Physics-informed neural networks for modeling atmospheric radiative transfer

Understanding the radiative transfer processes in the Earth’s atmosphere is crucial for accurate climate modeling and climate change predictions. These processes are governed by complex physical phenomena , which can be generally modeled by the radiative transfer equation (RTE). Solutions to the RTE are obtained by various methods including numerical (standard RTE solvers), stochastic (Monte-Carlo), and data-driven (machine-learning) approaches. This paper introduces a novel numerical approach utilizing a Physics-Informed Neural Network (PINN) to solve the RTE in atmospheric scenarios, applying physics constraints in a machine-learning framework. We show that our PINN model offers a flexible and efficient solution, enabling the simulation of radiance values using plane-parallel atmosphere, and under diverse conditions, including clouds and aerosols. • Introduced a novel neural network model for atmospheric radiative transfer . • Achieved high precision, matching state-of-the-art solvers for aerosols and clouds. • Proved theoretical bounds ensuring method convergence and derived accuracy metrics.