gaussregular 0.1.3

A lightweight, high-performance Python library for converting ECMWF reduced Gaussian grids (N-grids and O-grids) to regular Gaussian grids. Designed for efficient in-memory processing with zero intermediate file I/O.

Overview

GaussRegular specializes in converting ECMWF reduced Gaussian grids to full (regular) Gaussian grids:

• N-grids: Classical reduced Gaussian grids (e.g., N320, N640)
• O-grids: Octahedral reduced Gaussian grids (e.g., O320, O640, O1280)

Both grid types carry a pl array (points per latitude row) and use identical row-wise interpolation logic, matching CDO's setgridtype,regular and setgridtype,regularnn behavior.

Features

• Dual grid support: Handles both classical N-grids (e.g., N320, N640) and octahedral O-grids (e.g., O320, O640, O1280)
• Two interpolation modes:
  • Bilinear (default) — matching CDO setgridtype,regular
  • Nearest-neighbor — matching CDO setgridtype,regularnn
• C-accelerated: Core interpolation in compiled C, ~10-100x faster than pure NumPy
• In-memory processing: Read GRIB2 data → convert → analyze → output, no intermediate files
• Configurable precision: precision="auto" preserves float32/float64 input precision; single and double force the output precision
• xarray/cfgrib integration: Direct support for GRIB2 data
• Batch processing: Efficient multi-dimensional array handling (time, level, etc.)

Installation

pip install gaussregular

With xarray support:

pip install "gaussregular[xarray]"

Quick Start

Case 1: Direct Numpy Array Conversion

If you have a NumPy array with known grid metadata:

import numpy as np
import gaussregular as gr

# Your reduced Gaussian data (1D array, one value per grid point)
values = np.random.rand(1040640)  # Example data

# Gaussian row length array (from GRIB metadata, e.g., GRIB_pl)
pl = np.array([20, 27, 36, 40, 45, ..., 360, 360], dtype=np.int32)  # 161 latitude rows in this example

# Create regularizer
regularizer = gr.GaussRegularizer(method="linear")

# Convert to regular Gaussian (2D array)
regular_data, lon = regularizer.regularize_values(
    values=values,
    pl=pl,
    missval=np.nan
)

print(f"Output shape: {regular_data.shape}")  # e.g. (161, 320) in this example
print(f"Longitude points: {lon.size}")

Case 2: xarray DataArray (from cfgrib)

If you're reading GRIB2 files with cfgrib:

import xarray as xr
import gaussregular as gr

# Open GRIB2 file with cfgrib backend
ds = xr.open_dataset("era5_sample.grib", engine="cfgrib")
da = ds["temperature"]  # xarray.DataArray

# Create regularizer
regularizer = gr.GaussRegularizer(method="linear", cache=True)

# Convert directly
regular_da = regularizer.regularize_xarray(da)

print(f"Regular grid shape: {regular_da.shape}")
print(f"Coordinates preserved: {list(regular_da.coords.keys())}")

The library automatically extracts GRIB_pl, grid type, and other metadata from the DataArray attributes.

Case 3: Multi-dimensional Data (Time series, Levels)

Batch processing with leading dimensions:

import xarray as xr
import gaussregular as gr

# Multi-dimensional xarray: (time, level, values)
ds = xr.open_dataset("era5_multi_level.grib", engine="cfgrib")
da = ds["temperature"]  # Shape: (24, 137, 1040640) in this example

regularizer = gr.GaussRegularizer(method="linear", cache=True)
regular_da = regularizer.regularize_xarray(da)

print(f"Output shape: {regular_da.shape}")  # e.g. (24, 137, 161, 320) in this example

The regularizer automatically handles reshaping and maintains coordinate dimensions.

Case 4: xarray Dataset (Multiple Variables)

You can pass a full xr.Dataset directly. All convertible reduced-Gaussian variables are regularized.

import xarray as xr
import gaussregular as gr

ds = xr.open_dataset("era5_multi_vars.grib", engine="cfgrib")

regularizer = gr.GaussRegularizer(method="linear", cache=True)
regular_ds = regularizer.regularize_dataset(ds)

# Or use auto dispatch:
# regular_ds = regularizer.regularize(ds)

print(regular_ds)

Case 5: Regional Sub-area Data

For data covering only a portion of the globe (e.g., a regional cut from a model run), specify the longitude bounds and grid number:

import numpy as np
import gaussregular as gr

# Regional reduced Gaussian data
values = np.random.rand(342080)  # Subset of N320 (e.g., Europe region)
pl = np.array([...])  # Row lengths for the region

regularizer = gr.GaussRegularizer(method="linear")

