geocat-rs 0.1.1

Rust-accelerated drop-in replacements for geocat-comp computational functions

geocat-rs

Rust-accelerated drop-in replacements for GeoCAT-comp computational functions.

GeoCAT-comp is NCAR's Python replacement for NCL. geocat-rs accelerates the compute-heavy functions using Rust + rayon parallelism, while leaving the xarray/scipy wrapper functions alone (they're already fast).

Performance

Real-world benchmarks on CESM-like 3D atmospheric fields. Every result verified identical to geocat.comp.

The #1 workflow: interp_hybrid_to_pressure

The function CESM researchers use most — converting hybrid-sigma vertical coordinates to standard pressure levels (replacing NCL's vinth2p).

Grid size	geocat.comp	geocat-rs	Speedup	Max diff
32 lev x 192 lat x 288 lon (1.8M pts)	161 ms	21 ms	7.5x	0.0
32 lev x 500 lat x 500 lon (8M pts)	724 ms	95 ms	7.6x	0.0

Tested with 14 drop-in compatibility tests: standard/full CESM grids, custom levels, multiple timesteps, mountain pressure extremes, 4 random seeds. All pass.

Meteorology functions on 3D fields (12.5M elements)

Function	geocat.comp	geocat-rs	Speedup
Dewpoint temperature	189 ms	12 ms	16x
Relative humidity (table lookup)	353 ms	17 ms	21x
Relative humidity over water	223 ms	16 ms	14x
Saturation vapor pressure	6.1 ms	4.5 ms	1.4x

WGS84 gradient functions (250K grid points)

Function	geocat.comp	geocat-rs	Speedup
rad_lat_wgs84 (degree-48 polynomial)	67 ms	1.0 ms	66x
arc_lat_wgs84 (degree-49 polynomial)	69 ms	1.0 ms	70x

Overall across all accelerated functions

	geocat.comp	geocat-rs
Total time	908 ms	53 ms
Overall speedup		17x

Accuracy

All functions verified against geocat.comp to machine precision:

Test suite	Tests	Result
interp_hybrid_to_pressure compatibility	14	All pass (max diff: 0.0)
Meteorology functions (scalar + array)	20	All pass (rtol < 1e-12)
WGS84 gradient functions	6	All pass (rtol < 1e-10)
Delta pressure	1	Pass
Total	44	All pass

pytest tests/ -v
python tests/bench_interp_hybrid.py
python tests/bench_workflow.py

Installation

pip install geocat-rs

Requires metrust for interp_hybrid_to_pressure (uses its Rust interpolation engine).

Usage

interp_hybrid_to_pressure (drop-in replacement)

# Before
from geocat.comp import interp_hybrid_to_pressure

# After (same API, 7.5x faster)
from geocat_rs.interp_hybrid import interp_hybrid_to_pressure

# Usage is identical
output = interp_hybrid_to_pressure(
    data, ps, hyam, hybm,
    new_levels=new_levels,
    lev_dim='lev',
)

Meteorology (array functions via Rust + rayon)

from geocat_rs._geocat_rs import meteorology as met

# Scalar
td = met.dewtemp_scalar(300.0, 50.0)  # T(K), RH(%) -> Td(K)
hi = met.heat_index_scalar(95.0, 50.0, False)  # T(F), RH(%) -> HI(F)
rh = met.relhum_scalar(300.0, 0.01, 101325.0)  # T(K), w(kg/kg), P(Pa) -> RH(%)

# Array (rayon-parallel)
import numpy as np
td_array = met.dewtemp_array(temp_k, rh_pct)
rh_array = met.relhum_array(temp_k, mixr, pressure)
svp_array = met.saturation_vapor_pressure_array(temp_f)

WGS84 gradient (66-70x faster)

from geocat_rs._geocat_rs import gradient as grad

radius = grad.rad_lat_wgs84_array(lat_grid.ravel())  # meters
arc = grad.arc_lat_wgs84_array(lat_grid.ravel())      # meters from equator

What's accelerated

Function	Source	Speedup	Notes
interp_hybrid_to_pressure	Rust column interpolator + rayon	7.5x	#1 most-used geocat function
dewtemp	Rust + rayon	16x	Dutton formula
relhum (table)	Rust + rayon	21x	NCL lookup table
relhum_ice / relhum_water	Rust + rayon	14x	Alduchov/Murray
heat_index	Rust + rayon	--	NWS Rothfusz regression
saturation_vapor_pressure	Rust + rayon	1.4x	Tetens/FAO-56
saturation_vapor_pressure_slope	Rust + rayon	--	FAO-56 Eq. 13
psychrometric_constant	Rust + rayon	--	FAO-56 Eq. 8
max_daylight	Rust + rayon	--	FAO-56 solar declination
rad_lat_wgs84	Rust + rayon	66x	Degree-48 polynomial
arc_lat_wgs84	Rust + rayon	70x	Degree-49 polynomial
arc_lon_wgs84	Rust + rayon	--	Radius * cos(lat) * lon
temp_extrapolate	Rust + rayon	--	ECMWF Eq. 16
geo_height_extrapolate	Rust + rayon	--	ECMWF Eq. 15
delta_pressure_1d	Rust	--	Simmons & Burridge

What's NOT accelerated (and why)

Function	Why not
climatology_average, calendar_average, month_to_season	xarray groupby — already optimized
eofunc_eofs, eofunc_pcs	Wraps eofs library (LAPACK)
fourier_filter, fourier_*_pass	numpy FFT (FFTW backend)
decomposition, recomposition	scipy spherical harmonics
interp_multidim, interp_sigma_to_hybrid	Wraps metpy (already Rust via metrust)

License

Apache-2.0 (same as GeoCAT-comp)