Track papers across labs, topics, and trends.
Search papers, labs, and topics across Lattice.
This paper introduces the Green's Observation Operator (GObO) to efficiently learn dynamics governed by ambient PDEs but observed on lower-dimensional submanifolds. GObO maps the ambient medium to a Green's kernel, enabling new source predictions via a single lower-dimensional integral, bypassing expensive full-domain solvers and static black-box surrogates. The method achieves exponential convergence for exact finite streaming states, demonstrating significant zero-shot prediction accuracy improvements over baselines for 3D heat conduction and advection-diffusion problems.
Predict PDE dynamics on submanifolds with zero-shot generalization to moving sources and mild nonlinearities, achieving 4-8x lower error than surrogates at millisecond query times.
Many physical systems are driven and observed only on lower-dimensional submanifolds of a larger spatial domain, while their dynamics are governed by the ambient medium occupying that domain. Examples include laser-heated parts imaged by an infrared camera, and ground-level emissions measured on a sensor plane. Full-domain solvers, however, compute the entire volume for every new source although only the observation submanifold is needed, and black-box surrogates do not exploit that the ambient medium remains fixed. We introduce the \emph{Green's Observation Operator (GObO)}, which maps the ambient medium once to the Green's kernel of a linear PDE restricted to the source and observation submanifolds. New sources then cost one lower-dimensional integral and no network evaluation. Exponential rates in the kernel yield an exact finite streaming state with horizon-independent memory; we prove its stability and an approximation rate for the restricted heat kernel. On three-dimensional heat conduction and advection--diffusion with collocated and distinct source and observation geometries, GObO trained on static sources predicts responses to moving sources zero-shot with 4--8$\times$ lower error than black-box surrogates, at 1.4\,ms per query after a single conditioning pass. The same kernel transfers across resolutions and admits corrections for mild nonlinearities, including radiative losses and temperature-dependent conductivity, without retraining, at the cost of lower in-distribution accuracy.