Physics-Informed Neural Networks for Statistical Analysis of Epidemic Dynamics on Complex Networks.

Source abstract

We develop a graph-aware physics-informed neural network (PINN) for reconstruction and parameter estimation in a network susceptible-infected-susceptible model under sparse noisy observations. The network represents node infection probabilities while the loss enforces adjacency-coupled dynamics, data fit, and initial conditions. In a 20-node Barabasi-Albert experiment with 10 observed nodes, three initialization seeds give mean transmission and recovery estimates 0.7352 and 0.5483 for true values 0.8 and 0.6, with mean full-network RMSE 1.83\times10^{-3}. Accurate trajectory recovery alongside rate bias illustrates the distinction between predictive accuracy and mechanistic identifiability.