Hybrid epidemiological models for efficient insight on the individual scale: a contribution to green computing

Abstract

Mathematical modeling has been proven to be of great aid for many different domains by complementing classical physical experiments. A large variety of models has already helped to create a better understanding of physical, biological, or epidemiological processes. As infectious disease dynamics are highly driven by heterogeneous human contact patterns and human behavior, agent based models are most suitable to investigate underlying structures and to generate insights into individual-scale properties. However, the additionally gained knowledge has to be paid by a huge computational effort as the complexity grows linearly or quadratically with the number of considered agents inside a region or location. In our paper, we show how hybrid epidemiological models can reduce the computational effort by more than 90 % without losing the required depth in information on the individual scale. In order to keep the computational effort small, agent based models are often only developed for a region or time-frame of interest, e.g., neglecting information from coupled regions. By using well established metapopulation models for coupled regions or time-windows with less stochastic influence, agent-based model predictions can be improved and individual-scale information of the focus area can be retained without substantial increase in effort or energy consumption. Although nowadays, high-perfomance computing (HPC) techniques allow the simulation of increasingly large problems, HPC also produces increasingly large CO2 footprints. Hybrid epidemiological models can complement software and hardware optimizations to reduce energy consumption by only computing the necessary level of detail where needed, using dynamically developing summary statistics where possible. In this talk, we will briefly explain the temporal and spatial hybrid models and then present results of a spatial hybrid model for the city of Munich and its neighboring or connecting counties. Although our demonstration is on the use case of Sars-CoV-2, the approach can be used for various sets of epidemiological models of different scale.