Coupling mixture designs, high-throughput experiments and machine learning for accelerated exploration of multinary systems

• Published in: Materials & Design 2023-07, Vol.231, p.112055, Article 112055
• Main Authors: Garel, Elise, Parouty, Jean-Luc, Van Landeghem, Hugo, Verdier, Marc, Robaut, Florence, Coindeau, Stéphane, Boichot, Raphaël
• Format: Article
• Language: English
• Subjects: Artificial Intelligence; Combinatorial design; Computer Science; Condensed Matter; High Throughput Experiment; Machine learning; Materials Science; Physics; PVD
• ISSN: 0264-1275; 0261-3069; 0264-1275
• DOI: 10.1016/j.matdes.2023.112055

[Display omitted] •We propose combining Mixture Design and Machine Learning to efficiently explore the Nb-Ti-Zr-Cr-Mo system in thin films.•Among all the models tested, Neural Networks are the more accurate to predict phase class, hardness and Young’s modulus.•Results show no clear indication of positive cocktail effect beyond three elements.•The optimal mechanical properties are systematically found far from the equimolar composition. Current societal challenges, such as climate change and resource depletion, highlight an unprecedented need for disruptive innovation in materials science. Significant breakthroughs are expected in multinary materials whose efficient exploration necessitates dedicated strategies. The exploration of a refractory high entropy alloy Nb-Ti-Zr-Cr-Mo is proposed here as test case for a new strategy. Based on the proven methodology of mixture design and on combinatorial thin film metallurgy, the composition space is explored by a limited number of chosen gradients to build an alloy library comprising hardness and ductility, two antagonistic properties. The workflow is showcased here by studying the properties of the as-grown graded film, which presents wide amorphous domains and contrasted mechanical properties. This experimental dataset then trains machine learning models to provide continuous predictions of the alloy properties over the entire composition space. We show that optimal alloy properties are expected close to the binary edges of the quinary.