In order to suppress pile-up effects and improve resolution, the ATLAS experiment at the LHC employs a suite of working points for missing transverse momentum ($p_\text{T}^\text{miss}$) reconstruction, and each is optimal for different event topologies and different beam conditions. A neural network (NN) can exploit various event properties to pick the optimal working point on an event-by-event basis, and also combine complementary information from each of the working points. The resulting regressed $p_\text{T}^\text{miss}$ (METNet) offers improved resolution and pile-up resistance across a number of different topologies compared to the current $p_\text{T}^\text{miss}$ working points. Additionally, by using the NN's confidence in its predictions, a machine learning-based $p_\text{T}^\text{miss}$ significance (`METNetSig') can be defined. This contribution presents simulation-based studies of the behaviour and performance of METNet and METNetSig for several topologies compared to current ATLAS $p_\text{T}^\text{miss}$ reconstruction methods.