Flow Reconstruction

Reconstructing turbulent flows sparse sensor measurements.

Publications:

*Mo, Y., Magri, L., 2025. Reconstructing unsteady flows from sparse, noisy measurements with a physics-constrained convolutional neural network. Phys. Rev. Fluids 10, 034901. https://doi.org/10.1103/PhysRevFluids.10.034901

*Go to branch paper/flow-reconstruction-2d.

**Mo, Y., Magri, L., 2026. Reconstruction of three-dimensional turbulent flows from sparse and noisy planar measurements: a weight-sharing neural network approach. DCE, 7, e5. https://doi.org/10.1017/dce.2026.10038

**Go to branch paper/flow-reconstruction-3d.

Major changes

28 Mar 2025: The current config file is no longer compatible with the old training script starting from commit e60831b779e301e9a3b5f88c0d43d4e47ba1455a.

Stage 1

Reconstruct the 2D flow behind a triangular shape, only limited amount of data is made available in training of the network.

The flow is simulated with XCompact3D, where a triangular cylinder was placed between two slip walls. Streamwise and wall-normal velocities, together with pressure are recorded. Spanwise velocity is $0$ in the computational domain. Pressure measurement at the base of the triangle are used as input to the network. The networks training methods are derived from PISR, a method for physics-informed super-resolution by Kelshaw, Rigas and Magri (2022).

Example of the generated wake behind a triangular shape.

Stage 2

Reconstruct 2D Kologorov flow from noisy, sparse sensors.

The Kolmogorov flows are generated using KolSol. Also starting from pressure measurements, and using sparse velocity & pressure measurements as collocation points. White noise are added to the measurements before training.

An example of the mean of a 2D Kolmogorov flow.

Stage 3

Reconstruction of 3D Kolmogorv flow. Staring from slices of measurements taken from within the domain.