In my PINN-to-BWLer experiments, imposing Dirichlet values at the transfer boundary increased the interpolant’s residual by about 30 times. A correction intended to make the boundary right gave the next optimizer a worse starting problem.

The proposed pipeline was appealing: use a neural network to find an approximate PDE solution, transfer it into a barycentric interpolant, and finish at higher precision. Each representation could do the part it was good at. The experiments made the handoff itself the central problem.

Agreement in a plot is weak evidence

Suppose the transferred function differs from the original by ε sin(kx). The function error is at most ε. The first derivative error scales as εk, and the second derivative error as εk². A difference too small to notice visually can dominate a second-order PDE residual.

This example does not diagnose every observed failure. It shows why matching sampled values cannot certify a useful transfer. The downstream solver uses a differential operator, so the transfer must be evaluated in the quantities that operator amplifies.

Boundary projection adds another complication. Enforcing a boundary value changes a representation whose interior derivatives depend on its coefficients. A local-looking adjustment in the values can have a nonlocal consequence in the interpolant. “The boundary is now exact” is only one part of the numerical state.

Put the first solver back on the clock

On the tested well-posed one-dimensional cases, cold-start BWLer reached high precision quickly enough that the neural warm start added roughly five seconds. Some higher-degree and irregular-domain settings benefited. The transfer did not work generally, including the tested Burgers setting.

It would be misleading to begin timing after the PINN had already trained. The comparison needs neural training, transfer, boundary handling, and the final solve. It also needs classical solvers, which were faster and more accurate than the learned pipelines on the tested problems.

The conclusion is narrower than “hybrid solvers do not work.” It is that this handoff did not produce a general advantage under the tested conditions. The useful next question is which properties predict a successful transfer and whether identifying those properties costs less than the benefit.

Residual minimization has a representation floor

A small residual is meaningful only relative to the equation, boundary conditions, sampling, and representation. If the interpolant cannot express a sufficiently good solution, more optimization may push it toward a residual minimum that is not closer to the reference truth.

This creates several distinct failure sources: the PINN approximation, the interpolation step, the boundary correction, and the second optimizer. A final error number cannot tell us which one dominated. Saving intermediate states and comparing them to a reference lets the experiment isolate the damaging transition.

A sensible switching policy should therefore observe more than the neural training loss. Candidate signals include derivative error estimates, boundary error, the transferred residual, and an estimate of what the receiving representation can express. A plateau in one solver is not evidence that another solver can exploit its current state.

The handoff is its own algorithm

I would now treat transfer as a procedure with an explicit acceptance test. Construct a candidate interpolant, evaluate the downstream quantities, and reject the handoff if it destroys too much useful structure. Keep the cost of that test in the total run.

There may be a regime where the neural approximation helps precisely because the conventional solver has a difficult initialization problem. There may also be regimes where a cold classical solve wins so decisively that a transfer policy has no useful job. Mapping that boundary is more informative than presenting one favorable hybrid example.

The broad ambition of combining learned representations with numerical methods remains attractive to me. This experiment gives it a less forgiving requirement: demonstrate that the information the second method needs actually survives the change in representation.

Source and related notes