A convolutional-neural-network estimator of CMB constraints on dark matter energy injection

We show that the impact of energy injection by dark matter annihilation on the cosmic microwave background power spectra can be apprehended via a residual likelihood map. By resorting to convolutional neural networks that can fully discover the underlying pattern of the map, we propose a novel way of constraining dark matter annihilation based on the Planck 2018 data. We demonstrate that the trained neural network can efficiently predict the likelihood and accurately place bounds on the annihilation cross-section in a model-independent fashion. The machinery will be made public in the near future. (paper)