Improving Spectral Synthesis Performance By Up To 5 Orders Of Magnitude With The Discrete Integral Transform Approach

van den Bekerom, Dirk

Permalink

https://hdl.handle.net/2142/111189 Copy

Description

Title

Improving Spectral Synthesis Performance By Up To 5 Orders Of Magnitude With The Discrete Integral Transform Approach

Author(s)

van den Bekerom, Dirk

Contributor(s)

• Pannier, Erwan
• Mishra, Pankaj

Issue Date

2021-06-23

Keyword(s)

Theory and Computation

Abstract

The increasing number and size of large Line-By-Line (LBL)-databases, routinely exceeding hundreds of millions of lines, calls for spectral synthesis algorithms that can process these large number of lines efficiently. In the special case where the lineshape is identical for every line in the database, the spectrum can be rapidly synthesized by the convolution of a stick-spectrum and the constant lineshape. We propose an integral transform that extends this procedure to databases with unique lineshapes for each line, by generalizing the stick spectrum to a 3D “lineshape distribution function”, which is a function of line position and Gaussian- $\&$ Lorentzian width axes. A fast discrete version of the integral transform based on the Fast Fourier Transform (FFT) is formulated for which the computation time scales with $c_1N_i + c_2N_\nu\log{N_\nu}$ (where $N_i$ is the number of lines, $N_\nu$ the number of spectral points, and $c_1$ and $c_2$ are small constants), which is much more efficient than the traditional $N_i \times N_\nu$ scaling. Furthermore, since the lineshape calculation is no longer a performance bottleneck, lineshapes such as exact Voigt profiles without cutoff can be computed with arbitrary precision. Our procedure can be implemented in existing spectral codes, providing a new level of accuracy and performance for high temperature, high resolution spectroscopy. The procedure was benchmarked against traditional methods within the open source package RADIS: A benchmark HITEMP CO$_2$ spectrum consisting of 1.8M lines and 200k spectral points took 3.1s, an improvement of $\sim300\times$. The new procedure has since been implemented as the default synthesis method for RADIS. The method's calculations can be easily divided up and are ideally suited for parallel computation. An experimental GPU procedure, also implemented in RADIS, was used to synthesize a CDSD-4000 CO$_2$ spectrum of 200M lines and 300k spectral points, which took only 150 ms per spectrum regardless of temperature (after an initialization of $\sim$5 s), or $4\cdot10^{14}$ lines$\times$spectral points/s -- an improvement of 5 orders of magnitude compared to the state-of-the-art -- with no appreciable loss in accuracy.

Publisher

International Symposium on Molecular Spectroscopy

Type of Resource

Text

Language

eng

Permalink

http://hdl.handle.net/2142/111189

DOI

10.15278/isms.2021.WE03

Owning Collections