<tt>COMET</tt>: Clustering observables modelled by emulated perturbation theory

In this paper we present COMET, a Gaussian process emulator of the galaxy power spectrum multipoles in redshift-space. The model predictions are based on one-loop perturbation theory and consider two alternative descriptions redshift-space distortions: one that performs full expansion real- to mapping, as recent effective field models, another preserves non-perturbative impact small-scale velocities by means an damping function. outputs COMET can be obtained at arbitrary redshifts (up $z \sim 3$), for fiducial background cosmologies, large parameter space covers shape parameters $\omega_c$, $\omega_b$, $n_s$, well evolution $h$, $A_s$, $\Omega_K$, $w_0$, $w_a$. This flexibility does not impair COMET's accuracy, since exploit exact degeneracy between allows us train significantly reduced space. While sped up least orders magnitude, validation tests reveal accuracy $0.1\,\%$ monopole quadrupole ($0.3\,\%$ hexadecapole), or alternatively, better than $0.25\,\sigma$ all three comparison statistical uncertainties expected Euclid survey with tenfold increase volume. We show these differences translate into shifts mean posterior values most same size, meaning used confidence underlying models. is publicly available Python package also provides tree-level bispectrum covariance matrices.