An approach for full space inverse materials design by combining universal machine learning potential, universal...

Area of Science:

• Materials Science
• Computational Materials Science
• Machine Learning in Materials

Background:

• Traditional materials design often requires prior knowledge of atomic composition and crystal structure.
• Automating the inverse design process for materials with specific physical properties is a significant challenge.

Purpose of the Study:

• To present a fully automated full space inverse materials design (FSIMD) approach.
• To enable the discovery of materials with target physical properties without predefined structural or compositional information.
• To demonstrate the application of FSIMD for optimizing cohesive energy and bulk modulus.

Main Methods:

• Training a universal machine learning potential (UPot) and a universal bulk modulus model (UBmod) using density functional theory data.
• Utilizing transfer learning to enhance model generalizability across diverse material systems (42 elements).
• Integrating UPot and UBmod with optimization algorithms and enhanced sampling techniques.

Main Results:

• The FSIMD approach successfully identified NaCl-type ZrC as the material with the highest cohesive energy.
• Diamond was identified as the material possessing the largest bulk modulus.
• The approach demonstrated capability for multi-objective property design, with accuracy dependent on training data quality.

Conclusions:

• The developed FSIMD approach offers a novel, automated pathway for inverse materials design.
• This method significantly reduces the prior knowledge required for discovering materials with targeted functionalities.
• FSIMD holds potential for accelerating the discovery of advanced materials for various practical applications.