Xarray Netcdf

xarray + NetCDF — Labeled N-D Arrays, Lazy I/O & CF Conventions

xarray brings the power of pandas-style labeled indexing to N-dimensional arrays. Combined with Dask for lazy out-of-core computation and Zarr for cloud-native storage, it forms the backbone of modern climate, oceanography, and remote-sensing workflows.

When to Use This Skill

Use this skill when you need to:

• Open, inspect, and process NetCDF4, HDF5, or GRIB files without loading all data to RAM.
• Work with coordinates (lat, lon, time, level) attached to array dimensions.
• Run climatological analyses: seasonal means, anomalies, rolling statistics.
• Rechunk large datasets for parallel processing with Dask workers.
• Convert legacy .nc files to chunked Zarr stores for fast cloud access.
• Attach CF-compliant metadata (units, standard_name, calendar) to output files.
• Merge, align, or interpolate multiple datasets with different grids.

Do not use this skill for purely tabular data (use pandas) or for small arrays that fit comfortably in memory without any need for coordinate labels (use NumPy).

Background & Key Concepts

xarray Data Structures

A DataArray records dimensions (e.g., "time", "lat", "lon"), coordinates (1-D arrays of tick values for each dim), and attributes (metadata dict).

NetCDF & CF Conventions

NetCDF-4 is a hierarchical binary format built on HDF5. The CF (Climate and Forecast) Metadata Conventions standardise:

• Coordinate names (latitude, longitude, time, pressure)
• units string format ("degrees_east", "K", "days since 1900-01-01")
• standard_name vocabulary (e.g., "air_temperature")
• Grid mapping attributes for projected grids

xarray decodes CF metadata automatically when opening NetCDF files.

Lazy Loading with Dask

When chunks={} or chunks="auto" is passed to xr.open_dataset, each variable becomes a dask.array — a graph of deferred operations. Nothing is loaded until .compute() or .load() is called, or data is written to disk.

Choosing good chunk sizes:

• Target ~100 MB per chunk.
• Chunk along the dimensions you will iterate over (usually time).
• Avoid tiny chunks (overhead) and huge chunks (memory spikes).

Zarr

Zarr stores each chunk as a separate compressed file (in a directory or cloud bucket). This enables concurrent reads by many Dask workers, far outperforming NetCDF for parallel workloads. The xarray.Dataset.to_zarr() method handles the conversion.

Environment Setup

Install dependencies

pip install "xarray>=2023.6" "dask[distributed]>=2023.6" "zarr>=2.15" \
            "netCDF4>=1.6" "matplotlib>=3.7" "numpy>=1.24" \
            bottleneck h5py scipy cftime

Optional: start a local Dask cluster

from dask.distributed import Client

client = Client(n_workers=4, threads_per_worker=2, memory_limit="4GB")
print(client.dashboard_link)   # open in browser for task graph visualization

Environment variables

# Path to your local data archive
export NC_DATA_DIR="/data/netcdf"
# Optional: fsspec token for cloud access (GCS, S3)
export FSSPEC_S3_KEY="<paste-your-key>"
export FSSPEC_S3_SECRET="<paste-your-secret>"

import os

data_dir = os.getenv("NC_DATA_DIR", "./data")
s3_key    = os.getenv("FSSPEC_S3_KEY", "")
s3_secret = os.getenv("FSSPEC_S3_SECRET", "")

Verify installation

import xarray as xr
import dask
import zarr
import netCDF4

print("xarray :", xr.__version__)
print("dask   :", dask.__version__)
print("zarr   :", zarr.__version__)
print("netCDF4:", netCDF4.__version__)

Core Workflow

Step 1 — Create, Inspect, and Index a DataArray

import numpy as np
import xarray as xr
import pandas as pd

# ── Build a synthetic 3-D temperature dataset ────────────────────────────────
rng = np.random.default_rng(42)

times = pd.date_range("2020-01-01", periods=365, freq="D")
lats  = np.linspace(-90, 90,  73)    # 2.5° grid
lons  = np.linspace(  0, 357.5, 144) # 2.5° grid

