• Python <3.9 is no longer supported.


  • Fixed small issue in Lightcone.lightcone_coords.


  • New Lightconer base class with user-facing classes RectilinearLightconer and AngularLightconer. These are now the recommended way to define how lightcones should be created. Note that these are not the final lightcone object, but simply a “definition” for how to create the lightcone.

  • The Lightcone class still exists, but is now joined by an AngularLightcone class that is the output lightcone when using the AngularLightconer definition.

  • New KEEP_3D_VELOCITIES user-param. This is required to be True if you want to apply RSDs to an AngularLightcone.

  • Ability to check-point lightcone creation, so that it can be resumed later. Simply pass the lightcone_fname argument to run_lightcone.

  • New cache_tools.get_boxes_at_redshift function to get all boxes at a given redshift, regardless of the parameters used to create them.

  • New CosmoParams.from_astropy() classmethod for constructing a cosmo params object directly from an astropy cosmology.


  • Refactored RSDs to the subcell_rsds.c C-module.

  • Added CI testing of the docs tutorials.

  • More precise debugging summaries of boxes in the C code.


  • Incorrect sign on adiabatic fluctuations.

v3.3.1 [24 May 2023]


  • Compilation of C code for some compilers (#330)

v3.3.0 [17 May 2023]


  • Refactored setting up of inputs to high-level functions so that there is less code repetition.


  • Running with R_BUBBLE_MAX too large auto-fixes it to be BOX_LEN (#112)

  • Bug in calling clear_cache.

  • Inconsistency in the way that the very highest redshift of an evolution is handled between low-level code (eg. spin_temperature()) and high-level code (eg. run_coeval()).


  • New validate_all_inputs function that cross-references the four main input structs and ensures all the parameters make sense together. Mostly for internal use.

  • Ability to save/read directly from an open HDF5 File (#170)

  • An implementation of cloud-in-cell to more accurately redistribute the perturbed mass across all neighbouring cells instead of the previous nearest cell approach

  • Changed PhotonConsEndCalibz from z = 5 -> z = 3.5 to handle later reionisation scenarios in line with current observations (#305)

  • Add in an initialisation check for the photon conservation to address some issues arising for early EOR histories (#311)

  • Added NON_CUBIC_FACTOR to UserParams to allow for non-cubic coeval boxes (#289)

v3.2.1 [13 Sep 2022]


  • Included log10_mturnovers(_mini) in lightcone class. Only useful when USE_MINI_HALOS

v3.2.0 [11 Jul 2022]


  • Floats are now represented to a specific number of significant digits in the hash of an output object. This fixes problems with very close redshifts not being read from cache (#80). Note that this means that very close astro/cosmo params will now be read from cache. This could cause issues when creating large databases with many random parameters. The behaviour can modified in the configuration by setting the cache_param_sigfigs and cache_redshift_sigfigs parameters (these are 6 and 4 by default, respectively). NOTE: updating to this version will cause your previous cached files to become unusable. Remove them before updating.


  • Added a missing C-based error to the known errors in Python.

v3.1.5 [27 Apr 2022]

v3.1.4 [10 Feb 2022]


  • error in FFT normalization in FindHaloes

  • docs not compiling on RTD due to missing scipy.integrate mock module

  • Updated matplotlib removed support for setting vmin/vmax and norm. Now passes vmin/vmax to the norm() constructor.

v3.1.3 [27 Oct 2021]

  • Fixed FAST_FCOLL_TABLES so it only affects MCGs and not ACGs. Added tests of this flag for high and low z separately.

v3.1.2 [14 Jul 2021]


  • MINIMIZE_MEMORY flag significantly reduces memory without affecting performance much, by changing the way some arrays are allocated and accessed in C. (#224)


  • Updated USE_INTERPOLATION_TABLES to be default True. This makes much more sense as a default value. Until v4, a warning will be raised if it is not set explicitly.

v3.1.1 [13 Jun 2021]


  • Bug in deployment to PyPI.

v3.1.0 [13 Jun 2021]


  • Ability to access all evolutionary Coeval components, either from the end Coeval class, or the Lightcone.

  • Ability to gather all evolutionary antecedents from a Coeval/Lightcone into the one file.

  • FAST_FCOLL_TABLES in UserParams which improves speeds quite significantly for ~<10% accuracy decrease.

  • Fast and low-memory generation of relative-velocity (vcb) initial conditions. Eliminated hi-res vcb boxes, as they are never needed.

