Conventional　hybrid　beamforming　(BF)　techniques　encounter　high　computational　complexity　(CC)　and　performance　loss　due　to　array　steering　vector　mismatches.　Therefore,　in　this　letter,　a　joint　robust　adaptive　BF　(RAB)　method　based　on　the　diagonal　loading　technique　along　with　phase-only　digital　beamformer　design　is　proposed.　In　addition,　with　the　aim　of　reducing　the　CC　of　the　system,　a　novel　deep-learning　model　is　proposed　to　estimate　the　digital　weights.　Simulations　demonstrated　that　the　proposed　deep　neural　network　(DNN)　model　can　have　similar　performance　for　digital　BF　weights　estimation　as　a　metaheuristic-based　one　with　significantly　lower　CC.