Using neural networks to calibrate agent based models enables improved regional evidence for vaccine strategy and policy.

Distribution and administration strategy are critical to successful population immunization efforts. Agent-based modeling (ABM) can reflect the complexity of real-world populations and can experimentally evaluate vaccine strategy and policy. However, ABMs historically have been limited in their time-to-development, long runtime, and difficulty calibrating. Our team had several technical advances in the development of our GradABMs: a novel class of scalable, fast and differentiable simulations. GradABMs can simulate million-size populations in a few seconds on commodity hardware, integrate with deep neural networks and ingest heterogeneous sources. This allows for rapid and real-world sensitivity analyses. Our first epidemiological GradABM (EpiABMv1) enabled simulation interventions over real million-scale populations and was used in vaccine strategy and policy during the COVID-19 pandemic. Literature suggests decisions aided by evidence from these models saved thousands of lives. Our most recent model (EpiABMv2) extends EpiABMv1 to allow improved regional calibration using deep neural networks to incorporate local population data, and in some cases different policy recommendations versus our prior models. This is an important advance for our model to be more effective at vaccine strategy and policy decisions at the local public health level.