Physics-Informed Neural Networks for Real-Time, Micro-scale Atmospheric Flow Prediction

The ability to precisely estimate velocity field in an atmospheric environment is useful to support flight operations. Conventional approaches require extensive sensor networks and resource-intensive numerical simulations, these combinations limit their viability for real-time operations. In this research, we introduce a framework based on Physics-Informed Neural Network (PINN) to enable real-time velocity field prediction while significantly curtailing the need for extensive sensor arrays and high computational cost. Combining the core physical principles or governing equations of atmospheric flow dynamics with PINN-based deep learning [7], offer a sensor-efficient way to estimate velocity field, which enables real-time atmospheric predictions. In this study, we will investigate the effectiveness of PINNs in different flow scenarios with various obstacle, evaluate their ability to produce reliable forecasts without any dependence on traditional methods. The outcomes aim to support various flight operations in the atmosphere.