Machine Learning-based Denoising of Surface Solar Irradiance simulated with Monte Carlo Ray Tracing

TL;DR

This paper demonstrates that machine learning denoising significantly accelerates Monte Carlo radiative transfer simulations of surface solar irradiance, maintaining accuracy with fewer samples and enabling more efficient atmospheric modeling.

Contribution

The study introduces a machine learning denoising autoencoder that reduces computational costs of Monte Carlo ray tracing for surface irradiance by an order of magnitude while preserving accuracy.

Findings

Denoising diffuse irradiance from 1 sample per pixel is faster and more accurate than traditional methods.

Errors in direct irradiance denoising are localized to cloud shadow edges.

Adding atmospheric parameters improves denoising accuracy for diffuse irradiance.

Abstract

Simulating radiative transfer in the atmosphere with Monte Carlo ray tracing provides realistic surface irradiance in cloud-resolving models. However, Monte Carlo methods are computationally expensive because large sampling budgets are required to obtain sufficient convergence. Here, we explore the use of machine learning for denoising direct and diffuse surface solar irradiance fields. We use Monte Carlo ray tracing to compute pairs of noisy and well-converged surface irradiance fields for an ensemble of cumulus cloud fields and solar angles, and train a denoising autoencoder to predict the well-converged irradiance fields from the noisy input. We demonstrate that denoising diffuse irradiance from 1 sample per pixel (per spectral quadrature point) is an order of magnitude faster and twice as accurate as ray tracing with 128 samples per pixel, illustrating the advantage of denoising over larger sampling budgets. Denoising of direct irradiance is effective in sunlit areas, while errors persist on the edges of cloud shadows. For diffuse irradiance, providing additional atmospheric information such as liquid water paths and solar angles to train the denoising algorithm reduces errors by approximately a factor of two. Our results open up possibilities for coupled Monte Carlo ray tracing with computational costs approaching those of two-stream-based radiative transfer solvers, although future work is needed to improve generalization across resolutions and cloud types.