Cross-Domain Pretraining for Steady-State Neural CFD Surrogates
Cross-domain pretraining cut CFD surrogate errors with far less task-specific data.
The paper tests neural CFD surrogates across geometries, boundary conditions, and fidelities. Its pretrained cross-domain models beat training from scratch and transfers from domain-specific experts in zero- and few-shot settings. With finetuning, the authors report 2-3x lower errors at the same sample size, or the same error using 8x fewer samples. They also find that simply pooling steady-state datasets is enough to make the approach work. ArXiv · AI/CL/LG's note
The paper tests neural CFD surrogates across geometries, boundary conditions, and fidelities. Its pretrained cross-domain models beat training from scratch and transfers from domain-specific experts in zero- and few-shot settings. With finetuning, the authors report 2-3x lower errors at the same sample size, or the same error using 8x fewer samples. They also find that simply pooling steady-state datasets is enough to make the approach work. ArXiv · AI/CL/LG's note
score 4