Active Learning for Rapid Targeted Synthesis of Compositionally Complex Alloys.

Area of Science:

• Materials Science
• Chemical Engineering
• Computational Materials Science

Background:

• Advanced materials synthesis is increasingly focused on complex compositions, posing significant challenges in achieving precise elemental ratios.
• Optimizing synthesis parameters for these materials is time-consuming and becomes exponentially more difficult with higher compositional complexity.
• Current methods, even with experienced operators, struggle with consistency when synthesis parameters are coupled, hindering rapid material exploration.

Purpose of the Study:

• To demonstrate an active learning (AL) approach for optimizing the physical vapor deposition (PVD) synthesis of thin-film alloys with up to five principal elements.
• To compare the efficacy of Gaussian process (GP) and random forest (RF) based AL models for materials synthesis optimization.
• To explore the application of transfer learning using pre-trained models for accelerating the discovery of novel material compositions.

Main Methods:

• An active learning (AL) framework was implemented to guide the optimization of physical vapor deposition (PVD) synthesis parameters.
• Gaussian process (GP) and random forest (RF) machine learning models were employed within the AL loop to predict optimal synthesis conditions.
• Transfer learning strategies were investigated, utilizing models trained on simpler (ternary, quaternary) or related compositions to pre-train models for complex (quinary) alloys.

Main Results:

• The best performing AL models successfully identified synthesis parameters for a target quinary alloy within 14 iterations.
• Random forest and Gaussian process models demonstrated improved prediction accuracy when pre-trained on lower-dimensional (ternary, quaternary) systems compared to models trained solely on quinary data.
• Pre-training with samples sharing common elements with the target composition also yielded performance improvements, showcasing the adaptability of the AL approach.

Conclusions:

• Active learning (AL) provides a powerful and adaptable strategy for accelerating the optimization of physical vapor deposition (PVD) synthesis for compositionally complex thin-film alloys.
• The use of transfer learning significantly enhances the efficiency of AL models by leveraging knowledge from related material systems.
• This approach holds broad potential for accelerating the exploration and discovery of a wide range of advanced, compositionally complex materials.