Selected research prototypes and scientific machine learning visualizations.
HyperFlowNet is a spatio-temporal neural operator for forecasting the transient evolution of velocity, pressure, and temperature fields in a high-pressure-difference hydrogen pipeline. It performs autoregressive prediction on a fixed irregular mesh and combines sliding-history attention, graph-aware interactions, rollout curriculum learning, and hard no-slip boundary projection.
StructFieldNet reconstructs full nodal von Mises stress fields for wing structures from grouped thickness variables and nodal coordinates on an unstructured finite-element mesh. It combines coordinate-design multiplicative lifting, Physics-Attention, and an SVD-initialized basis-plus-residual decoder to preserve global trends and localized stress hotspots.