# Shape: (time=365, lat=73, lon=144)
data = (
    280
    + 30 * np.cos(np.deg2rad(lats))[:, None]         # latitude gradient
    + 10 * np.sin(2 * np.pi * np.arange(365) / 365)[None, :, None]  # seasonal cycle
    + rng.standard_normal((73, 365, 144))              # noise
)
# Re-order to (time, lat, lon)
data = data.transpose(1, 0, 2)

temp = xr.DataArray(
    data,
    dims=["time", "lat", "lon"],
    coords={
        "time": times,
        "lat":  xr.Variable("lat",  lats,  attrs={"units": "degrees_north",
                                                    "standard_name": "latitude"}),
        "lon":  xr.Variable("lon",  lons,  attrs={"units": "degrees_east",
                                                    "standard_name": "longitude"}),
    },
    attrs={
        "long_name": "Air Temperature",
        "units": "K",
        "standard_name": "air_temperature",
    },
    name="tas",
)

print(temp)
print("\nDimensions:", dict(temp.dims))
print("Coordinates:", list(temp.coords))

# ── Label-based indexing ──────────────────────────────────────────────────────
# Select a single point by coordinate value
london = temp.sel(lat=51.5, lon=0.0, method="nearest")
print("\nLondon time series shape:", london.shape)

# Select a spatial slice
europe = temp.sel(lat=slice(35, 70), lon=slice(345, 360))
print("Europe slice shape:", europe.shape)

# Select by index position
first_week = temp.isel(time=slice(0, 7))
print("First week shape:", first_week.shape)

# Fancy: nearest-neighbor interpolation to irregular stations
station_lats = xr.DataArray([48.9, 51.5, 55.8], dims="station")
station_lons = xr.DataArray([2.3,  0.0, 37.6],  dims="station")
stations = temp.sel(lat=station_lats, lon=station_lons, method="nearest")
print("Station extraction shape:", stations.shape)

Step 2 — NetCDF I/O and CF Metadata

import xarray as xr
import numpy as np
import pandas as pd
import os

DATA_DIR = os.getenv("NC_DATA_DIR", "./data")
os.makedirs(DATA_DIR, exist_ok=True)
NC_PATH = os.path.join(DATA_DIR, "temperature.nc")

# ── Write a CF-compliant NetCDF file ─────────────────────────────────────────
times = pd.date_range("2000-01-01", periods=12, freq="ME")
lats  = np.linspace(-90, 90,  37)
lons  = np.linspace(  0, 355, 72)
data  = 280 + 20 * np.cos(np.deg2rad(lats))[:, None] * np.ones((37, 12, 72))
data  = data.transpose(1, 0, 2)

ds = xr.Dataset(
    {
        "tas": xr.DataArray(
            data.astype("float32"),
            dims=["time", "lat", "lon"],
            attrs={
                "long_name":     "Near-Surface Air Temperature",
                "standard_name": "air_temperature",
                "units":         "K",
                "cell_methods":  "time: mean",
            },
        )
    },
    coords={
        "time": xr.Variable("time", times,
                             attrs={"long_name": "time",
                                    "axis": "T"}),
        "lat": xr.Variable("lat", lats,
                            attrs={"long_name": "latitude",
                                   "standard_name": "latitude",
                                   "units": "degrees_north",
                                   "axis": "Y"}),
        "lon": xr.Variable("lon", lons,
                            attrs={"long_name": "longitude",
                                   "standard_name": "longitude",
                                   "units": "degrees_east",
                                   "axis": "X"}),
    },
    attrs={
        "Conventions":   "CF-1.10",
        "title":         "Synthetic monthly temperature",
        "institution":   "Demo",
        "source":        "Generated by xarray-netcdf SKILL",
        "history":       "Created 2026-03-17",
        "references":    "",
        "comment":       "Illustrative dataset only.",
    },
)

