Enhancing Cosmological Model Selection with Interpretable Machine Learning

We propose a novel approach using neural networks (NNs) to differentiate between cosmological models, especially in the case where they are nested and the additional model parameters are close to zero, making it difficult to discriminate them with traditional approaches. Our method complements Bayesian analyses for cosmological model selection, which heavily depend on the chosen priors and average the unnormalized posterior over potentially large prior volumes. By analyzing simulated realistic data sets of the growth rate of the large scale structure (LSS) of the Universe, based on current galaxy-clustering survey specifications, for the cosmological constant and cold dark matter (ΛCDM) model and the Hu-Sawicki $f(R)$ model, we demonstrate the potential of NNs to enhance the extraction of meaningful information from cosmological LSS data. We find that the NN can successfully distinguish between ΛCDM and the $f(R)$ models, by predicting the correct model with approximately $97%$ overall accuracy, thus demonstrating that NNs can maximise the potential of current and next generation surveys to probe for deviations from general relativity.