Physics-informed machine learning for the COVID-19 pandemic: Adherence to social distancing and short-term predictions for eight countries

Abstract

Background: The analysis of COVID-19 infection data through the eye of Physics-inspired Artificial Intelligence leads to a clearer understanding of the infection dynamics and assists in predicting future evolution. The spreading of the pandemic during the first half of 2020 was curtailed to a larger or lesser extent through social distancing measures imposed by most countries. In the context of the standard Susceptible-Infected-Recovered (SIR) model, changes in social distancing enter through time-dependent infection rates. Methods: In this work we use machine learning and the infection dynamical equations of SIR to extract from the infection data the degree of social distancing and, through it, assess the effectiveness of the imposed measures. Results: Quantitative machine learning analysis is applied to eight countries with infection data from the first viral wave. We find as two extremes Greece and USA where the measures were successful and unsuccessful, respectively, in limiting spreading. This physics-based neural network approach is employed to the second wave of the infection, and by training the network with the new data, we extract the time-dependent infection rate and make short-term predictions with a week-long or even longer horizon. This algorithmic approach is applied to all eight countries with good short-term results. The data for Greece is analyzed in more detail from August to December 2020. Conclusions: The model captures the essential spreading dynamics and gives useful projections for the spreading, both in the short-term but also for a more intermediate horizon, based on specific social distancing measures that are extracted directly from the data.