# Encode time and apply compression
encoding = {
    "tas":  {"zlib": True, "complevel": 4, "dtype": "float32",
             "_FillValue": 1e20},
    "time": {"dtype": "float64",
              "units": "days since 2000-01-01",
              "calendar": "proleptic_gregorian"},
}
ds.to_netcdf(NC_PATH, encoding=encoding)
print(f"Saved to {NC_PATH}")

# ── Re-open and inspect ───────────────────────────────────────────────────────
ds2 = xr.open_dataset(NC_PATH)
print(ds2)
print("\nVariable attrs:", ds2["tas"].attrs)
print("Time dtype:", ds2["time"].dtype)
ds2.close()

# ── Open multiple files with glob (MFDataset) ─────────────────────────────────
# xr.open_mfdataset concatenates files along a shared dimension automatically
# ds_all = xr.open_mfdataset(
#     os.path.join(DATA_DIR, "temperature_*.nc"),
#     combine="by_coords",
#     parallel=True,     # uses dask
# )

Step 3 — Lazy Dask Operations and Rechunking

import xarray as xr
import numpy as np
import pandas as pd
import dask
import os

DATA_DIR = os.getenv("NC_DATA_DIR", "./data")
NC_PATH  = os.path.join(DATA_DIR, "temperature.nc")

# ── Open with Dask chunking ───────────────────────────────────────────────────
ds = xr.open_dataset(
    NC_PATH,
    chunks={"time": 4, "lat": -1, "lon": -1},   # chunk over time
)
print(ds["tas"])                   # shows dask-backed array; no data read yet
print(ds["tas"].chunks)

# ── Compute seasonal mean (lazy until .compute()) ─────────────────────────────
seasonal_mean = ds["tas"].groupby("time.season").mean("time")
print("\nSeasonal mean (lazy):", seasonal_mean)
# Trigger computation
result = seasonal_mean.compute()
print("Seasonal mean (computed):", result)

# ── Rechunk for better parallelism ───────────────────────────────────────────
# Current: chunked over time → good for time slices
# New: chunked over space  → good for spatial operations
ds_rechunked = ds.chunk({"time": -1, "lat": 9, "lon": 18})
print("\nRechunked chunks:", ds_rechunked["tas"].chunks)

# ── Apply a custom ufunc lazily ───────────────────────────────────────────────
def compute_anomaly(x, clim):
    """Subtract climatology from each month."""
    return x - clim

climatology = ds["tas"].mean("time")   # spatial mean map (lazy)
anomaly = xr.apply_ufunc(
    compute_anomaly,
    ds["tas"],
    climatology,
    dask="parallelized",
    output_dtypes=[float],
)
print("\nAnomaly (lazy):", anomaly)
anomaly_computed = anomaly.compute()
print("Anomaly std:", float(anomaly_computed.std()))

# ── Rolling statistics ────────────────────────────────────────────────────────
rolling_mean = ds["tas"].rolling(time=3, center=True).mean()
print("\n3-month rolling mean shape:", rolling_mean.compute().shape)

# ── Resample to quarterly ─────────────────────────────────────────────────────
quarterly = ds["tas"].resample(time="QS").mean()
print("\nQuarterly shape:", quarterly.compute().shape)

ds.close()

Advanced Usage

Convert NetCDF to Zarr for Cloud Storage

import xarray as xr
import zarr
import os
import numpy as np
import pandas as pd

DATA_DIR  = os.getenv("NC_DATA_DIR", "./data")
NC_PATH   = os.path.join(DATA_DIR, "temperature.nc")
ZARR_PATH = os.path.join(DATA_DIR, "temperature.zarr")

ds = xr.open_dataset(NC_PATH, chunks={"time": 4})

# Write to Zarr with explicit chunk encoding
zarr_encoding = {
    "tas": {
        "chunks":    (4, 37, 72),      # must match or divide dask chunks
        "compressor": zarr.Blosc(cname="lz4", clevel=5, shuffle=zarr.Blosc.SHUFFLE),
        "dtype":     "float32",
    }
}
ds.to_zarr(ZARR_PATH, encoding=zarr_encoding, mode="w")
print(f"Zarr store written to {ZARR_PATH}")