  • Also output the mean free path (i.e. MFP_box in IonizedBox).

  • Added the effect of DM-baryon relative velocities on PopIII-forming minihaloes. This now provides the correct background evolution jointly with LW feedback. It gives rise to velocity-induced acoustic oscillations (VAOs) from the relative-velocity fluctuations. We also follow a more flexible parametrization for LW feedback in minihaloes, following new simulation results, and add a new index ALPHA_STAR_MINI for minihaloes, now independent of regular ACGs.

  • New hooks keyword to high-level functions, that are run on the completion of each computational step, and can be used to more generically write parts of the data to file.

  • Ability to pass a function to write= to write more specific aspects of the data (internally, this will be put into the hooks dictionary).

  • run_lightcone and run_coeval use significantly less memory by offloading initial conditions and perturb_field instances to disk if possible.


  • Bug in 2LPT when USE_RELATIVE_VELOCITIES=True [Issue #191, PR #192]

  • Error raised when redshifts are not in ascending order [Issue #176, PR #177]

  • Errors when USE_FFTW_WISDOM is used on some systems [Issue #174, PR #199]

  • Bug in ComputeIonizedBox causing negative recombination rate and ring structure in Gamma12_box [Issue #194, PR #210]

  • Error in determining the wisdom file name [Issue #209, PR#210]

  • Bug in which cached C-based memory would be read in and free’d twice.


  • Added dft.c, which makes doing all the cubic FFTs a lot easier and more consistent. [PR #199]

  • More generic way of keeping track of arrays to be passed between C and Python, and their shape in Python, using _get_box_structures. This also means that the various boxes can be queried before they are initialized and computed.

  • More stringent integration tests that test each array, not just the final brightness temperature.

  • Ability to plot the integration test data to more easily identify where things have gone wrong (use --plots in the pytest invocation).

  • Nicer CLI interface for New options to clean the test_data/ directory, and also test data is saved by user-defined key rather than massive string of variables.

  • Nicer debug statements before calls to C, for easily comparing between versions.

  • Much nicer methods of keeping track of array state (in memory, on disk, c-controlled, etc.)

  • Ability to free C-based pointers in a more granular way.



  • coeval_callback and coeval_callback_redshifts flags to the run_lightcone. Gives the ability to run arbitrary code on Coeval boxes.

  • JOSS paper!

  • get_fields classmethod on all output classes, so that one can easily figure out what fields are computed (and available) for that class.


  • Only raise error on non-available external_table_path when actually going to use it.



  • Added prototype functions to enable compilation for some standard compilers on MacOS.


Modifications to the internal code structure of 21cmFAST


  • Refactor FFTW wisdom creation to be a python callable function


Complete overhaul of 21cmFAST, including a robust python-wrapper and interface, caching mechanisms, and public repository with continuous integration. Changes and equations for minihalo features in this version are found in

All functionality of the original 21cmFAST v2 C-code has been implemented in this version, including USE_HALO_FIELD and performing full integration instead of using the interpolation tables (which are faster).


  • Updated the radiation source model: (i) all radiation fields including X-rays, UV ionizing, Lyman Werner and Lyman alpha are considered from two seperated population namely atomic-cooling (ACGs) and minihalo-hosted molecular-cooling galaxies (MCGs); (ii) the turn-over masses of ACGs and MCGs are estimated with cooling efficiency and feedback from reionization and lyman werner suppression (Qin et al. 2020). This can be switched on using new flag_options USE_MINI_HALOS.

  • Updated kinetic temperature of the IGM with fully ionized cells following equation 6 of McQuinn (2015) and partially ionized cells having the volume-weightied temperature between the ionized (volume: 1-xHI; temperature T_RE ) and neutral components (volume: xHI; temperature: temperature of HI). This is stored in IonizedBox as temp_kinetic_all_gas. Note that Tk in TsBox remains to be the kinetic temperature of HI.

  • Tests: many unit tests, and also some regression tests.

  • CLI: run 21cmFAST boxes from the command line, query the cache database, and produce plots for standard comparison runs.

  • Documentation: Jupyter notebook demos and tutorials, FAQs, installation instructions.

  • Plotting routines: a number of general plotting routines designed to plot coeval and lightcone slices.

