Surrogate modeling of fluid flow under different conditions using physics-informed Deep Operator Networks

We applied and evaluated a physics-informed Deep Operator Network (PI-DeepONet) for modeling incompressible two-dimensional steady flows under varying Reynolds numbers and inlet boundary conditions. By combining the generalization capability of DeepONets with the physical constraints imposed by physics-informed neural networks (PINNs), the framework enables flow field prediction without relying on labeled data. Two types of input variations are considered: parametric variation in Reynolds numbers and functional variation in inlet velocity profiles. The results show that PI-DeepONet successfully generalizes across both scenarios, accurately predicting velocity and pressure fields even for unseen configurations. Furthermore, we explored the impact of architectural design on performance and found that shared-network variants significantly reduce computational cost without sacrificing accuracy. These results highlight both the potential and limitations of PI-DeepONet as a practical surrogate modeling tool for scientific computing.