In the process of quenching heat treatment, it is critical to establish the optimal process parameters producing the least residual stress magnitudes and related distortions and/or cracking, to reduce the cost of manufacturing high-quality components with intricate and durable designs and meet the stringent requirements of a broad range of high-performance industries. Because such e ects occur as the result of uneven cooling in di erent regions of the quenched part, a feasible control objective is thus to enhance the spatial uniformity of heat removal, to prevent spatial gradients of irreversible strains typically originating from heterogeneous plastic deformation. For decades this process has been largely driven by trial and error, intuition and experience. In this study, a single-step Deep Reinforcement Learning (DRL) algorithm is used to provide the best possible cooling rate in industrial quenching processes governed by coupled pseudo-compressible Navier-Stokes and heat equations, along with latent heat formulation. The numerical reward fed to the neural network is computed with an in-house stabilized finite elements environment combining variational multi-scale (VMS) modeling of the governing equations, immerse volume method, and multi-component anisotropic mesh adaptation. A case of Rayleigh-Bénard convection in a highaspect ratio, closed cavity is used first as testbed for the proposed methodology. In a second phase, we tackle several quenching numerical experiments aiming at improving temperature homogeneity within two-dimensional components in various shapes, whose results showcase the potential of DRL to produce unanticipated solutions by learning the e ect of highly unsteady boiling flow physics on the temperature distribution.