  • New power spectrum option (POWER_SPECTRUM=5) that uses a CLASS-based transfer function. WARNING: If POWER_SPECTRUM==5 the cosmo parameters cannot be altered, they are set to the Planck2018 best-fit values for now (until CLASS is added): (omegab=0.02237, omegac= 0.120, hubble=0.6736 (the rest are irrelevant for the transfer functions, but in case: A_s=2.100e-9, n_s=0.9649, z_reio = 11.357)

  • New user_params option USE_RELATIVE_VELOCITIES, which produces initial relative velocity cubes (option implemented, but not the actual computation yet).

  • Configuration management.

  • global params now has a context manager for changing parameters temporarily.

  • Vastly improved error handling: exceptions can be caught in C code and propagated to Python to inform the user of what’s going wrong.

  • Ability to write high-level data (Coeval and Lightcone objects) directly to file in a simple portable format.


  • POWER_SPECTRUM option moved from global_params to user_params.

  • Default cosmology updated to Planck18.


All changes and equations for this version are found in


  • Updated the ionizing source model: (i) the star formation rates and ionizing escape fraction are scaled with the masses of dark matter halos and (ii) the abundance of active star forming galaxies is exponentially suppressed below the turn-over halo mass, M_{turn}, according to a duty cycle of exp(−M_{turn}/M_{h}), where M_{h} is a halo mass.

  • Removed the mean free path parameter, R_{mfp}. Instead, directly computes inhomogeneous, sub-grid recombinations in the intergalactic medium following the approach of Sobacchi & Mesinger (2014)



  • Support for a halo mass dependent ionizing efficiency: zeta = zeta_0 (M/Mmin)^alpha, where zeta_0 corresponds to HII_EFF_FACTOR, Mmin –> ION_M_MIN, alpha –> EFF_FACTOR_PL_INDEX in ANAL_PARAMS.H



  • Code ‘redshift_interpolate_boxes.c’ to interpolate between comoving cubes, creating comoving light cone boxes.

  • Enabled openMP threading for SMP machines. You can specify the number of threads (for best performace, do not exceed the number of processors) in INIT_PARAMS.H. You do not need to have an SMP machine to run the code. NOTE: YOU SHOULD RE-INSTALL FFTW to use openMP (see INSTALL file)

  • Included a threaded driver file ‘drive_zscroll_reion_param.c’ set-up to perform astrophysical parameter studies of reionization

  • Included explicit support for WDM cosmologies; see COSMOLOGY.H. The prescription is similar to that discussed in Barkana+2001; Mesinger+2005, madifying the (i) transfer function (according to the Bode+2001 formula; and (ii) including the effective pressure term of WDM using a Jeans mass analogy. (ii) is approximated with a sharp cuttoff in the EPS barrier, using 60* M_J found in Barkana+2001 (the 60 is an adjustment factor found by fitting to the WDM collapsed fraction).

  • A Gaussian filtering step of the PT fields to perturb_field.c, in addition to the implicit boxcar smoothing. This avoids having”empty” density cells, i.e. delta=-1, with some small loss in resolution. Although for most uses delta=-1 is ok, some Lya forest statistics do not like it.

  • Added treatment of the risidual electron fraction from X-ray heating when computing the ionization field. Relatedly, modified Ts.c to output all intermediate evolution boxes, Tk and x_e.

  • Added a missing factor of Omega_b in Ts.c corresponding to eq. 18 in MFC11. Users who used a previous version should note that their results just effecively correspond to a higher effective X-ray efficiency, scaled by 1/Omega_baryon.

  • Normalization optimization to Ts.c, increasing performace on arge resolution boxes


  • GSL interpolation error in kappa_elec_pH for GSL versions > 1.15

  • Typo in macro definition, which impacted the Lya background calculation in v1.11 (not applicable to earlier releases)

  • Outdated filename sytax when calling gen_size_distr in drive_xHIscroll

  • Redshift scrolling so that drive_logZscroll_Ts.c and Ts.c are in sync.


  • Output format to avoid FFT padding for all boxes

  • Filename conventions to be more explicit.

  • Small changes to organization and structure



  • Wrapper functions mod_fwrite() and mod_fread() in Cosmo_c_progs/misc.c, which should fix problems with the library fwrite() and fread() for large files (>4GB) on certain operating systems.

  • Included print_power_spectrum_ICs.c program which reads in high resolution initial conditions and prints out an ASCII file with the associated power spectrum.

  • Parameter in Ts.c for the maximum allowed kinetic temperature, which increases stability of the code when the redshift step size and the X-ray efficiencies are large.


  • Oversight adding support for a Gaussian filter for the lower resolution field.