# Re-open from Zarr
ds_z = xr.open_zarr(ZARR_PATH)
print(ds_z)

# ── Inspect raw Zarr store ────────────────────────────────────────────────────
store = zarr.open(ZARR_PATH, mode="r")
print("\nZarr store keys:", list(store.keys()))
print("tas array info :", store["tas"].info)

ds.close(); ds_z.close()

Interpolation and Grid Remapping

import xarray as xr
import numpy as np
import pandas as pd

# Create two datasets on different grids
rng = np.random.default_rng(0)

lats_coarse = np.linspace(-90, 90, 19)    # 10° grid
lons_coarse = np.linspace(0, 350, 36)
times = pd.date_range("2020-01", periods=6, freq="ME")
data_coarse = rng.standard_normal((6, 19, 36)).astype("float32") + 280

ds_coarse = xr.Dataset(
    {"tas": (["time", "lat", "lon"], data_coarse)},
    coords={"time": times, "lat": lats_coarse, "lon": lons_coarse},
)

# Fine target grid
lats_fine = np.linspace(-90, 90, 73)    # 2.5° grid
lons_fine = np.linspace(0, 357.5, 144)

# Bi-linear interpolation
ds_fine = ds_coarse.interp(lat=lats_fine, lon=lons_fine, method="linear")
print("Coarse shape:", ds_coarse["tas"].shape)
print("Fine shape:  ", ds_fine["tas"].shape)

# Nearest-neighbor for categorical/mask data
ds_nn = ds_coarse.interp(lat=lats_fine, lon=lons_fine, method="nearest")

# ── Merge two datasets with different times ───────────────────────────────────
times2 = pd.date_range("2020-07", periods=6, freq="ME")
data2  = rng.standard_normal((6, 73, 144)).astype("float32") + 280
ds2    = xr.Dataset(
    {"tas": (["time", "lat", "lon"], data2)},
    coords={"time": times2, "lat": lats_fine, "lon": lons_fine},
)

ds_merged = xr.concat([ds_fine, ds2], dim="time")
print("Merged shape:", ds_merged["tas"].shape)

Weighted Averaging (Latitude Weights)

Area-weighted global mean accounts for the fact that grid cells near the poles are smaller than those at the equator.

import xarray as xr
import numpy as np
import pandas as pd

rng = np.random.default_rng(1)
lats = np.linspace(-90, 90, 73)
lons = np.linspace(0, 357.5, 144)
times = pd.date_range("2020-01-01", periods=12, freq="ME")
data  = (280 + 20 * np.cos(np.deg2rad(lats))[:, None]
         + rng.standard_normal((73, 12, 144))).transpose(1, 0, 2)

da = xr.DataArray(data, dims=["time", "lat", "lon"],
                  coords={"time": times, "lat": lats, "lon": lons},
                  name="tas")

# Cosine-latitude weights
weights = np.cos(np.deg2rad(da.lat))
weights.name = "weights"

# Weighted mean over lat and lon
global_mean = da.weighted(weights).mean(dim=["lat", "lon"])
print("Global mean time series:", global_mean.values)
print("Annual mean:", float(global_mean.mean()))

Working with CF Time Calendars

import xarray as xr
import cftime
import numpy as np

# 360-day calendar (used in many climate models)
times_360 = xr.cftime_range(
    start="1850-01-01", periods=120, freq="MS", calendar="360_day"
)
data_360 = np.random.rand(120, 10, 20).astype("float32")

da_360 = xr.DataArray(
    data_360,
    dims=["time", "lat", "lon"],
    coords={
        "time": times_360,
        "lat": np.linspace(-90, 90, 10),
        "lon": np.linspace(0, 342, 20),
    },
)
print("360-day calendar time range:", da_360.time.values[[0, -1]])
print("Number of time steps:", len(da_360.time))

# Select a decade
decade = da_360.sel(time=slice("1900-01-01", "1909-12-30"))
print("Decade shape:", decade.shape)

# Convert to standard Gregorian calendar by resampling is not trivial;
# the recommended path is to convert units after writing back to NetCDF.

