Past rewinding of fluid dynamics from noisy observation via physics-informed neural computing

Area of Science:

• Fluid dynamics
• Computational physics
• Machine learning

Background:

• Reconstructing fluid past is ill-posed due to noise and resolution limits.
• Traditional methods are unstable and diverge with noisy, distorted observations.
• Existing inverse problem techniques lack robustness against observational errors in time-reversed simulations.

Purpose of the Study:

• To present a novel method for robust time-reversed fluid simulation.
• To overcome the instability and error sensitivity of traditional fluid reconstruction techniques.
• To enable accurate reconstruction of fluid behavior from noisy and incomplete data.

Main Methods:

• Utilizing physics-informed neural computing (PINC).
• Developing a PINC approach that satisfies physics and observations while tolerating errors.
• Applying the method to extreme fluid scenarios.

Main Results:

• Achieved robust time-reversed fluid simulation using PINC.
• Demonstrated successful reconstruction of fluid past from noisy observations.
• Showcased time rewinding in scenarios including shock, instability, blast, and magnetohydrodynamic vortex.

Conclusions:

• PINC offers a robust solution for time-reversed fluid dynamics.
• The method accurately reconstructs the most probable past from imperfect observations.
• Potential applications include tracing interstellar evolution and understanding fusion plasma instabilities.