Dirty engineering data-driven inverse prediction machine learning model.

Area of Science:

• Metallurgy and Materials Science
• Computational Materials Design
• Machine Learning Applications

Background:

• Current machine learning (ML) in metallurgy primarily uses forward models to predict properties from conditions.
• These forward models often have a high input feature dimension (material conditions) versus a low output feature dimension (material properties).
• A need exists for inverse design models that deduce material conditions from desired properties.

Purpose of the Study:

• To develop and demonstrate a novel inverse design strategy for materials discovery.
• To identify new thermo-mechanically controlled processed (TMCP) steel alloy candidates.
• To propose a method for handling non-independent and identically distributed (IID) industrial engineering data.

Main Methods:

• Employed two independent inverse design approaches: metaheuristics-assisted inverse reading of forward ML models and a modified variational autoencoder.
• Utilized a rule-based thermodynamic calculation tool (Thermo-Calc.) for validation.
• Developed a protocol for treating non-IID industrial engineering data.

Main Results:

• Successfully implemented both inverse design strategies, yielding overlapping and consistent results.
• Pinpointed several novel TMCP steel alloy candidates.
• Validated the identified alloy candidates using thermodynamic calculations.

Conclusions:

• The developed inverse design strategy is effective for materials discovery in metallurgy.
• The combination of metaheuristics and variational autoencoders offers a powerful approach to inverse problems.
• The study provides a practical framework for utilizing industrial engineering data in ML-driven materials design.