Submesoscale fronts and filaments in the ocean mixed layer generate strong vertical motions that can substantially influence vertical heat exchange, yet these processes are largely unresolved in climate-scale ocean models. As a result, accurate parameterization of submesoscale vertical heat flux is essential. Here, we aim to learn the relationships between submesoscale heat fluxes and commonly available ocean variables by training machine-learning models on submesoscale-resolving data. To emulate the information available to coarse-resolution models, high-resolution fields are temporally filtered and spatially coarse-grained to a target grid scale. From the resulting coarse mixed-layer fields, we construct physically interpretable predictors that characterize frontal strength and local kinematics, including buoyancy, horizontal gradients, relative vorticity, divergence, strain, mixed-layer depth, and horizontal velocity magnitude. The learning target is the submesoscale contribution to the mixed-layer vertical heat flux, diagnosed as the residual between the high-resolution flux and its mesoscale component. Given the strong spatial coherence and temporal persistence of submesoscale signals in the mixed layer, we adopt a UNet–LSTM architecture. The UNet maps coarse-resolution inputs to spatially distributed unresolved heat-flux patterns, while the LSTM incorporates short-term temporal context. Results demonstrate accurate and stable predictions on an unseen test dataset and support the feasibility of machine-learning-based parameterizations for representing submesoscale heat fluxes in climate models.