21cmKAN

This is the initial release of the 21cmKAN emulator code from GitHub (https://github.com/jdorigojones/21cmKAN). Please see the GitHub page for further information about how to install, use, and adapt 21cmKAN. This code was created by @jdorigojones and @b-reyes. Please contact me (Johnny; johnny.dorigojones@colorado.edu) regarding any suggestions or issues you may have when using or adapting this code!

21cmKAN is an emulator of the global 21 cm cosmological signal based on the Kolmogorov-Arnold Network. KANs are a novel type of fully-connected neural network that capture complex relationships by learning data-driven functional transformations, or activation functions, as opposed to using fixed, pre-determined activations (see figure below). The expressivity of KANs makes them useful for modeling certain structured, lower-dimensional functions or PDEs often found in science, and their transparent architecture makes it easy to interpret and verify their predictions.

21cmKAN has similar accuracy as the most accurate current emulator of the global 21 cm signal, 21cmLSTM (Dorigo Jones et al. 2024), while training 75 times faster and predicting each signal in 3.7 milliseconds on average, when utilizing the same typical A100 GPU and training on the same data. 21cmKAN can be trained and used to obtain unbiased physical parameter constraints altogether in under 30 minutes. The speed-accuracy combination of 21cmKAN enables producing many emulator models that can constrain complex feature spaces and covariances across different physical models and parameterizations to fully exploit upcoming observations.

The tutorial notebooks and commented scripts provided in the GitHub repository make 21cmKAN simple to train, evaluate, employ in Bayesian inference analyses, and apply to different physical models and data sets. Please see the associated paper -- Dorigo Jones et al. 2025 -- for details on the architecture, training, and interpretation of 21cmKAN, as well as high-level and in-depth descriptions of the unique differences and advantages of KANs compared to traditional fully-connected neural networks. 21cmKAN is free to use on the MIT open source license. All of the data used to train and test 21cmKAN in DJ+25 is publicly-available on Zenodo: 21cmGEM/21cmVAE data set; ARES data set.