arxiv: 2605.10417 · v1 · submitted 2026-05-11 · 🧮 math.NA · cs.NA

Recognition: 2 theorem links

· Lean Theorem

PCELM: Perturbation-Correction Extreme Learning Machine for the Stefan problem

Siyuan Lang, Wenjie Liu, Zhiyue Zhang

Pith reviewed 2026-05-12 05:28 UTC · model grok-4.3

classification 🧮 math.NA cs.NA

keywords Stefan problemextreme learning machineperturbation correctionconvex optimizationmoving boundaryphase changenumerical PDEneural network approximation

0 comments

The pith

PCELM converts nonconvex Stefan optimization into a convex subproblem via perturbation correction around an initial approximation.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Stefan problems involve moving boundaries and discontinuous material properties from phase changes, making the residual minimization nonconvex and prone to optimization plateaus when using randomized neural networks. The PCELM method first computes a moderate-accuracy basic solution by directly minimizing the original nonconvex residual. It then forms a correction term through a perturbation expansion centered on that basic solution, which linearizes the problem and yields a convex optimization task for the output coefficients. A convexity proof supports this transformation, and tests across one- and multi-dimensional, single- and multi-phase Stefan problems show the correction consistently raises relative L2 accuracy by 2-6 orders of magnitude.

Core claim

The PCELM framework obtains an initial moderate-accuracy approximation by minimizing the nonconvex residual of the Stefan problem, then derives a correction term from a first-order perturbation expansion around this approximation. The resulting subproblem for the correction coefficients is convex, admits efficient solution, and produces large accuracy gains while overcoming optimization difficulties inherent to the original nonconvex formulation.

What carries the argument

Perturbation expansion around the basic nonconvex solution, which linearizes the residual into a convex optimization problem for the output-layer coefficients of the extreme learning machine.

If this is right

The correction step reliably escapes optimization plateaus for both single- and multi-phase Stefan problems.
Accuracy gains of 2-6 orders hold across one- and higher-dimensional geometries.
The proven convexity of the correction subproblem guarantees fast, reliable solution of the linear system for output weights.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The two-step structure may transfer to other moving-boundary or free-boundary PDEs where an inexpensive initial guess can be obtained.
Perturbation corrections could serve as a general technique to convexify nonconvex residuals in physics-informed neural networks when a moderate starting point exists.
Theoretical convergence rates for the perturbation correction could be derived by quantifying the distance of the basic approximation from the true solution.

Load-bearing premise

The initial basic approximation must lie close enough to the true solution that the first-order perturbation expansion introduces only small truncation error.

What would settle it

Run the correction step on deliberately poor basic approximations and check whether accuracy fails to improve or the subproblem loses convexity in the reported test cases.

Figures

Figures reproduced from arXiv: 2605.10417 by Siyuan Lang, Wenjie Liu, Zhiyue Zhang.

**Figure 2.** Figure 2: Structure of the extreme learning machine. [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗

**Figure 3.** Figure 3: Numerical results for the one-dimensional one-phase Stefan problem: (a) Ab [PITH_FULL_IMAGE:figures/full_fig_p021_3.png] view at source ↗

**Figure 4.** Figure 4: Numerical results for the one-dimensional two-phase Stefan problem: (a) Ab [PITH_FULL_IMAGE:figures/full_fig_p023_4.png] view at source ↗

**Figure 5.** Figure 5: Numerical results for the two-dimensional one-phase Stefan problem: (a) Point [PITH_FULL_IMAGE:figures/full_fig_p026_5.png] view at source ↗

**Figure 6.** Figure 6: Numerical results for the two-dimensional Frank-sphere problem: (a) Point-wise [PITH_FULL_IMAGE:figures/full_fig_p029_6.png] view at source ↗