# Specify regional bounds and grid number for proper output sizing
regular_data, lon = regularizer.regularize_values(
    values=values,
    pl=pl,
    missval=np.nan,
    grid_number=320,  # N320 grid
    xfirst=0.0,       # Start longitude (degrees east)
    xlast=40.0,       # End longitude (degrees east)
)

print(f"Regional grid shape: {regular_data.shape}")
print(f"Longitude range: {lon[0]:.2f}–{lon[-1]:.2f}°E")

Parameters for regional data:

• grid_number (int): Gaussian truncation number (e.g., 320 for N320). Required to distinguish global vs. regional grids.
• xfirst (float, default=0.0): Longitude of the first data point (degrees east, range 0–360).
• xlast (float, default=359.9999): Longitude of the last data point (degrees east).

The library automatically infers the output longitude count based on the regional bounds.

Advanced: Fast Mode (Default)

By default fast=True is used for best performance. The linear interpolation kernel handles missing values (NaN / sentinel) correctly even in fast mode — if both interpolation neighbors are missing the output is missing, otherwise the valid neighbor is used. Set fast=False only if you need per-row missing-value scanning for strict correctness on exotic sentinel values:

import gaussregular as gr

engine = gr.GaussRegularizer(method="linear", cache=True)

# Default: fast=True — recommended for all ERA5/GRIB data
regular_ds = engine.regularize_dataset(ds)

# Safe mode: per-row missing-value scan (slower, ~1.8x)
regular_ds_safe = engine.regularize_dataset(ds, fast=False)

Advanced: Precision

By default precision="auto" is used. Float32 inputs use the single-precision C path and return float32 outputs; float64 inputs use the double-precision C path and return float64 outputs. Use precision="single" or precision="double" to force a specific output precision:

import gaussregular as gr

engine = gr.GaussRegularizer(method="linear", cache=True)

regular_auto = engine.regularize_xarray(da)  # follows da.dtype for float32/float64
regular_f32 = engine.regularize_xarray(da, precision="single")
regular_f64 = engine.regularize_xarray(da, precision="double")

API Reference

GaussRegularizer

Main converter class.

Parameters:

• method (str, default="linear"): Interpolation method, either "linear" or "nearest"
• precision (str, default="auto"): Precision policy, one of "auto", "single", or "double". Auto preserves float32/float64 input precision
• grid_number (int, optional): Default Gaussian truncation number N (e.g., 320 for N320/O320). Used when GRIB_N is missing for xarray inputs and as a fallback when regularize_values is called without a per-call grid_number. When neither metadata nor this attribute provide N, a heuristic based on the equatorial row length is used
• cache (bool, default=False): Enable xarray metadata plan caching for repeated calls on same-structure data
• max_plan_cache (int, default=32): Maximum number of cached conversion plans

Methods:

regularize(data, nlon=None, grid_type_hint=None, method=None)

• Auto-detects input type (xr.Dataset, xr.DataArray, or NumPy) and calls appropriate method
• Returns regularized data in same type as input

regularize_values(values, pl, missval, nlon=None, method=None, fast=True)

• values: 1D NumPy array (len(values) == sum(pl))
• pl: 1D NumPy array of row lengths (from GRIB_pl)
• missval: Missing-value sentinel (float). Points equal to this value (or NaN, if missval is NaN) are treated as missing and handled by the interpolation kernel
• nlon (optional): Number of columns in output regular grid. Auto-inferred if not provided
• method (optional): Per-call method override: "linear" or "nearest"
• fast (default=True): Per-row missing-value scan toggle. True (recommended) lets the interpolation kernel handle NaN/sentinel natively. Set False for explicit per-row scanning
• Returns: (out, lon) where out is NumPy array and lon is longitude vector

regularize_xarray(dataarray, nlon=None, grid_type_hint=None, method=None, fast=True)

• dataarray: xarray.DataArray with GRIB metadata in .attrs
• Requires: GRIB_pl attribute. If GRIB_missingValue is present it is used as missval; otherwise NaN values are treated as missing
• method (optional): Per-call method override: "linear" or "nearest"
• fast (default=True): Per-row missing-value scan toggle
• Returns: xarray.DataArray with regularized data and preserved coordinates

regularize_dataset(dataset, nlon=None, grid_type_hint=None, method=None, fast=True)

