Architecture of our best performing neural network model. The input data is a light curve of 480 timesteps, the output is a list of two real numbers between 0 and 1: one for the probability of detection of an exocometary transit and one for the position of the transit in the light curve. Convolutional layers are denoted Conv <kernel size> <number of feature maps>, max pooling layers are denoted MaxPool <window length> <stride length>, fully connected layers are denoted Dense <number of units>, LSTM layers are denoted LSTM <units> <dropout>, GRU layers are denoted GRU <unit> <droupout> <recurent dropout>, and the squeezed-excitation blocks are denoted SEB.