HELIOS-K: GPU Opacity Calculator

• HELIOS-K is an open-source, GPU-accelerated opacity calculator that condenses extensive spectroscopic line lists into efficient opacity and k-distribution tables.
• It employs advanced Voigt-profile algorithms and k-distribution methods to process up to 10^8 spectral lines per second on NVIDIA GPUs.
• Optimized in CUDA C++, HELIOS-K features self-tuning kernels and asynchronous I/O, significantly accelerating opacity computations for exoplanet and climate models.

HELIOS-K is an open-source, GPU-accelerated opacity calculator designed for the rapid computation of molecular and atomic opacities required in exoplanetary atmosphere modeling, climate simulations, and general radiative transfer applications. Written in CUDA C++ and optimized for NVIDIA GPUs, HELIOS-K achieves throughput of $\sim 10^7$–$10^8$ lines per second, enabling it to condense exceptionally large spectroscopic line lists into usable opacity tables or correlated-k-distribution tables for forward modeling, retrievals, and high-dimensional circulation models. The code supports a broad array of line-list sources (e.g., ExoMol, HITRAN, HITEMP, NIST, Kurucz, VALD3), implements state-of-the-art Voigt-profile algorithms, and delivers both high-resolution line-by-line opacity outputs and compressed k-tables. The software, now at version 2.0, integrates a self-tuning GPU kernel and asynchronous I/O pipeline, and is coupled to an open-access opacity database, DACE, that provides ready-to-use cross sections for hundreds of species across extensive $(T,P,\nu)$ grids (Grimm et al., 2015, Grimm et al., 2021, Malik et al., 2016, Simonetti et al., 2021, Deitrick et al., 2022).

1. Physical and Computational Context

Opacity calculations are central to exoplanet atmosphere modeling, including 1D radiative–convective equilibrium, retrieval pipelines, and 3D general circulation models (GCMs). Temperatures above $\sim500$ K activate billions of spectral lines per molecule (e.g., H$_2$O transitions $\sim10^9$ at $T\approx1000$ K), and these transitions must be resolved across fine wavenumber and $(T,P)$ grids. Classic CPU-based approaches cannot accommodate $\gtrsim 10^8$ lines practically, necessitating GPU acceleration and algorithmic optimizations (Grimm et al., 2015, Grimm et al., 2021). HELIOS-K was developed to deliver opacity computations that would otherwise demand hours or days, within seconds to minutes on modern GPU hardware.

Typical workflows consist of: (i) preprocessing line-list data (e.g., conversion, error checking), (ii) computation of line-by-line cross sections and blending with continuum opacities, (iii) binning and transformation to k-distributions, and (iv) integration with radiative-transfer solvers (e.g., HELIOS, EOS, or GCMs such as THOR+HELIOS) (Grimm et al., 2015, Deitrick et al., 2022).

2. Inputs, Supported Databases, and Preprocessing

HELIOS-K accepts molecular and atomic line lists in standard formats (e.g., ExoMol .states/.trans, HITRAN/HITEMP .par, NIST, Kurucz GFNEW, VALD3). Supported spectroscopic quantities per transition include:

• Line-center wavenumber $\nu_0$
• Einstein $10^8$0 coefficient
• Lower-state energy $10^8$1 or energy difference $10^8$2
• Statistical weight $10^8$3
• Air/self broadening half-widths, temperature exponents, and pressure shifts
• Reference line strength $10^8$4 (commonly at $10^8$5 K)

Partition functions are read from accompanying files, with database-specific corrections implemented (e.g., isotope fractions, broadening flags, missing $10^8$6).

Preprocessing via Python scripts includes downloading, parsing, applying database corrections, computing $10^8$7 from spectroscopic data, and writing the results in a uniform binary block format for GPU consumption (Grimm et al., 2021). For continuum absorption (e.g., CIA H$10^8$8–H$10^8$9, H$(T,P,\nu)$0–He, or H$(T,P,\nu)$1O/CO$(T,P,\nu)$2 continua), tabular data from sources such as MT-CKD or Gruszka-Borysow-Baranov are merged as needed (Simonetti et al., 2021, Deitrick et al., 2022).

3. Voigt-Profile Computation and Line-wing Treatment

Line shapes are modeled as Voigt profiles, the convolution of Gaussian (thermal Doppler) and Lorentzian (pressure/natural) broadening,

$(T,P,\nu)$3

where $(T,P,\nu)$4 and $(T,P,\nu)$5 (Grimm et al., 2015, Grimm et al., 2021). HELIOS-K employs a hybrid of Algorithm 916 (Zaghloul & Ali 2012) for low $(T,P,\nu)$6, and Gauss–Hermite quadrature for large $(T,P,\nu)$7 regimes, as described in (Grimm et al., 2015). This approach balances speed and numerical precision over the entire line profile.

