Illustration of our LSTM-FCNN composed of a Siamese LSTM network and a FCNN with two hidden layers. Details about the LSTM cells are shown in Fig. 6. Both sub-LSTM networks for image one (top, input denoted by S _1k) and image two (bottom, input denoted by S_2k) are the same LSTM network (i.e., with the same weights in the neural network) composed of 2 layers (LSTM A and LSTM B) where during the training process, a dropout rate of 20% is used. We use only the last outputs $h_{1N_{\text{sl},1}}\text{\hspace{0.17em}and\hspace{0.17em}}{h}_{2N_{\text{sl},2}}$ from the second layer. These outputs are concatenated to be the input for the FCNN with two hidden layers using ReLU activation functions and a sigmoid activation for the final output. Given that the data is normalized the output of the network, t_norm and Var, needs to be converted to get the time delay Δt (see Eq. (21)) and the uncertainty of the time delay σ_Δt (see Sect. 4). The initial cell state c₀ and hidden state h₀ have the size N_hidden with all elements being equal to zero.