Float32 and Float64

All RRTMGP.jl kernels are generic in the working precision: the float type chosen at construction (through RRTMGPGridParams, solve, or solve_gray) flows through the lookup tables, the optics, and the RTE solve. Float64 is the reference; Float32 halves the memory traffic of the bandwidth-bound kernels and is the precision at which the CliMA atmosphere runs radiation on GPUs.

Accuracy achieved at Float32

A consistency test (test/float32_consistency.jl) solves the same gray, clear-sky, and cloudy problems at both precisions and measures the end-to-end difference: input rounding, gas and cloud optics, and the RTE solve together. The measured maximum broadband flux differences are:

Longwave: about 2–5 × 10⁻⁴ W/m², 25–90 times smaller than before the single-precision measures described below and reaching the interpolation-noise floor.
Shortwave: about 1–2 × 10⁻² W/m² in clear skies and about 5 × 10⁻² W/m² with clouds. These were verified to be per-g-point interpolation and coefficient noise rather than accumulation error: repeating the runs with full Float64 broadband accumulation left them unchanged, so compensated summation was unnecessary.

For scale, these differences are orders of magnitude below both typical flux magnitudes (hundreds of W/m²) and the accuracy of the correlated-$k$ approximation itself; the Fortran reference (rte-rrtmgp) accepts an absolute flux error of 0.35 W/m² in its own validation. CI enforces ratcheting thresholds (gray 10⁻³ W/m² and 10⁻⁸ K/s for the heating rate; longwave 10⁻³ W/m²; shortwave 3 × 10⁻² clear and 1.2 × 10⁻¹ W/m² cloudy) that are tightened as numerics improvements land. The comparisons against the rte-rrtmgp reference results also run at both precisions.

How: the single-precision numerics

The Float32 accuracy rests on evaluating exact reformulations instead of cancellation-prone ones, and on switching to series expansions where cancellation is unavoidable. The measures were, roughly in the order radiation flows:

Gas optics. The interpolation fraction in the binary species parameter $η$ is computed relative to its clamped table cell, so it stays continuous at $η = 1$, a value to which Float32 rounds more often than Float64.
Longwave, no scattering. The linear-in-$τ$ source factor $(1 - t)/\tau - t$ (Clough et al. [16], Eq. 13) loses $\sim\!\mathrm{eps}/\tau^2$ to cancellation for thin layers; it switches to its third-order series at $\tau = \mathrm{eps}^{1/4}$, where the two error curves cross.
Two-stream coefficients. The diffusion eigenvalue $k = \sqrt{(\gamma_1-\gamma_2)(\gamma_1+\gamma_2)}$ uses exact $\gamma_1 - \gamma_2$ identities rather than subtracting near-equal numbers, with $k^2$ floored at $\sqrt{\mathrm{eps}}$; thin-layer factors $1 - e^{-k\tau}$ are evaluated with expm1.
Longwave two-stream source. The Toon et al. [17] source terms are factored so that no term divides by $\tau$ and no small difference of near-equal quantities appears (exact $1 \mp R_{\mathrm{dif}} - T_{\mathrm{dif}}$ factorizations).
Shortwave direct beam. The direct reflectance and transmittance of Meador and Weaver [13] (Eqs. 14–15) have a removable singularity at $k\mu_0 = 1$; $k\mu_0$ is nudged off resonance symmetrically within a $\sqrt{\mathrm{eps}}$ window, keeping numerator and denominator mutually consistent. An energy-conservation clamp then bounds the direct reflectance and transmittance by the energy the unscattered beam leaves behind, following Ukkonen and Hogan [21]; and the direct-beam profile itself is built by accumulating optical depth downward rather than by repeated multiplication.
Delta scaling. The $f = g^2$ similarity scaling of cloud and aerosol optics (Joseph et al. [20]) is evaluated in exact, non-cancelling forms (for example $g' = g/(1+g)$ instead of $(g - g^2)/(1 - g^2)$).

The guard constants involved (k_min, τ_thresh, resonance_window, μ₀_min) live in the RRTMGP.Numerics module, where their values are derived; all scale with eps(FT), so the kernels behave consistently at either precision, and the same logic would extend to other float types.