Physics-informed conditional neural processes for reconstructing transient temperature fields of turbine blades from sparse maximum temperature data

Reconstructing the full transient temperature field of a turbine blade from sparse maximum temperature data obtained via micro-crystal measurements remains a significant engineering challenge. To address this, the present study first performs multi-condition conjugate heat transfer simulations on a periodic single-passage blade model. The simulation model is validated against quasi-steady-state temperature distributions measured by dual infrared thermal imaging cameras on both the suction and pressure sides, achieving a mean absolute percentage error (MAPE) of less than 3.35%. A conditional neural process network is then constructed to infer the maximum temperature across the entire blade point cloud and to simultaneously retrieve the four inlet operating condition parameters from only a few measurement points, achieving a MAPE of 1.01% for the retrieved parameters. By integrating turbine blade heat transfer mechanisms with simulation-based prior knowledge, this study proposes an innovative temperature time-history model. In this model, the parameters γ , q , τ and β comprehensively capture the transient temperature evolution at measurement points under varying conditions. Furthermore, a physics-informed neural network (PINN) framework for conjugate heat transfer problems is established. This framework couples the fluid and solid domains by enforcing temperature and heat flux continuity at the interface, while allowing for the flexible incorporation of diverse physical models. Case studies demonstrate that, compared with the CFD results, the maximum relative error is less than 6.59% and mean relative error is less than 1.66% for all nodes at quasi-steady state. The proposed approach offers a novel methodology for accurately reconstructing the full transient temperature field of a turbine blade from limited maximum temperature measurements. While further validation under realistic service temperatures is still required, the results obtained in this study demonstrate the practical potential of the proposed method.