Machine-Learning-Assisted Discovery and Accelerated Synthesis of Metal Phosphosulfides

Area of Science:

• Materials Science
• Inorganic Chemistry
• Computational Materials Science
• Machine Learning in Materials Discovery

Background:

• Metal phosphosulfides are promising multifunctional materials but face significant synthesis challenges.
• Previous research focused on individual compounds, hindering broad exploration of the phosphosulfide material space.
• Accelerated discovery workflows are needed for complex inorganic materials.

Purpose of the Study:

• To computationally screen hypothetical ternary phosphosulfides for stability and electronic properties.
• To develop a machine learning model for accurate band gap prediction.
• To establish a high-throughput experimental synthesis and characterization method for phosphosulfides.

Main Methods:

• Density Functional Theory (DFT) calculations to evaluate 909 hypothetical ternary phosphosulfides.
• Development of a multi-fidelity machine learning model for band gap prediction.
• Thin-film combinatorial libraries for high-throughput synthesis and characterization.

Main Results:

• Identified 19 new thermodynamically stable ternary phosphosulfides, including novel Si- and Ge-based compounds.
• The machine learning model accurately predicted experimentally relevant band gaps.
• Successfully synthesized over 100 unique compositions per experiment, yielding four distinct phosphosulfide compounds in four experiments.

Conclusions:

• Accelerated workflows integrating theory, AI, and high-throughput experimentation are effective for discovering challenging inorganic materials like phosphosulfides.
• The developed methods enable rapid exploration and synthesis of novel phosphosulfide compositions.
• This approach paves the way for broader discovery of functional phosphosulfide materials.