From optics to atmospheric models: Deep learning helps parameterize light scattering by non-spherical particles

Aerosol particles suspended in the atmosphere are important components of the earth-atmosphere system. These particles play a significant role in atmospheric physics and radiation processes. However, accurately and efficiently modeling light scattering by the particles, which are mostly non-spherical, is a fundamental problem according to Dr. Lei Bi, a scientist studying atmospheric radiation from Zhejiang University, China.

“In the past, the traditional look-up table (LUT) method was used to improve the efficiency of electromagnetic scattering calculation…” said Dr. Wei Han from the China Meteorological Administration (CMA). “…but with the increase of particle parameters, the volume of LUT becomes larger, which is quite inconvenient for further applications.”

Dr. Han, along with Xiaoye Zhang from the CMA as well as Prof. Bi and his team collaborated to develop a new compression algorithm for modeling that accounts for particle shape. They recently published their results in Advances in Atmospheric Sciences. “…we developed a deep learning (DL) method for computing the optical properties of non-spherical particles.” remarked Dr. Han.

By designing an optimal neural network architecture, the researchers successfully compressed the optical properties of the atmospheric particles in a super-spheroid shape space. This process is known as creating a deep neural network, or DNN, which can, with high accuracy, substantially compress high volumes of data into a manageable file size.

One of the collaborating scientists, Jinhe Yu stated “Using DNN model parameters, the file size is only 20MB, nearly 7000 times smaller than the original database (127GB), and the determination coefficient between the computed results and the truth values can be as high as 0.999.”

This accuracy rating allows users to store the database much easier without losing important data. Better storage techniques, such as DNN, can make database applications much more portable and suitable for current and future modeling endeavors. With this information at hand, the researchers examined DNN performance with unknown atmospheric particle sizes and shapes, finding that the predicted results have sufficient accuracy.

“Our findings are important for practical applications.” added Xiaoye Zhang, an academician of the Chinese Academy of Engineering. “Computing non-spherical particles' optical properties efficiently and accurately is critical to chemical data assimilation and chemical weather forecasting research.”

To further their research goals, this group of scientists plans to implement the DL scheme and DNN models into the Global/Regional Assimilation and Prediction Enhanced System (GRAPES) and the Chinese Unified Atmospheric Chemistry Environment (CUACE) systems to develop even more applications for this groundbreaking algorithm.