Deep Learning Prediction and Visualization Analysis of Microstructure—Tensile Properties of Medium Manganese Steel

In recent years, many studies have focused on modeling the prediction of microstructure and mechanical properties. Traditional methods mainly rely on physical models and empirical formulas, typically requiring extensive experimental data and struggling to describe complex microstructural characteristics accurately. This paper proposes a deep learning model based on an FPN (Feature Pyramid Network) and CNN (Convolutional Neural Network) for extracting features from microstructural images of Fe-C-Mn-Al-Cr steel processed by vacuum smelting and subjected to eight different heat treatment temperatures to predict its mechanical tensile properties. The experimental results show that the EfficientNetFPN model achieves R2 values of 0.96, 0.97, and 0.97 in predicting tensile strength (UTS), yield strength (YS), and elongation at break (TEL), respectively, outperforming models such as EfficientNet, ResNet50, ResNet18, and VGG16 in prediction accuracy. Compared to traditional CNN models, this method fully leverages the feature extraction advantages of EfficientNetFPN, enabling more comprehensive utilization of microstructural information and thereby improving prediction accuracy. Through feature visualization from low to high levels, the model’s understanding of microstructural images can be elucidated, and key microstructural features affecting material performance can be identified.