Parallel Processing with Dask Distributed

import xarray as xr
import numpy as np
import pandas as pd
from dask.distributed import Client, LocalCluster
import os

DATA_DIR = os.getenv("NC_DATA_DIR", "./data")

def run_parallel_analysis():
    cluster = LocalCluster(n_workers=2, threads_per_worker=2)
    client  = Client(cluster)
    print("Dashboard:", client.dashboard_link)

    # Open a large (synthetic) dataset with dask
    times = pd.date_range("1950-01-01", periods=840, freq="ME")
    lats  = np.linspace(-90, 90, 73)
    lons  = np.linspace(0, 357.5, 144)
    rng   = np.random.default_rng(7)
    data  = rng.standard_normal((840, 73, 144)).astype("float32") + 280

    ds = xr.Dataset(
        {"tas": (["time", "lat", "lon"], data)},
        coords={"time": times, "lat": lats, "lon": lons},
    ).chunk({"time": 60, "lat": -1, "lon": -1})

    # Compute global mean in parallel
    weights = np.cos(np.deg2rad(ds["tas"].lat))
    global_mean = ds["tas"].weighted(weights).mean(["lat", "lon"])
    result = global_mean.compute()

    print("Global mean std (70 years):", float(result.std()))

    client.close(); cluster.close()
    return result

# result = run_parallel_analysis()

Troubleshooting

File opens but all variables show NaN

Cause: _FillValue is being decoded but the compression codec does not match the version of netCDF4 used to write the file.

Fix: Open with mask_and_scale=False to inspect raw values first.

ds_raw = xr.open_dataset("file.nc", mask_and_scale=False)
print(ds_raw["tas"].values[:5])    # inspect without NaN substitution

Then re-encode with explicit _FillValue:

encoding = {"tas": {"_FillValue": 1e20, "zlib": True}}
ds.to_netcdf("fixed.nc", encoding=encoding)

ValueError: conflicting sizes for dimension on open_mfdataset

Cause: Files have slightly different grid sizes or coordinate values.

Fix: Use join="override" and verify the coordinate mismatch first.

import xarray as xr

ds = xr.open_mfdataset(
    "data/*.nc",
    combine="by_coords",
    join="override",     # trust first file's coords
    compat="override",
)

Dask workers run out of memory

Cause: Chunks are too large or too many tasks run in parallel.

Fix: Reduce chunk size and limit concurrency.

from dask.distributed import Client

client = Client(n_workers=4, threads_per_worker=1, memory_limit="2GB")

# Rechunk to smaller pieces
ds = ds.chunk({"time": 12, "lat": 20, "lon": 36})

RuntimeError: NetCDF: HDF error on parallel read

Cause: HDF5 (the backend for NetCDF4) is not built with parallel I/O support by default; multiple workers cannot open the same file simultaneously.

Fix: Convert to Zarr first, then open with Dask.

import xarray as xr

ds = xr.open_dataset("big_file.nc")
ds.to_zarr("big_file.zarr", mode="w")
ds_z = xr.open_zarr("big_file.zarr")   # Zarr supports concurrent reads

Slow Zarr writes

Cause: Too many small chunks produce millions of tiny files.

