Neural surrogates for pathogen-importation forecasting
Making a global, coupled ODE model tractable for Bayesian inference and probabilistic forecasting.
The problem
Forecasting pathogen importation across a connected world requires both mechanistic realism and repeated model evaluation. The underlying metapopulation model links 238 countries and territories through air and ground transportation, making direct likelihood-free inference computationally demanding.
My contribution
I conceived and developed a mechanistically informed neural-surrogate methodology to test whether machine-learning approximations can preserve the useful behaviour of the coupled ODE model while reducing evaluation cost. I designed temporally conditioned surrogate architectures and comparative graph-attention and recurrent models that incorporate transportation-network and location-level features.
I built the end-to-end Python research pipeline: mechanistic simulation, parameter-space sampling, surrogate training, SMC-ABC inference, posterior uncertainty propagation, validation against observed first-case records, and computational benchmarking.
Methods and tools: Python, PyTorch, PyTorch Geometric, graph attention networks, LSTM/GRU models, SMC-ABC, ODE simulation, uncertainty quantification, spatiotemporal networks.
Status: Postdoctoral research; manuscript in preparation.