How the Model Works

CityCOVID includes the places that people go to according to their hourly schedules for the activities that they engage in during the course of their day. Individual people are represented as software ​“agents” in the model, each agent having a set of characteristics (such as socio-demographics, neighborhood of residence) and behaviors (such as their schedules of where and when people go outside the home, how they react to having a disease, or in response to a stay-at-home-order). During a simulated day, agents move from place-to-place, hour-by-hour, engaging in social activities and interactions with agents who are located at the same place and engaged in the same activity.

Each agent in the model includes their own individualized disease progression dynamic that determines transitions between possible COVID-19 disease states. Transitions have functional dependence on heterogeneous agent attributes, exposure through co-location over time with infected individuals, and other factors such as the effects of self-protective behaviors on viral transmissibility.

CityCOVID is being used to understand the spread of COVID-19 through the population by forecasting new infections, hospitalizations, and deaths. The model provides a computational platform for investigating the potential impacts of non-pharmaceutical interventions (NPIs) for mitigating the spread of COVID-19.

Key Features

Agent-based Modeling Approach

With our detailed model approach we are able to simulate specific and realistic NPIs, such as imposing restrictions on specific types of activities, or exploring alternate school learning models, closely resembling those being considered by public health officials. Unlike statistical approaches, agent-based modeling allows us to directly investigate novel interventions that have not been implemented yet. CityCOVID is built using Argonne’s Repast agent-based modeling toolkit and the Chicago Social Interaction Model (ChiSIM) framework.

Machine Learning + HPC

We are developing large-scale machine learning algorithms on high-performance computing resources to estimate key CityCOVID model parameters, such as how social distancing affects disease transmission, and to quantify the uncertainties in the epidemiologic forecasts we produce. This is enabled by Argonne HPC software technologies, including the EMEWS framework and the Swift/T workflow language.

Crisis Response

The questions being asked by public health officials have evolved over the course of the pandemic and change daily. The need is for robust modeling approaches that can quickly respond to shifting areas of concern. By combining sophisticated agent-based modeling, high-performance computing, and machine learning we are developing just-in-time capabilities to address this need.