Combining Simulated and Real Data for Near-Term Forecasting of Nonstationary Dynamic Processes

Our team is developing algorithms and techniques to model the COVID-19 outbreak. Using a hybrid physics-based and machine learning model, we have developed algorithms that utilize mobility data and high-fidelity simulations to model how the pandemic spread has occurred. Our algorithmic framework already demonstrated the ability to accurately backward predict nonstationary data on short timescales (days to weeks) across the country. We are currently improving our predictive algorithms that allow for accurate forward predictions in nonstationary environments without the need for extensive training at the onset of a new event. Our general solution to modeling complex dynamic systems, as applied to the COVID pandemic, is to combine physics-based individual infection simulation data with observational data at the individual and population level, to develop flexible deep learning models with built-in uncertainty quantification mechanisms, to cope with partially observed complex dynamics for accurate near-term prediction. This technique is broadly applicable for predicting behaviors of non-stationary dynamical systems, such as disease or behavior models, without reliance on stationary longitudinal observation data over a time horizon. We will analyze the likely behavioral patterns of interactions among individuals and how it relates to the spread of COVID-19 (and potentially related infectious diseases) during the modeling stage. The project will generate real-time predictions for new cases and hotspots within 7 days for less than 10% prediction error.  In particular, we will establish a connection between the agent-based spatial movement and encounter models with the population level infection dynamics, where the former provides mechanistic foundation to constrain the possible model class and the latter relates directly to the observations on the population level. We use Bayesian inference to learn the model and quantify uncertainty of it. Predictions and credible intervals are formed by forward simulating the learned model. We will deliver visualized results and prognoses of future infection behaviors including number of infected in a region and locations of possible outbreaks.