| Summary: | In precipitation hardening metallic materials, the size and volume fraction of precipitation phases are regarded as primary microstructural parameters to control the strength instead of others. Why? In this research, a supervised learning approach was developed to correlate γ′ precipitation microstructures with hardness based on experimentally observed 483 scanning electron microscope (SEM) images comprised with different γ′ precipitates. First, up to 23 descriptors were defined and extracted numerically as training inputs from SEM images by pattern recognition techniques. Then, 10 descriptors were further selected to reduce computational cost of deep neural network (DNN) with the assistance of shallow neural network (SNN). Furthermore, to improve the accuracy of DNN, new training sets were proposed to combine these 10 descriptors with two more descriptors: area distribution and one heat treatment parameter - cooling rate. In conclusion, the supervised learning approach was proven to outperform the prediction of existing physics-based constitutive models.
|
|---|
