Model data format

This section describes how all input model data must be “formatted” for use by the framework. By “format” we mean not only the binary file format, but also the organization of data within and across files and metadata conventions.

All these questions are answered by the choice of data source: different data sources can, in principle, process data from different formats. The format information is presented in this section because, currently, all data sources require input data in the format described below. This accommodates the majority of use cases encountered at NCAR and GFDL; in particular, any CF-compliant data, or data published as part of CMIP6, should satisfy the requirements below.

A core design goal of this project is the ability to run diagnostics seamlessly on data from a wide variety of sources, including different formats. If you would like the package to support formats or metadata conventions that aren’t currently supported, please make a request in the appropriate GitHub discussion thread.

Model data format requirements

File organization

  • Model data must be supplied in the form of a set of netCDF files.

  • We support reading from all netCDF-3 and netCDF-4 binary formats, as defined in the netCDF FAQ, via xarray (see below), with the exception that variables nested under netCDF-4 groups are not currently supported.

  • Each netCDF file should only contain one variable (i.e., an array with the values of a single dependent variable, along with all the values of the coordinates at which the dependent variable was sampled). Additional variables (coordinate bounds, auxiliary or transformed coordinates) may be present in the file, but will be ignored by the framework.

  • The data for one variable may be spread across multiple netCDF files, but this must take the form of contiguous chunks by date (e.g., one file for 2000-2009, another for 2010-2019, etc.). The spatial coordinates in each file in a series of chunks must be identical.


  • The framework currently only supports model data provided on a latitude-longitude grid.

  • The framework currently only supports vertical coordinates given in terms of pressure. The pressure coordinate may be in any units (mb, Pa, atm, …). We hope to offer support for parametric vertical coordinates in the near future.

  • The time coordinate of the data must follow the CF conventions; in particular, it must have a calendar attribute which matches one of the CF conventions’ recognized calendars (case-insensitive).

  • The framework doesn’t impose any limitations on the minimum or maximum resolution of model data, beyond the storage and memory available on the machine where the PODs are run.


The framework currently makes use of two pieces of metadata (attributes for each variable in the netCDF header), in addition to the calendar attribute on the time coordinate:

  • units: Required for all variables and coordinates. This should be a string of the form recognized by UDUNITS2, specifically the python cfunits package (which improves CF convention support, e.g. by recognizing 'psu' as “practical salinity units.”)

    This attribute is required because we allow PODs to request model data with specific units, rather than requiring each POD to implement and debug redundant unit conversion logic. Instead, unit checking and conversion is done by the framework. This can’t be done if it’s not clear what units the input data are in.

  • standard_name: If present, should be set to a recognized CF convention standard name. This is used to confirm that the framework has downloaded the physical quantity that the POD has requested, independently of what name the model has given to the variable.

    If the user or data source has specified a naming convention, missing values for this attribute will be filled in based on the variable names used in that convention.

Many utilities exist for editing metadata in netCDF headers. Popular examples are the ncatted tool in the NCO utilities and the setattribute operator in CDO, as well as the functionality provided by xarray itself. Additionally, the Explicit file data source provides limited functionality for overwriting metadata attributes.

In situations where none of the above options are feasible, the --disable-preprocessor command-line flag may be used to disable all functionality based on this metadata. In this case, the user is fully responsible for ensuring that the input model data has the units required by each POD, is provided on the correct pressure levels, etc.

xarray reference implementation

The framework uses xarray to preprocess and validate model data before the PODs are run; specifically using the netcdf4 engine and with CF convention support provided via the cftime library. We also use cf_xarray to access data attributes in a more convention-independent way.

If you’re deciding how to post-process your model’s data for use by the MDTF package, or are debugging issues with your model’s data format, it may be simpler to load and examine your data with these packages interactively, rather than by invoking the entire MDTF package. The following python snippet approximates how the framework loads datasets for preprocessing. Use the _MDTF_base conda environment to install the correct versions of each package.

import cftime, cf_xarray
import xarray as xr

ds = xr.open_mfdataset(
    [<path to first file>, <second file>, ...],
    data_vars='minimal', coords='minimal',
    compat='equals', join='exact',
    decode_times=True, use_cftime=True
# match coordinates to X/Y/Z/T axes using cf_xarray:
ds =
# print summary

The framework has additional logic for cleaning up noncompliant metadata (e.g., stripping whitespace from netCDF headers), but if you can load a dataset with the above commands, the framework should be able to deal with it as well.

If the framework runs into errors when run on a dataset that meets the criteria above, please file a bug report via the gitHub issue tracker.