• dataset: xarray.Dataset with one or more reduced Gaussian variables
• Converts each convertible variable and keeps non-convertible variables unchanged
• method (optional): Per-call method override: "linear" or "nearest"
• fast (default=True): Per-row missing-value scan toggle
• Returns: xarray.Dataset

clear_cache()

• Clears internal xarray plan cache

Module-level Functions

regularize_values(values, pl, missval, nlon=None, method="linear", fast=True)

• Standalone function using default regularizer configuration

regularize_xarray(dataarray, grid_type_hint=None, method="linear", fast=True)

• Standalone function using default regularizer configuration

regularize_dataset(dataset, grid_type_hint=None, method="linear", fast=True)

• Standalone function using default regularizer configuration

Grid Specifics

N-grids (Classical Gaussian)

• Regular spacing in longitude
• Example: N320 → 4×N = 1280 columns at the equator in the regular grid
• Total reduced-grid points: 4×N×(2N+1)

O-grids (Octahedral)

• Refinement at equator for better accuracy
• Example: O320, O640, O1280
• Formula: nlon = 4×N + 16 for O-grid

The library automatically detects grid type from GRIB metadata. You can also override with grid_type_hint parameter.

Supported Grid Types

GaussRegular recognizes many GRIB metadata variants:

• reduced_gg
• reduced-gg
• reduced gaussian
• octahedral_gg
• octahedral_reduced_gg
• And several aliases

See GRID_TYPE_FULL_NAMES in the API for the full list.

Performance Tuning

Interpolation Method

• method="linear" (default): Bilinear interpolation. Slightly slower but more accurate.
• method="nearest": Nearest-neighbor. Faster, less smooth.

Missing-Value Detection

• fast=True (default): Recommended mode. The linear interpolation kernel correctly propagates missing values at boundaries (both neighbors missing → output missing). Provides maximum throughput.
• fast=False: Per-row missing-value scan before interpolation. Use only if you have exotic non-NaN sentinel values that require explicit detection. Typical slowdown: ~1.8x.

Caching

• Enable cache=True when processing multiple DataArrays with identical structure to reuse parsed metadata.

Typical throughput:

• On a typical ERA5 model-level example (4 time steps × 137 levels, N96/O96 grids), a full-field linear interpolation runs in ~0.24 s on a modern desktop CPU

Limitations

1. Missing values are propagated by the interpolation kernel — NaN/sentinel inputs produce NaN outputs at affected grid points by default (fast=True). For strict per-row scanning use fast=False
2. Requires GRIB metadata — pl array must be provided or in DataArray attrs
3. Supports only Gaussian grids — other map projections not supported
4. Precision: Outputs double-precision computation; results cast back to input dtype
5. No projection support — only Gaussian grid conversions

Testing & Validation

GaussRegular has been validated against CDO setgridtype,regular and setgridtype,regularnn with double-precision agreement to machine epsilon.

Test with real ERA5 data:

python examples/run_era5_demo.py

Requirements

• Python: 3.9+
• NumPy: ≥1.23
• Optional: xarray ≥2023.1, cfgrib ≥0.9.10

Citation

If you use GaussRegular in research, please cite:

@software{gaussregular,
  author = {Qianye Su},
  title = {GaussRegular: ECMWF Reduced Gaussian to Regular Grid Converter},
  year = {2026},
  url = {https://github.com/QianyeSu/GaussRegular}
}

Troubleshooting

"Missing GRIB_pl in DataArray attrs"

• Ensure you opened the file with engine="cfgrib"
• Check that cfgrib correctly parsed the GRIB2 file

"Input does not look like reduced Gaussian grid"

• Verify your file contains a reduced Gaussian grid (not full Gaussian, lat-lon, or other projection)
• Pass grid_type_hint if metadata is nonstandard

"Results look wrong near missing values"

• Ensure you are not using fast=True when missing values are present
• Verify that missing values are either NaN (and missval is left as default) or match the GRIB_missingValue/missval you pass
• If necessary, preprocess your data（例如插值或填充）以减少大块缺测区域对结果的影响

Performance is slower than expected

• Use method="linear" unless you specifically want nearest-neighbour behavior
• Use caching: regularizer.cache=True for repeated calls
• Profile with larger batch operations instead of single conversions

Contributing

Contributions welcome! Please:

1. Read the source code and existing patterns
2. Write tests for new features
3. Ensure CI passes before submitting PR