A compact and efficient physics-informed architecture for reconstructing and predicting complex physical fields from...

Area of Science:

• Computational Fluid Dynamics
• Artificial Intelligence
• Scientific Machine Learning

Background:

• Traditional methods for characterizing complex physical fields are computationally expensive.
• Existing AI models for field reconstruction/forecasting require large datasets and complex architectures.
• There is a need for efficient AI models that can handle limited or low-accuracy data.

Purpose of the Study:

• To develop compact and computationally efficient physics-informed neural networks (PINNs).
• To demonstrate the capability of these PINNs in reconstructing and forecasting complex physical fields with multiscale spatiotemporal dynamics.
• To overcome the limitations of data-intensiveness and high computational cost associated with current AI models.

Main Methods:

• Introduction of compact and computationally efficient physics-informed neural networks (PINNs).
• Demonstration using laminar and turbulent flows as benchmarks, including lid-driven cavity flows and 3D Homogeneous Isotropic Turbulence (HIT).
• Evaluation of super-resolution, forecasting from temporal snapshots, and spatial domain extension capabilities.

Main Results:

• Accurate super-resolution of lid-driven cavity flows (Re=1000), reconstructing high-resolution solutions from coarse inputs.
• Successful recovery of fine-scale structures and preservation of physical properties in 3D HIT from inaccurate data.
• Effective forecasting from limited temporal data with controlled error growth and accurate turbulent kinetic energy prediction.

Conclusions:

• The proposed compact PINN architecture offers a unified, training-efficient approach for physics-informed generalization.
• The framework enables reconstruction and prediction of complex physical fields with multiscale dynamics from limited observations.
• Physically informed sampling enhances reconstruction fidelity and energy consistency in spatial domain extension tasks.