This research aims to develop a surrogate model of agent-based simulations, particularly for on-demand mobility services. A large-scale agent-based simulator, called MATSim, will be adopted for generating scenarios for calibrating the surrogate model. MATSim simulation with a 10% population sample in a scenario like Stockholm might take several hours. Since large scenarios like an entire city can involve several computation, running a job with MATSIM is likely to consume large amounts of memory (RAM). MATSim uses multi-threading to accelerate computing speeds. Francesco Guglielmi from Politecnico Milano, visiting doctoral students at KTH, will join us to collaborate on the analysis.