FE reduced-order model-informed neural operator for structural dynamic response prediction.

Area of Science:

• Structural Dynamics
• Computational Mechanics
• Machine Learning

Background:

• Physics-Informed Neural Networks (PINN) show promise for differential equations but face accuracy and efficiency challenges in structural dynamics.
• Directly embedding large structural models as constraints in neural networks hinders trainability and precision.

Purpose of the Study:

• To introduce a novel Fourier Neural Operator (FNO)-based method, FRINO, for high-precision, low-cost, and versatile structural dynamic response prediction.
• To overcome the limitations of PINNs in handling complex structural dynamic models.

Main Methods:

• Employed Fourier Neural Operator (FNO) to capture frequency-domain features of structural dynamics.
• Integrated a reduced-order model (ROM) via proper orthogonal decomposition to enforce physical constraints and reduce computational cost.
• Validated the FRINO method using cantilever beam dynamic response prediction under various excitations.

Main Results:

• FRINO accurately predicts structural dynamic responses and inherent dynamic characteristics.
• Achieved prediction accuracy up to two orders of magnitude higher than PINNs.
• Demonstrated computation speed enhancement of up to three orders of magnitude compared to PINNs.
• FRINO shows broad versatility for predicting responses under diverse unknown excitations.

Conclusions:

• FRINO offers a significant advancement over PINNs for structural dynamic response prediction.
• The method provides high precision, computational efficiency, and versatility.
• Optimal FRINO performance requires careful consideration of physical loss, data resolution, and network architecture.