**Figure 7.** Figure 7: Numerical results for the three-dimensional Frank-sphere problem: (a) Point [PITH_FULL_IMAGE:figures/full_fig_p032_7.png] view at source ↗

read the original abstract

For Stefan problems, characterized by moving boundaries and discontinuous coefficients due to phase changes, the inherent nonconvexity of the objective functional frequently causes optimization difficulty in randomized neural network approximations; to address this, we propose a Perturbation-Correction Extreme Learning Machine (PCELM) framework, built upon the extreme learning machine framework. This method first establishes a basic approximation during an initialization step by minimizing the original nonconvex residual, typically achieving only moderate accuracy, and then, in a subsequent correction step, determines a correction term by solving a subproblem based on a perturbation expansion around this basic approximation, thereby transforming it into a convex optimization problem for the output coefficients that ensures rapid convergence. We further provide a rigorous a convexity analysis, demonstrating that PCELM method solves a convex sub-problem. Numerical experiments on various Stefan problems, including multi-phase and multi-dimensional Stefan problems, confirm that the proposed PCELM method successfully overcomes optimization plateaus, with the correction step consistently delivering a significant improvement of 2-6 orders of magnitude in the relative L2 accuracy.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

PCELM gives a workable two-step fix for nonconvex ELM residuals on Stefan problems and reports large accuracy lifts, but the perturbation step around a moderate-accuracy base solution leaves truncation error uncontrolled.

read the letter

The new piece here is the explicit perturbation-correction split: run ordinary ELM on the nonconvex residual to get a rough u0, then linearize around it to turn the correction into a convex problem in the output weights. That framing is not in the earlier ELM or Stefan papers they cite, and it directly targets the optimization stalls that come from the unknown interface and the jump in coefficients. The convexity argument is stated as rigorous, and the experiments cover multi-phase and multi-dimensional cases, which is more than a toy demonstration. The reported 2-6 order drop in relative L2 error after the correction step is the main empirical result, and it looks consistent across the tests they show. That is useful practical evidence even if the absolute errors are not yet at the level of specialized moving-mesh codes. The soft spot is exactly the one the stress-test flags. Because u0 is only moderately accurate by construction, any O(1) mismatch in interface location or jump height makes the neglected higher-order terms comparable to the target accuracy. The abstract gives no explicit bound on ||u0 - u*|| or any a-posteriori residual check that would guarantee the convex subproblem stays close to the original PDE. Without that, the convexity is real but the solution it produces may solve a different problem. The numerics would need to be re-checked with proper baseline comparisons and with the perturbation radius quantified. This is a methods paper aimed at people who already use randomized networks for PDEs and want a cheap way to stabilize the Stefan case. It has enough concrete construction and numerical support to deserve referee time; the main revision needed is a clearer error analysis around the linearization step. I would send it out rather than desk-reject.

Referee Report

2 major / 2 minor

Summary. The manuscript proposes the Perturbation-Correction Extreme Learning Machine (PCELM) for Stefan problems. A basic approximation is first obtained by minimizing the original nonconvex residual within the ELM framework (moderate accuracy only). A first-order perturbation expansion around this basic solution is then used to formulate a convex subproblem whose solution yields the correction to the output weights. The authors assert a rigorous convexity analysis for the subproblem and present numerical experiments on multi-phase and multi-dimensional Stefan problems claiming consistent 2-6 orders of magnitude improvement in relative L2 accuracy after the correction step.

Significance. If the truncation error of the perturbation expansion can be rigorously controlled and the reported accuracy gains survive proper baseline comparisons, the method would provide a useful route to reliable convex optimization within randomized neural-network solvers for free-boundary problems with discontinuous coefficients. The explicit two-step construction and the convexity result are potentially valuable contributions to the ELM literature for nonconvex PDE residuals.

major comments (2)

[Abstract / PCELM construction] Abstract and method description: the first-order perturbation expansion around the moderate-accuracy basic approximation u0 is asserted to produce a convex subproblem whose solution delivers 2-6 orders of accuracy gain, yet no explicit bound on ||u0 - u*|| (or on the interface deviation) is supplied to guarantee that the neglected higher-order terms remain smaller than the target accuracy. For Stefan problems the residual contains discontinuous coefficients across an unknown moving interface; without such a bound the convex correction may solve a different problem than the original PDE, undermining the central claim.
[Numerical experiments] Numerical experiments: the abstract states that the correction step consistently improves relative L2 accuracy by 2-6 orders, but the manuscript supplies neither tables reporting the accuracy of the basic approximation alone nor comparisons against standard ELM baselines or other convexification techniques. This makes it impossible to isolate the contribution of the perturbation step or to rule out that the gains arise from re-optimization rather than the claimed mechanism.

