Microstructure-Mediated Inverse Design of Thermal Barrier Coatings via Interpretable, Data-Efficient Machine Learning

Area of Science:

• Materials Science
• Mechanical Engineering
• Computational Materials Science

Background:

• Microstructure control is vital for thermal barrier coating (TBC) performance.
• Complex process-structure relationships hinder predictive TBC design.
• Data-driven methods face limitations due to scarce data and poor interpretability.

Purpose of the Study:

• To develop a closed-loop machine learning framework for TBC design.
• To decouple process-property relationships using microstructure as an interpretable intermediary.
• To enable accurate TBC optimization with limited experimental data.

Main Methods:

• Implemented a deep neural network for automated microstructure feature extraction from SEM images.
• Developed forward prediction models for microstructure features (pores, unmelted regions, cracks).
• Created an inverse design engine to translate target microstructures into manufacturing parameters.

Main Results:

• Achieved high R-squared values for microstructure feature prediction (0.896 for pores, 0.878 for unmelted regions, 0.727 for cracks).
• Uncovered dominant mechanisms governing thermal energy distribution and particle trajectory.
• Demonstrated high accuracy in structural reproduction (92%) and parameter prediction (88%).

Conclusions:

• The framework enables data-efficient, interpretable materials design for TBCs.
• Designed coatings achieved 35-46% reduction in thermal conductivity at 1000°C.
• This approach shifts TBC development towards intelligent precision design.