A truncation or "cutting length" is applied to Voigt wings—typically |$(T,P,\nu)$8| ≤ 25–100 cm$(T,P,\nu)$9 for molecules (the default at DACE is 100 cm$\sim500$0; for atoms, no cut is used)—due to uncertainty in far-wing broadening physics (Grimm et al., 2021). The choice of cutoff can affect radiative fluxes by up to $\sim500$1; this is currently an irreducible error source independent of computational methodology (Grimm et al., 2015).

4. k-Distribution and Correlated-k Methods

To reduce computational cost, HELIOS-K supports the k-distribution method, where line-by-line opacity $\sim500$2 in a spectral interval is reordered into a cumulative distribution: $\sim500$3 This enables radiative transfer integrals to be recast from $\sim500$4-space to $\sim500$5-space: $\sim500$6 HELIOS-K accomplishes this by binning $\sim500$7 within each spectral interval, sorting to obtain $\sim500$8, and resampling the result using Chebyshev polynomials or Gaussian quadrature points for efficient storage and rapid radiative-transfer evaluation (Grimm et al., 2015, Malik et al., 2016, Deitrick et al., 2022). The correlated-k approximation is invoked to enable bin-wise rank-ordering to be used across different atmospheric layers or species, yielding flux errors typically $\sim500$91–10% when bins are sufficiently finely resolved (Grimm et al., 2015).

5. GPU Implementation and Performance

HELIOS-K code is structured for efficient CUDA execution. Key design features:

• Each CUDA thread handles one $_2$0 point; lines are accumulated via registers and shared memory.
• Line parameters are bulk-loaded into device memory; constants are placed in CUDA's constant cache.
• Sorting (e.g., in k-distribution construction) leverages NVIDIA's Thrust library.
• Overlapping host/device data transfers are managed through CUDA streams for maximum throughput.
• The GPU plane is tiled to optimize memory locality and minimize warp divergence by region decomposition in $_2$1 space (Grimm et al., 2021).

Benchmarks on NVIDIA Tesla cards report the following:

• $_2$2–$_2$3 lines: $_2$41 second for line-by-line opacity function with $_2$5, $_2$6 (Grimm et al., 2015).
• ExoMol BT2 H$_2$7O, $_2$8–$_2$9 cm$\sim10^9$0, $\sim10^9$1 lines: $\sim10^9$2 s at $\sim10^9$3 cm$\sim10^9$4, $\sim10^9$5 s at $\sim10^9$6 cm$\sim10^9$7 (Grimm et al., 2021).
• Effective GPU throughput is up to $\sim10^9$8 lines s$\sim10^9$9, yielding an $T\approx1000$0 acceleration over traditional Voigt integrators (Grimm et al., 2021).

6. Applications and Integration in Atmospheric Modeling

HELIOS-K outputs high-resolution $T\approx1000$1 and/or precomputed k-distribution tables. These are directly ingestible by radiative-transfer codes such as HELIOS (for self-consistent 1D and 3D modeling) (Malik et al., 2016, Deitrick et al., 2022) and EOS (Simonetti et al., 2021), as well as GCMs such as THOR+HELIOS that require multi-wavelength, multi-species, and cloud/haze radiative transfer (Deitrick et al., 2022). HELIOS-K is callable from Python and Fortran through APIs, and can interface with radiative transfer solvers using "perfectly correlated" or "random overlap" k-coefficient treatments.

Precomputed opacity grids are made available via the DACE portal (https://dace.unige.ch/opacityDatabase), supporting direct download or server-side $T\approx1000$2 extraction and visualization, with file metadata including species, isotopologue, resolution, and line-wing cut. This facilitates reproducible, rapid prototyping and benchmarking for exoplanet models (Grimm et al., 2021).

7. Limitations, Uncertainties, and Future Enhancements

The principal source of uncertainty in HELIOS-K lies in the choice of line-wing cutoff, as Lorentzian/Voigt profiles systematically overestimate far-wing absorption under high pressure. Quantum-mechanical broadening models remain incomplete; thus, errors in computed fluxes and spectra are at the $T\approx1000$3 level and propagate into atmospheric structure and retrievals (Grimm et al., 2015). Computational performance is now dominated by I/O and database parsing, not GPU-side evaluation.

HELIOS-K 2.0 incorporates modular architecture with pipeline parallelism, self-tuning kernel sizing, and advanced file I/O. Planned future developments include support for non-Voigt line profiles (e.g., Allard or Fano), scaling to multi-GPU and cluster environments, and versioned, on-the-fly k-distribution construction (Grimm et al., 2021).

---

References:

(Grimm et al., 2015, Malik et al., 2016, Grimm et al., 2021, Simonetti et al., 2021, Deitrick et al., 2022)