A high-performance framework for SARS-CoV-2 via data-driven agent-based modeling of infectious dynamics

RWTH Publications (RWTH Aachen)

Abstract

Mathematical modeling is a central tool to aid policy-makers, informing the design and implementation of effective measures to contain pathogen spread.Aggregated compartmental models offer computational efficiency, but typically assume homogeneous mixing.Agent-based models (ABMs) overcome this limitation by simulating individuals explicitly, capturing heterogeneous contact patterns.The aim of this thesis is to advance and apply a data-driven ABM for infectious disease simulations with a particular focus on Severe Acute Respiratory Syndrome Coronavirus 2 (SARS-CoV-2).We address challenges such as computational feasibility, faithful implementation of non-pharmaceutical interventions (NPIs), and the sensitivity of epidemic predictions to model initialization assumptions.The central contribution of this dissertation is the substantial advancement of an agent-based model, the MEmilio-ABM, which follows a location-based design in which agents move between discrete location types such as households, workplaces, schools, and social venues. Agent mobility is implemented through either trip chains or mobility rules, and individual infectiousness is modeled through a continuous viral load trajectory.The model is implemented in C++20 with shared- and distributed-memory parallelization and achieves linear algorithmic complexity in the number of agents, with wall-clock runtime per time step scaling linearly up to 256 million agents on a CPU-based HPC infrastructure, enabling large-scale ensemble analyses.We demonstrate policy analysis capabilities through a retrospective Brunswick, Germany case study during the Alpha variant period, with key parameters fitted to reported public health data. Testing of symptomatic individuals proves central to outbreak control in this setting, and a counterfactual analysis indicates that a five-fold increase in symptomatic testing, without a lockdown, could have achieved comparable reductions in the number of deaths.We show that quarantine duration is more consequential than the degree of contact reduction during isolation, with five to eleven days achieving substantial mitigation largely independent of degree of contact reduction.A novel coupling framework between the pedestrian dynamics model Vadere and MEmilio-ABM enables a controlled comparison between two initialization strategies: transmission-informed initialization, which preserves the social clustering of initial infections, and standard uniform initialization based on case numbers.Across four outbreak scenarios, 100 stochastic replications, and demographic structures from Germany, France, and the United States, uniform initialization systematically overestimates epidemic growth, with prediction errors reaching up to 46% in cumulative infections after a 10-day simulation period for the most clustered settings. Our analysis identifies social clustering of the initial outbreak event as the key determinant. Transmission-informed initialization produces local saturation within connected social units, constraining epidemic growth — a dynamic that uniform initialization cannot replicate.This dissertation advances an HPC-capable, data-driven ABM framework that is applied to NPI policy evaluation and quantifies initialization-driven prediction bias in epidemic simulations.

Authors 1

1. Sascha Korf corresponding Aachen

   RWTH Aachen University

   Affiliation as printed

   RWTH Aachen