Fig. 5: Example of data distribution shift resulted machine learning model mistakes.
The task is to predict material mechanical strength from feedstock material SEM images. The authors take different scans of the same microstructure using different microscope settings. Ideally, ML model predictions should only depend on the microstructure content, not the microscope settings. However, results show that darker images are consistently predicted to have bigger ultimate compressive strength (UCS) values, even with image normalization61. The x-axis shows experiment id. The first row shows example images from the given experiment. The second row illustrates the effect of one image normalization method (i.e., histogram equalization).
