelity evaluations, in contrast to single-fidelity methods that needed 100% high-fidelity assessments. Compared to trial-and-error approaches, MFAL delivers five-fold improvement in optimization speed while requiring 40% fewer high-fidelity evaluations than conventional methods. It demonstrates optimization of experimental/computational resources by strategically focusing expensive evaluations on the most promising compositional areas. MFAL has the potential for rapid development of next-generation alloys with customized properties and insights into composition-8093property relationships. • A multi-fidelity active learning (MFAL) framework is developed for HEA alloy design. • MFAL identifies the alloy with minimum coeAicient of thermal expansion. • MFAL provides significant resource savings compared to global optimizer.