Driven by the need to reduce emissions from aviation, improving efficiency is the central challenge in modern gas turbine design. Applying Machine Learning (ML) techniques in place of direct simulation in the loop is of increasing prominence due to near instant results. However, they have not yet achieved the robustness necessary for industrial application in the turbomachinery field. This work aims to detail a generalisable ML framework applied to a highly-loaded, transonic compressor – NASA Rotor37. The model targets achieving equivalent accuracy of flow structures and extracted performance parameters to steady state simulations for a given physics fluid regime, with less than 200 training data points for 18 varied Engineering Parameters. Central to the framework is a novel, geometry-agnostic processing technique that bypasses the limitations of traditional blade parameterisation, enabling seamless scalability across diverse topologies. By encoding raw blade geometry and boundary conditions directly into an asymmetric UNET architecture [1], the model performs a complex dimensional mapping to generate precise 2D flow fields. Unlike prevalent works restricted to fixed geometric templates [2], this method leverages physics-informed data preparation to ensure the model remains generalisable and capable of predicting out-of-distribution blade designs. The model accurately predicts five flow state variables across an off-design blade characteristic, capturing both stall and choked flow behaviours. The values consistently exceeding 0.99 are achieved for local state variables, while derived isentropic efficiency maintains a mean of 0.96 across the full off-design location. The proposed methodology exhibits the robustness required for industrial application, accurately predicting off- design performance for a novel geometries containing topological features absent from the training set. When integrated into an optimisation workflow, the framework identifies optimal configurations that closely align with computational fluid dynamics results, reducing computational overhead to near negligible.