Sparse Surrogate Modelling for Combustion Chemistry via Time-Lag Autoencoders and Gradient-Based Clustering | Synapse