minor comments (2)

[Abstract] Abstract contains the phrase 'rigorous a convexity analysis'; this should be corrected to 'a rigorous convexity analysis'.
[Method] The explicit algebraic form of the first-order perturbation expansion and the resulting convex objective functional should be displayed as an equation for clarity.

Simulated Author's Rebuttal

2 responses · 1 unresolved

We thank the referee for the constructive and detailed comments. We respond to each major comment below and indicate the revisions we will make to address them.

read point-by-point responses

Referee: [Abstract / PCELM construction] Abstract and method description: the first-order perturbation expansion around the moderate-accuracy basic approximation u0 is asserted to produce a convex subproblem whose solution delivers 2-6 orders of accuracy gain, yet no explicit bound on ||u0 - u*|| (or on the interface deviation) is supplied to guarantee that the neglected higher-order terms remain smaller than the target accuracy. For Stefan problems the residual contains discontinuous coefficients across an unknown moving interface; without such a bound the convex correction may solve a different problem than the original PDE, undermining the central claim.

Authors: We appreciate the referee pointing out the absence of an explicit a priori bound on the error of the basic approximation. The manuscript proves convexity of the derived subproblem but does not supply a rigorous bound on ||u0 - u*|| or interface deviation that would guarantee the neglected higher-order terms are controlled for arbitrary Stefan problems. This is a genuine limitation of the current analysis. In the revision we will add a dedicated paragraph in Section 3 discussing the practical size of the correction term (supported by additional numerical diagnostics) and the conditions under which the first-order model is expected to remain valid, while explicitly noting that a fully rigorous truncation-error bound for moving interfaces with discontinuous coefficients is left for future work. revision: partial
Referee: [Numerical experiments] Numerical experiments: the abstract states that the correction step consistently improves relative L2 accuracy by 2-6 orders, but the manuscript supplies neither tables reporting the accuracy of the basic approximation alone nor comparisons against standard ELM baselines or other convexification techniques. This makes it impossible to isolate the contribution of the perturbation step or to rule out that the gains arise from re-optimization rather than the claimed mechanism.

Authors: We agree that the current experimental section does not isolate the effect of the correction step sufficiently. In the revised manuscript we will insert new tables (in the numerical section) that explicitly report the relative L2 error of the basic ELM approximation before correction for every example. We will also add direct comparisons against a standard ELM solver applied to the identical nonconvex residual, confirming that the basic step alone typically plateaus at moderate accuracy while the perturbation correction produces the reported gains. A short discussion of related convexification strategies from the ELM literature will be added to the introduction for context. revision: yes

standing simulated objections not resolved

A rigorous a priori bound on ||u0 - u*|| (or interface deviation) that guarantees control of higher-order perturbation terms for general Stefan problems with unknown moving interfaces and discontinuous coefficients.

Circularity Check

0 steps flagged

No significant circularity; derivation remains self-contained

full rationale

The PCELM construction begins with a basic approximation obtained by direct minimization of the original nonconvex residual, followed by an explicit first-order perturbation expansion that produces a separate convex subproblem in the output weights. Numerical accuracy gains of 2-6 orders are reported from independent experiments on multi-phase and multi-dimensional Stefan problems rather than being recovered by construction from the same fitted quantities. No load-bearing self-citation, self-definitional loop, or renaming of a known result appears in the provided derivation chain; the convexity claim is asserted via a separate analysis whose validity is external to the input data or prior fitted values.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 0 invented entities

The method relies on standard ELM random-feature assumptions and the validity of a first-order perturbation expansion around an approximate solution; no new physical entities are introduced.

