CNN dataset for Enhancing spatial-spectral interferograms of attosecond pulses via weak value amplification and deep learning

Numerical simulations of WVA-based frequency-domain interferometry under shot noise and readout noise, covering time delays in the range of 0 as $< \tau <$ 10 as. Both networks were trained under identical conditions: an initial learning rate of 0.001, a maximum of 1000 iterations, and Training:Validation ratios of 8000:2000. For the CNN-regressor, each grayscale interferogram is assigned a continuous $\tau$ value as its label. In contrast, the CNN-classifier categorizes $\tau$ into 100 discrete classes (0 as, 0.1 as 0.2 as... 9.9 as), where each class represents a bin centered at a specific $\tau$ value. To emulate realistic experimental variability, the actual $\tau$ values within each class are randomized uniformly around the bin center with a tolerance of $\pm$ 0.05 as, representing the time jitter.