Track papers across labs, topics, and trends.
Search papers, labs, and topics across Lattice.
The paper introduces Manifold Constraining based on Lie group (MCL), a plug-and-play module that enforces geometric inductive bias in neural operators by constraining latent space updates to manifolds parameterized by low-rank Lie algebras. This addresses instability issues in multi-layer iteration and long-horizon rollout of existing neural operators that arise from unconstrained Euclidean latent space updates violating geometric and conservation laws. Experiments on 1-D Burgers and 2-D Navier-Stokes equations demonstrate a 30-50% reduction in relative prediction error with only a 2.26% increase in parameters.
Neural operators get a 30-50% accuracy boost in long-term PDE prediction with a simple, geometrically-principled "plug-in" that barely increases parameter count.
Neural operators offer an effective framework for learning solutions of partial differential equations for many physical systems in a resolution-invariant and data-driven manner. Existing neural operators, however, often suffer from instability in multi-layer iteration and long-horizon rollout, which stems from the unconstrained Euclidean latent space updates that violate the geometric and conservation laws. To address this challenge, we propose to constrain manifolds with low-rank Lie algebra parameterization that performs group action updates on the latent representation. Our method, termed Manifold Constraining based on Lie group (MCL), acts as an efficient \emph{plug-and-play} module that enforces geometric inductive bias to existing neural operators. Extensive experiments on various partial differential equations, such as 1-D Burgers and 2-D Navier-Stokes, over a wide range of parameters and steps demonstrate that our method effectively lowers the relative prediction error by 30-50\% at the cost of 2.26\% of parameter increase. The results show that our approach provides a scalable solution for improving long-term prediction fidelity by addressing the principled geometric constraints absent in the neural operator updates.