axioms (2)

domain assumption The initial ELM approximation obtained by minimizing the nonconvex residual is close enough to the true solution for the perturbation series to be truncated after the linear term while preserving accuracy.
Invoked in the description of the correction step.
standard math The subproblem obtained after the perturbation expansion is convex in the output coefficients.
Claimed to be demonstrated by rigorous analysis.

pith-pipeline@v0.9.0 · 5487 in / 1402 out tokens · 67318 ms · 2026-05-12T05:28:44.382354+00:00 · methodology

discussion (0)

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

IndisputableMonolith/Cost/FunctionalEquation.lean washburn_uniqueness_aczel unclear
This method first establishes a basic approximation during an initialization step by minimizing the original nonconvex residual, typically achieving only moderate accuracy, and then, in a subsequent correction step, determines a correction term by solving a subproblem based on a perturbation expansion around this basic approximation, thereby transforming it into a convex optimization problem for the output coefficients
IndisputableMonolith/Foundation/BranchSelection.lean branch_selection unclear
We further provide a rigorous a convexity analysis, demonstrating that PCELM method solves a convex sub-problem.

Reference graph

Works this paper leans on

45 extracted references · 45 canonical work pages

[1]

Kumar, V

S. Kumar, V. K. Katiyar, Mathematical modeling of freezing and thaw- ing process in tissues: a porous media approach, Int. J. Appl. Mech. 2 (2010) 617-633

work page 2010
[2]

J. A. Sethian, J. Straint, Crystal growth and dendritic solidification, J. Comput. Phys. 98 (1992) 231-253

work page 1992
[3]

W. J. Boettinger, J. A. Warren, C. Beckermann, et al, Phase-field sim- ulation of solidification, Annual review of materials research 32 (2002) 163-194

work page 2002
[4]

Kumar, A Stefan problem with moving phase change material, vari- ablethermalconductivityandperiodicboundarycondition, Appl.Math

A. Kumar, A Stefan problem with moving phase change material, vari- ablethermalconductivityandperiodicboundarycondition, Appl.Math. Comput. 386 (2020) 125490

work page 2020
[5]

Rubinshte˘ ın, The Stefan Problem, American Mathematical Soc, 1971

L. Rubinshte˘ ın, The Stefan Problem, American Mathematical Soc, 1971

work page 1971
[6]

E. I. Hanzawa, Classical solutions of the Stefan problem, Tohoku Math. J., Second Series 33 (1981) 297–335

work page 1981
[7]

D. E. Womble, A front-tracking method for multiphase free boundary problems, SIAM J. Numer. Anal. 26 (1989) 380-396

work page 1989
[8]

S. O. Unverdi, G. Tryggvason, A front-tracking method for viscous, in- compressible, multi-fluid flows, J. Comput. Phys. 100 (1992) 25-37

work page 1992
[9]

M.Muradoglu, G.Tryggvason, Afront-trackingmethodforcomputation of interfacial flows with soluble surfactants, J. Comput. Phys. 227 (2008) 2238-2262

work page 2008
[10]

Juric, G

D. Juric, G. Tryggvason, A front-tracking method for dendritic solidifi- cation, J. Comput. Phys. 123 (1996) 127-148

work page 1996
[11]

J.Glimm, J.W.Grove, X.L.Li., etal, Three-dimensionalfronttracking, SIAM J. Sci. Comput. 19 (1998) 703-727

work page 1998
[12]

Leung, H

S. Leung, H. Zhao, A grid based particle method for moving interface problems, J. Comput. Phys. 228 (2009) 2993-3024. 33

work page 2009
[13]

H. S. Udaykumar, R. Mittal, P. Rampunggoon, et al, A sharp inter- face Cartesian grid method for simulating flows with complex moving boundaries, J. Comput. Phys. 174 (2001) 345-380

work page 2001
[14]

Voller, M

V. Voller, M. Cross, Accurate solutions of moving boundary problems using the enthalpy method, Int. J. Heat. Mass Transf. 24 (1981) 545-556

work page 1981
[15]

