Architecture used to retrieve planetary albedo maps from normalized phase curves. Several logic layers with ReLU activation functions are used to scale the output map by adding and multiplying by two single values. The scaling part of the network decreases the loss from 0.0197 to 0.0162 (for the ideal geometry as shown in Fig. 17). The output of the neural network is between 500 and 1000 values, depending on the number of facets that is visible for the specific geometry. Neurons in the bottom layers have biases but neurons in the top layers do not.