Network architecture follows coupling in multiphysics systems: single vs. multiple branches in DeepONet and...

Area of Science:

• Computational Physics
• Machine Learning for Science

Background:

• Surrogate models are crucial for real-time prediction of complex physical systems.
• Deep Operator Networks (DeepONets) show promise but their efficacy in multiphysics coupling is underexplored.

Purpose of the Study:

• To investigate whether neural operator architecture should mirror the strength of physical coupling.
• To compare single-branch and multi-branch DeepONet designs for modeling coupled physical systems.

Main Methods:

• Compared feedforward and recurrent single-branch vs. multi-branch DeepONets.
• Evaluated models on reaction-diffusion, thermo-electrical, and thermo-mechanical systems.
• Assessed performance based on accuracy and prediction speed.

Main Results:

• Single-branch networks outperformed multi-branch variants in tightly coupled systems.
• Multi-branch designs were more suitable for decoupled or single-physics problems.
• Trained surrogates achieved prediction speeds up to 1.8 × 10⁴ times faster than physics-based solvers.

Conclusions:

• Neural operator architecture design is critical for effective multiphysics modeling.
• Single-branch architectures promote shared representations beneficial for strong coupling.
• DeepONets offer significant speedups for complex physical system simulations.