Martensite Start Temperature Prediction through a Deep Learning Strategy Using Both Microstructure Images and...

Area of Science:

• Materials Science
• Metallurgy
• Computational Materials Science

Background:

• Martensite start temperature (Ms) prediction is crucial for steel properties.
• Existing models often neglect microstructural factors like morphology, limiting accuracy.
• Digitizing microstructural data for prediction remains a significant challenge.

Purpose of the Study:

• To develop an advanced predictive model for Ms in medium-Mn steels.
• To integrate both chemical composition and microstructure image data for enhanced prediction accuracy.
• To overcome the limitations of traditional methods in handling complex microstructural features.

Main Methods:

• Established an experimental database for medium-Mn steels.
• Developed and trained a convolutional neural network (CNN) model using composition and microstructure images.
• Preprocessed data to create interaction matrices between numerical and image data.

Main Results:

• The CNN model demonstrated superior accuracy and rationality compared to traditional AI models.
• The model effectively utilizes microstructure images, mitigating overfitting through data augmentation.
• Data preprocessing to capture interaction information proved more effective than direct data linking.

Conclusions:

• The developed CNN model offers a more accurate and comprehensive approach to Ms prediction.
• Integrating microstructural images alongside composition significantly improves predictive capabilities.
• This deep learning strategy provides a robust framework for materials property prediction using multimodal data.