Notes. These architectures are tested for the best combination of inclination and rotation axis (shown in Fig. 17),for which 681 facets are visible. The inputs to the neural network are 64 Lambertian fluxes (8 orbit locations ☓8 rotation phases) for the wavelength of 550 nm. The architectures are trained until the validation loss does not decrease for 10 consecutive epochs. The periodic convolution uses a 1 ☓ 3 kernel, as shown in Fig. 10.