Porous corrosion deposits (CRUD) on PWR fuel cladding alter the heat path from the cladding to the coolant. Under high heat flux a vapor film forms inside the deposit, and that film thickness controls cladding superheat and boric-acid enrichment. High-fidelity two-phase film-boiling models obtain the film thickness from a hydrodynamic force balance, but only through an outer iteration on the interface saturation temperature. Linear correlations of film thickness cannot represent parameter interactions. This paper develops a physics-informed machine-learning model for vapor-film thickness as a function of CRUD porosity, chimney density, deposit thickness, particle size and heat flux. The network is trained on two-phase film-boiling solutions from a fully coupled heat, flow, chemistry and radiolysis framework and is constrained by capillary (and, where active, disjoining) interface balance, energy consistency across the liquid- and vapor-saturated regions, and the geometric bound that the film cannot exceed the deposit thickness. The model recovers the dependence of film thickness on the CRUD properties and operating conditions and remains consistent with the teacher solver when the predicted thickness is returned to the two-phase model. Film thickness is obtained in a single forward pass, replacing the 3–8 outer Python–COMSOL iterations (typically under 245 s per case) required by the existing coupled solver. The approach is intended for rapid screening of partial dry-out and CRUD-induced power-shift risk.