Zero-shot physics-consistent imputation for spatio-temporal sensor databases

Missing data in spatio-temporal sensor databases impairs downstream analytics in industrial and urban systems. Traditional imputation methods, such as K-Nearest Neighbors (KNN) and localized data imputation (LDI), rely primarily on statistical or geometric proximity and often ignore governing physical laws. We propose PhysImpute, a zero-shot physics-consistent imputation framework based on Sparse Finite-Element Operator Networks (Sparse FEONet). By modeling sensors as a sparse nite-element mesh, PhysImpute learns a continuous operator that recovers missing values while reducing physical residuals. The framework is designed for deployment without per-mask retraining and supports e cient inference across varying missingness patterns. Experiments on two synthetic PDE datasets and two real traffic datasets show a favorable accuracyphysics trade-o : PhysImpute achieves top performing pointwise accuracy (best RMSE on three datasets and close second on one) while yielding stronger physical consistency than interpolation-driven baselines. These results indicate that PhysImpute is a practical approach for integrating physical priors into database cleaning pipelines for digital twin applications.