A new computational method for learning the behavior of multiscale partial differential equations reported lower errors than the tested alternatives on most benchmark problems. Called the Frame Kernel Method, or FKM, it had the lowest reported relative ℓ2 error on three of four standard PDE tests and three of four multiscale tests.
The study asks whether a native multiscale representation can support surrogate learning for PDE solvers while retaining a scale-by-scale breakdown of predicted outputs. The work is an arXiv preprint, version 2 dated 1 September 2026, so its results have not yet been presented as a journal publication.
A representation built around spatial scales
FKM is designed for problems whose solutions contain features at more than one spatial scale. It uses compactly supported radial basis functions, localized mathematical building blocks, on nested subsets of observation sites. Each level has its own center density and support radius, and the levels together form a finite redundant dictionary.
The learning step uses coefficients from the input frame as the kernel features. Predicted output coefficients are organized by frame level, allowing the result to be synthesized separately at each level. These are scale-indexed frame contributions, not independent frequency bands, because the frame is nonorthogonal.
The paper also gives a mathematical guarantee under stated assumptions. When the observation sites are distinct and the finest-level kernel is strictly positive definite, the frame spans the sampled data space. A thin QR construction then gives exact interpolation and a unique coefficient vector with the smallest possible ℓ2 norm.
The fastest rates are not yet explained
The synthetic tests used analytic and finitely smooth target functions. Decomposition sites and evaluation sites were sampled independently and uniformly, without a minimum-separation requirement. Under a fixed-density protocol, evaluation-site error fell rapidly as the number of sites increased.
For kernels labelled C2, C4 and C6, the observed discrete relative ℓ2 orders were about 5.3, 7.3 and 8.75. These values were numerically close to twice the baseline exponents, a pattern the paper presents as evidence consistent with doubled convergence.
That interpretation is not yet a theorem. The baseline theory is stated in continuous Sobolev norms and does not explain the faster rates seen for smooth targets. The proposed doubled-rate explanation requires uniform stability of the QR projectors and a localized doubled-source approximation estimate, but the paper states those requirements as a central conjecture rather than proved results.
The synthetic results varied across the tests. Analytic slopes were close to the conjectured doubled rates, C2 targets were faster than expected over the tested range, and the C6 kernel levelled off in three dimensions. Some of this behavior was interpreted as preasymptotic.
Lower errors on most benchmark problems
On the standard PDE suite, FKM had the lowest reported relative ℓ2 error on three of four problems. The strongest gains were reported for cavity flow, triangular Darcy flow and reaction-diffusion. For cavity flow, FKM's error was 4.06e-7, compared with 1.45e-4 for the vanilla kernel method, 1.89e-4 for Transolver and 2.19e-4 for Geo-FNO.
The exception was piecewise-constant Darcy flow. FKM's reported error was 0.0328, slightly above the vanilla kernel method's 0.0306 and above the other listed comparisons. The paper attributes this result to discontinuities that violate the smoothness assumptions behind the kernel-frame representation.
On the multiscale suite, FKM again had the lowest reported error on three of four problems. It performed best on laminar cylinder flow, vortex shedding and species transport. In compressible Navier-Stokes at Mach 1.0, it was competitive but not best, with a reported relative ℓ2 error of 0.4937, compared with 0.481 for Transolver, 0.541 for the vanilla kernel method and 0.581 for Geo-FNO.
Accuracy comes with a computational tradeoff
The design involves a choice between density and sparsity. In a density sweep, analytic evaluation-site error fell by more than five orders of magnitude as density increased. The study selected a design-matrix density of 20% as a balance among accuracy, sparsity and preservation of multiscale structure.
The main practical constraint is scale. The authors report that sparse QR currently limits the three-dimensional implementation to approximately O(105) points. They identify sketching, preconditioning and block QR as possible ways to improve scaling, while also listing a proof of the doubled-rate conjecture as future work.
An early result with open questions
The comparisons show where FKM performed best in the tested suites, but the unproved convergence explanation remains central to the paper's interpretation. The authors also leave open how reliably the method can scale beyond the reported three-dimensional point limit and how it will perform on problems that do not fit its smoothness assumptions.
The reported funding came from AFOSR LRIR grants FA9550-25-1-0042 and 24RXCOR005, together with an NSF Division of Mathematical Sciences award numbered 2505986.
Paper data and sources
Original title: The Frame Kernel Method for Multiscale Operator Learning
Authors: Branden Frieden, Ryan Whitehead, M. Keith Ballard et al.
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-25
DOI: Not available
Original paper · Full text