Application of a Neural Network to Store and Compute the Optical Properties of Non-Spherical Particles

Abstract

Radiative transfer simulations and remote sensing studies fundamentally require accurate and efficient computation of the optical properties of non-spherical particles. This paper proposes a deep learning (DL) scheme in conjunction with an optical property database to achieve this goal. Deep neural network (DNN) architectures were obtained from a dataset of the optical properties of super-spheroids with extensive shape parameters, size parameters, and refractive indices. The dataset was computed through the invariant imbedding T-matrix method. Four separate DNN architectures were created to compute the extinction efficiency factor, single-scattering albedo, asymmetry factor, and phase matrix. The criterion for designing these neural networks was the achievement of the highest prediction accuracy with minimal DNN parameters. The numerical results demonstrate that the determination coefficients are greater than 0.999 between the prediction values from the neural networks and the truth values from the database, which indicates that the DNN can reproduce the optical properties in the dataset with high accuracy. In addition, the DNN model can robustly predict the optical properties of particles with high accuracy for shape parameters or refractive indices that are unavailable in the database. Importantly, the ratio of the database size (∼127 GB) to that of the DNN parameters (∼20 MB) is approximately 6810, implying that the DNN model can be treated as a highly compressed database that can be used as an alternative to the original database for real-time computing of the optical properties of non-spherical particles in radiative transfer and atmospheric models.

Analytical Prediction of Scattering Properties of Spheroidal Dust Particles With Machine Learning

“…F 12 /F 11 and F 34 /F 11 are the most difficult to train, due in part to many small values close to zero, and their MREs are 5.41% and 9.40%, respectively. Even so, the RMSEs of F 12 /F 11 , F 34 /F 11 , F 33 /F 11 , and F 44 /F 11 are all comparable and lower than their counterparts in recent work using a deep learning model to train a similar LUT for super-spheroids aerosols (Yu et al, 2022), illustrating that our NN performs well in predicting each single-scattering property.…”

“…This study is the first application of ML methods for aerosol particle scattering and their derivatives approximation at the same time, which differs from the past study that only predicts aerosol scattering properties with ML (Yu et al, 2022). Here, the rigor and constraints of aerosol optical properties (e.g., the normalization property of phase function) are retained to a degree at least comparable to the LUT or FD method.…”

The uncertainties in the laboratory-measured short-wave refractive indices of mineral dust aerosols and the derived optical properties: A theoretical assessment

“…Further efforts to improve the computational efficiency are crucial for calculating the optical properties of nonspherical particles with large size parameters. The use of GPU-accelerated computing and data-driven techniques shows promise in this regard (Bi et al, 2022;Yu et al, 2022).…”

“…The uncertainties caused by the calculation of optical properties are negligible compared to those in experiments. To reduce the computational burden, the neural network developed by Yu et al (2022) was used. The single particle optical properties of sphere models were calculated using the Lorenz-Mie theory (Bohren and Huffman, 2008).…”

NeuralMie (v1.0): An Aerosol Optics Emulator

“…Several past studies have applied machine learning to aerosol scattering and optics. Chen et al (2022) used a neural network to represent scattering by spheroidal dust particles, Yu et al (2022) trained a large neural network on a database of non-spherical particles to predict particle optics, and Ren et al (2020) trained a neural network to predict information about aerosol size distributions from photometer observations of AOPs. Both Thong and Yoon (2022) and Stremme (2019) trained neural networks to directly emulate a Mie scattering model.…”