Learnable composition for neural operators

The paper introduces LatentDDM, a neural operator framework that first pretrains on small subdomains and then adapts to new settings by training only a lightweight composition module. Experiments on steady Darcy flow and unsteady airfoil flow show that this approach reduces error by 36‑56% on larger domains and improves 20‑step rollouts, outperforming capacity‑matched full‑domain models. The study highlights co‑designed local pretraining and composition‑level transfer as a promising design principle for physical foundation models.

Machine-generated by The Flow from the publisher's headline and feed description — not written or checked by a human. The full article lives at arXiv Machine Learning.