Fix: Use at least 100 MB per chunk and pick a fast compressor.

import zarr

enc = {
    "tas": {
        "chunks": (120, 73, 144),
        "compressor": zarr.Blosc(cname="lz4", clevel=3),
    }
}
ds.to_zarr("output.zarr", encoding=enc, mode="w")

External Resources

• xarray documentation: https://docs.xarray.dev/en/stable/
• xarray tutorial: https://tutorial.xarray.dev/
• CF Conventions 1.11: https://cfconventions.org/cf-conventions/cf-conventions.html
• Dask documentation: https://docs.dask.org/en/stable/
• Zarr specification v2: https://zarr.readthedocs.io/en/stable/spec/v2.html
• netCDF4-python: https://unidata.github.io/netcdf4-python/
• Pangeo community (cloud-native geoscience): https://pangeo.io/
• xarray-tutorial notebooks: https://github.com/xarray-contrib/xarray-tutorial

Examples

Example 1 — ENSO Index from SST Data

Compute the Nino 3.4 sea-surface temperature anomaly index, a standard metric for El Nino / La Nina monitoring.

import xarray as xr
import numpy as np
import pandas as pd
import matplotlib.pyplot as plt

# ── Synthetic SST dataset ────────────────────────────────────────────────────
rng = np.random.default_rng(99)
times = pd.date_range("1980-01-01", periods=504, freq="ME")  # 42 years
lats  = np.linspace(-30, 30, 61)   # tropical band
lons  = np.linspace(120, 290, 171) # Pacific basin

# Add an El Nino-like signal with period ~4 years
enso_signal = 1.5 * np.sin(2 * np.pi * np.arange(504) / 48)
seasonal     = 0.5 * np.cos(2 * np.pi * np.arange(504) / 12)
base_sst     = 28.0 - 2.0 * np.abs(lats / 30)[:, None]  # (lat, 1)

# Broadcast: (time, lat, lon)
sst = (base_sst[None, :, :]
       + enso_signal[:, None, None] * np.cos(np.deg2rad(lats))[None, :, None]
       + seasonal[:, None, None]
       + 0.3 * rng.standard_normal((504, 61, 171))).astype("float32")

ds = xr.Dataset(
    {"sst": (["time", "lat", "lon"], sst,
              {"units": "degC", "long_name": "Sea Surface Temperature"})},
    coords={"time": times, "lat": lats, "lon": lons},
    attrs={"Conventions": "CF-1.10", "title": "Synthetic tropical SST"},
)

# ── Compute SST climatology (1980-2009 baseline) ─────────────────────────────
clim_period = ds.sel(time=slice("1980-01-01", "2009-12-31"))
climatology  = clim_period["sst"].groupby("time.month").mean("time")
print("Climatology shape:", climatology.shape)   # (12, lat, lon)

# ── Monthly anomaly ───────────────────────────────────────────────────────────
anomaly = ds["sst"].groupby("time.month") - climatology
print("Anomaly shape:", anomaly.shape)

# ── Nino 3.4 index: area mean in 5°S-5°N, 190°-240°E ─────────────────────────
nino34_box = anomaly.sel(lat=slice(-5, 5), lon=slice(190, 240))
weights    = np.cos(np.deg2rad(nino34_box.lat))
nino34     = nino34_box.weighted(weights).mean(["lat", "lon"])

# 5-month running mean (standard smoothing)
nino34_smooth = nino34.rolling(time=5, center=True, min_periods=3).mean()

# ── Plot ──────────────────────────────────────────────────────────────────────
fig, ax = plt.subplots(figsize=(14, 4))
ax.fill_between(nino34.time.values, nino34_smooth.values,
                where=nino34_smooth.values > 0.5,
                color="red", alpha=0.5, label="El Nino (>0.5°C)")
ax.fill_between(nino34.time.values, nino34_smooth.values,
                where=nino34_smooth.values < -0.5,
                color="blue", alpha=0.5, label="La Nina (<-0.5°C)")
ax.plot(nino34.time.values, nino34_smooth.values, "k", lw=1)
ax.axhline(0, color="k", lw=0.5)
ax.set_ylabel("Nino 3.4 SST anomaly (°C)")
ax.set_title("Nino 3.4 index — synthetic dataset")
ax.legend(); plt.tight_layout()
plt.savefig("enso_index.png", dpi=150)
print("Peak Nino 3.4 value:", float(nino34_smooth.max()))

