Scalable Agent-Based Simulations for SARS-COV-2: A Case Study for Testing Strategies in the City of Brunswick
Please login to view abstract download link
Accurately modeling infectious disease spread and dynamics remains central to global public health research, particularly in light of the SARS-CoV-2 pandemic. Agent-based models (ABMs) are well- suited for studying complex individual-level dynamics and the effects of non-pharmaceutical interven- tions (NPIs). However, their computational intensity necessitates advanced high-performance computing (HPC) methods to enable realistic, large-scale simulations. We present a highly scalable ABM integrated into the MEmilio pandemic simulation framework, specif- ically developed for high-resolution pandemic modeling at city and regional scales. The model supports a comprehensive range of NPIs implemented in a highly customizable and modular manner, including detailed testing strategies, mask mandates, remote work policies, school closures, and quarantine mea- sures. This modularity allows researchers and policymakers to evaluate diverse intervention scenarios and their combinations efficiently. In a case study of Brunswick, Germany during the Alpha variant wave (March 1 to May 30, 2021), we simulated over 370,000 agents and calibrated the model to real-world epidemiological data including ICU occupancy and mortality rates. Results indicate that asymptomatic testing effectiveness depends strongly on symptomatic testing capacity, with untargeted testing becoming more crucial as targeted testing increases, while contributing little when symptomatic testing is limited. Furthermore, we employ sophisticated hybrid parallelization using OpenMP and MPI in an optimized C++ implementation to address key computational challenges such as individualized transmission risk calculations across diverse location types. The model demonstrates exceptional performance, simulat- ing one million agents over 30 days in under a minute on a standard consumer laptop. Weak scaling efficiency reaches 82% from 1 to 4 cores and 33% from 1 to 32 cores, enabling thousands of parallel simulations within hours for large-scale scenarios and extensive batch runs. Overall, our model enables rapid, detailed scenario analysis at city, regional, and national scales, of- fering substantial modeling flexibility combined with computational efficiency while reliably reflecting observed pandemic behavior and representing a significant advancement in agent-based epidemic simu- lation.
