Ensemble methodology incorporating uncertainty quantification applied to modeling COVID-19 outbreaks

Resumo

Mathematical models are essential for understanding physical phenomena and making predictions. However, complex systems often exhibit dynamics that a single model cannot fully capture. To address this, ensemble methods combine the individual outputs of multiple models to improve performance, specifically aiming to enhance the accuracy of the final prediction. In this work,  we construct an ensemble integrating different types of models, including mechanistic and machine learning models. The mechanistic component comprises epidemiological compartmental models, while the machine learning component employs a physics-informed neural network (PINN). Each individual model was calibrated based on real COVID-19 data from the city of Recife, with the mechanistic models specifically being treated by the Markov Chain Monte Carlo (MCMC) method. The uncertainty measures of the models were evaluated to verify the robustness of the simulations. The final response was then determined by combining the individual outputs of each model. According to our preliminary results, the ensemble demonstrated an advantage by effectively highlighting the strengths of each model.