Neural network and graph neural network surrogate models for spatially resolved models in computational epidemiology

Kurzfassung

Expert compartment models based on ordinary differential equations (ODE) are fundamental tools in the field of epidemiological modeling, providing high quality predictions and assessment of containment strategies as became evident once more with the emergence of the SARS-CoV-2 virus. Solving the differential equations of a spatially resolved compartment model for all 400 counties of Germany with included contact-reducing measures requires significant computation time when done thousands and thousands of times. In contrast, machine learning models, after extensive training, can offer rapid predictions. We suggest to replace the ODE-based model and the Graph-ODE model with neural networks and a Graph Neural Networks trained on data generated from the expert model to provide single or multiple predictions very fast with only a little trade off in accuracy. This suggested approach yields great potential for fast approximation of complex simulations, facilitating the development of public web applications that provide user-friendly and time-efficient simulations for decision-makers.