Example 2 — Vertical Profile Analysis with Pressure Levels

Work with 4-D (time, level, lat, lon) atmospheric data, compute geopotential thickness, and save diagnostics to Zarr.

import xarray as xr
import numpy as np
import pandas as pd
import matplotlib.pyplot as plt
import os

DATA_DIR  = os.getenv("NC_DATA_DIR", "./data")
ZARR_OUT  = os.path.join(DATA_DIR, "thickness.zarr")
os.makedirs(DATA_DIR, exist_ok=True)

rng = np.random.default_rng(42)

# ── Create synthetic 4-D pressure-level dataset ───────────────────────────────
pressure_levels = np.array([1000, 925, 850, 700, 500, 300, 200, 100],
                            dtype="float32")  # hPa
times = pd.date_range("2020-01-01", periods=60, freq="D")
lats  = np.linspace(-90, 90, 37)
lons  = np.linspace(0, 355, 72)

# Temperature: lapse rate + random noise + seasonal
T_base   = 290 - 6.5e-3 * 8000 * (1 - pressure_levels / 1000)[:, None, None]
T_data   = (T_base[None, :, :, :]
            + 10 * np.cos(np.deg2rad(lats))[None, None, :, None]
            + rng.standard_normal((60, 8, 37, 72)).astype("float32") * 2)

# Geopotential height (synthetic, proportional to log-pressure)
Z_data = (-(8e3 * np.log(pressure_levels / 1013.25))[:, None, None]
           + rng.standard_normal((60, 8, 37, 72)).astype("float32") * 50).astype("float32")

ds = xr.Dataset(
    {
        "ta": (["time", "plev", "lat", "lon"], T_data.astype("float32"),
               {"units": "K", "standard_name": "air_temperature",
                "long_name": "Air Temperature"}),
        "zg": (["time", "plev", "lat", "lon"], Z_data,
               {"units": "m", "standard_name": "geopotential_height",
                "long_name": "Geopotential Height"}),
    },
    coords={
        "time": times,
        "plev": xr.Variable("plev", pressure_levels,
                             attrs={"units": "hPa", "axis": "Z",
                                    "positive": "down"}),
        "lat": lats, "lon": lons,
    },
    attrs={"Conventions": "CF-1.10"},
)

# ── 1000-500 hPa thickness ────────────────────────────────────────────────────
thickness = (ds["zg"].sel(plev=500)
             - ds["zg"].sel(plev=1000))
thickness.attrs = {"units": "m", "long_name": "1000-500 hPa Thickness"}

# ── Tropopause temperature (minimum T in column above 200 hPa) ────────────────
strat = ds["ta"].sel(plev=slice(None, 300))   # 300 hPa and above (lower pressure)
tropopause_T = strat.min(dim="plev")

# ── Zonal mean cross-section ──────────────────────────────────────────────────
zonal_mean_T = ds["ta"].mean("lon").mean("time")  # (plev, lat)

fig, ax = plt.subplots(figsize=(10, 6))
cf = ax.contourf(lats, pressure_levels, zonal_mean_T.values,
                 levels=20, cmap="RdBu_r")
ax.set_ylim(1000, 100)
ax.set_xlabel("Latitude (°N)")
ax.set_ylabel("Pressure (hPa)")
ax.set_title("Zonal mean temperature (K)")
plt.colorbar(cf, ax=ax, label="K")
plt.tight_layout()
plt.savefig("zonal_mean_T.png", dpi=150)

# ── Save diagnostics to Zarr ─────────────────────────────────────────────────
diag_ds = xr.Dataset(
    {"thickness":    thickness,
     "tropopause_T": tropopause_T},
)
diag_ds.to_zarr(ZARR_OUT, mode="w")
print(f"Diagnostics saved to {ZARR_OUT}")

# Verify round-trip
check = xr.open_zarr(ZARR_OUT)
print("Zarr thickness range:",
      float(check["thickness"].min()), "to",
      float(check["thickness"].max()), "m")
check.close()