Skip to main content
Back to timeline
arXivSource publication:

FDM+ drops 2D radiative transport control below 3% error on two heterogeneous media, without repeated forward solves

Synopsis

Constrained control of the 2D radiative transport equation has relied on repeatedly solving the forward problem, which is costly and leaves multiple solutions unaddressed. FDM+ extends the frozen diffusion model to two dimensions, using decoupled prediction to generate the low-dimensional boundary control and the high-dimensional flux field separately, and parameterizing the control by a small set of nodes. On a layered strong-absorption medium and a maze medium, FDM+ keeps reconstruction error below 3%, while compared methods are several times higher on some cases.

Source-provided article image: FDM+: A Framework for Constrained Control of the 2D Radiative Transport Equation
arXiv

Interpretation

Constrained control of 2D radiative transport can now be produced directly by a generative model that outputs the boundary control, instead of calling a high-fidelity solver for every target state. The earlier frozen diffusion model was validated only on one-dimensional radiative transport; moving to two dimensions runs into a dimension mismatch between control and state field, a resolution gap between solver time steps and diffusion sequence length, and non-unique solutions. On the central strong-absorption layered medium, FDM+ gives relative errors of 0.97%, 0.80% and 0.92% across three cases, against 2.45%, 3.10% and 2.78% for DiffPhyCon, 5.86%, 7.94% and 6.12% for FNO, and 4.88%, 6.54% and 7.35% for DeepONet.

Two parallel output pathways handle the high-dimensional flux field and the low-dimensional boundary control separately, with the flux field still predicting noise and the control predicted directly as a clean value. Expanding the control across spatial dimensions and predicting noise jointly, as the standard route does, contaminates the low-dimensional control signal with spatial noise; coupled prediction exceeds 100% error in the ablation. In the ablation setting with a raised initial temperature that makes an unparameterized baseline feasible, decoupled plus parameterized reaches 0.95%, 0.64%, 0.35% and 1.54% on four cases, decoupled plus unparameterized reaches 1.97%, 0.82%, 0.58% and 3.43%, and coupled plus parameterized reaches 145.60%, 15.20%, 30.80% and 295.10%.

The control function is no longer generated step by step; the model learns a small set of low-dimensional nodes, and piecewise-linear interpolation restores the high-resolution boundary input the solver needs. At low temperature the high-fidelity solver needs thousands of adaptive time steps, so generating a control sequence of the same length is computationally infeasible. A node-count scan shows reconstruction error falling and then rising with node count, and the authors combine interpolation error and distribution-estimation error into one total error bound, noting that the dimension-dependent factor increases monotonically with node count but that this does not establish monotonicity of the total error or the existence of an optimal node count.

Perspective

The demonstrated setting is the 2D grey radiative transport equation with control imposed on the top boundary, for two heterogeneous configurations, a layered strong-absorption medium and a maze medium, both with known and stationary media; training and test data come from a high-fidelity solver based on the diamond-difference scheme. The authors propose extending the framework to dynamic or uncertain media and to 3D anisotropic scattering.

The dimension-dependent factor in the total error bound increases monotonically with node count, but the authors state explicitly that this does not establish monotonicity of the total error or the existence of an optimal node count; the turn where error falls and then rises in the node-count scan is a numerical observation. The multi-solution experiment switches to a 2D nonlinear diffusion equation as a representative model in the diffusion limit of radiative transport, and the 50 generated candidate sources match the target within the constraint region to about 1% error.