Topology optimisation provides powerful design tools to engineer efficient industrial fluid systems, enabling the automated generation of high-performance configurations. However, designing new configurations with modified parameter settings (e.g., physical properties, operating conditions, …) typically requires executing computation and optimisation tasks from scratch. This leads to computationally expensive pipelines, slowing the design process and limiting the exploration of the design space in practical workflows. In this work, we rely on a recent extension of the surrogate modelling concept, namely a surrogate model of optimised topologies [1]. Instead of accelerating the optimisation loop by replacing full-order simulations with surrogate models, the proposed approach directly learns the map between user-defined design parameters and the corresponding optimised topologies. In this context, Deep Reinforcement Learning (DRL) is employed as a derivative-free optimisation strategy based on convolutional neural networks [2], to construct the aforementioned surrogate mapping for geometrically-parameterised fluid problems. Unlike conventional supervised learning approaches, DRL does not require a precomputed dataset, making it particularly suitable for problems where data generation is costly, such as topology optimisation. The resulting surrogate model thus provides an informed initial guess for traditional topology optimisation pipelines, accelerating the early stages of the design process in parameterised settings and allowing an efficient exploration of the design space. The capabilities of the proposed DRL-based surrogate model for topology optimisation are demonstrated through two numerical examples: the optimisation of a pipe bend to minimise pressure drop and the design of a fluid diode to maximise diodicity.