V. R. Voller, J. B. Swenson, W. Kim, et al, An enthalpy method for mov- ing boundary problems on the earth’s surface, Int. J. Numer. Methods. Heat. Fluid Flow. 16 (2006) 641-654

work page 2006
[16]

C. R. Swaminathan, V. R. Voller, A general enthalpy method for model- ing solidification processes, Metallurgical transactions B 23 (1992) 651- 664

work page 1992
[17]

S. Chen, B. Merriman, S. Osher, et al, A simple level set method for solving Stefan problems, J. Comput. Phys. 135 (1997) 8-29

work page 1997
[18]

Olsson, G

E. Olsson, G. Kreiss, A conservative level set method for two phase flow, J. Comput. Phys. 210 (2005) 225-246

work page 2005
[19]

D. Peng, B. Merriman, S. Osher, et al, A PDE-based fast local level set method, J. Comput. Phys. 155 (1999) 410-438

work page 1999
[20]

G. J. Fix, Phase field methods for free boundary problems, 1982

work page 1982
[21]

J. A. Mackenzie, M. L. Robertson, A moving mesh method for the solu- tion of the one-dimensional phase-field equations, J. Comput. Phys. 181 (2002) 526–544

work page 2002
[22]

J. T. Lin, The numerical analysis of a phase field model in moving boundary problems, SIAM J. Numer. Anal. 25 (1998) 1015-1031

work page 1998
[23]

Z. Zhou, W. Jiang, T. Qian, et al, A new phase-field model for anisotropic surface diffusion: anisotropic Cahn–Hilliard equation with improved conservation, Proceedings of the Royal Society A: Math. Phys. Eng. Sci. 481 (2025)

work page 2025
[24]

Cybenko, Approximation by superpositions of a sigmoidal function, Math

G. Cybenko, Approximation by superpositions of a sigmoidal function, Math. Control. Signals. Systems. 2 (1989) 303-314. 34

work page 1989
[25]

Raissi, P

M. Raissi, P. Perdikaris, G. E. Karniadakis, Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations, J. Comput. Phys. 378 (2019) 686–707

work page 2019
[26]

S. Wang, P. Perdikaris, Deep learning of free boundary and Stefan prob- lems, J. Comput. Phys. 428 (2021) 109914

work page 2021
[27]

J. Li, W. Wu, X. Feng, Improved physics-informed neural networks com- bined with small sample learning to solve two-dimensional Stefan prob- lem, Entropy 25 (2023) 675

work page 2023
[28]

B. E. Madir, F. Luddens, C. Lothodé, I. Danaila, Physics informed neural networks for heat conduction with phase change, International Journal of Heat and Mass Transfer 252 (2025) 127430

work page 2025
[29]

L. A. Larios-Cardenas, F. Gibou, Error-correcting neural networks for semi-Lagrangian advection in the level-set method, J. Comput. Phys. 471 (2022) 111623

work page 2022
[30]

Shkolnikov, H

M. Shkolnikov, H. M. Soner, V. Tissot-Daguette, Deep level-set method for Stefan problems, J. Comput. Phys. 503 (2024) 112828

work page 2024
[31]

J. Chen, X. Chi, Z. Yang, et al, Bridging traditional and ma- chine learning-based algorithms for solving PDEs: the random feature method, J. Mach. Learn. 1 (2022) 268–298

work page 2022
[32]

X. Chi, J. Chen, Z. Yang, The random feature method for solving inter- face problems, Comput. Methods Appl. Mech. Eng. 420 (2024) 116719

work page 2024
[33]

Huang, Q

G. Huang, Q. Zhu, C. Siew, Extreme learning machine: theory and applications, Neurocomputing 70 (2006) 489-501

work page 2006
[34]

Dwivedi, B

V. Dwivedi, B. Srinivasan, Physics informed extreme learning machine (PIELM)–a rapid method for the numer ical solution of partial differen- tial equations, Neurocomputing 391 (2020) 96–118

work page 2020
[35]

S. Lang, Z. Zhang, A Perturbation-Correction Method Based on Lo- cal Randomized Neural Networks for Quasi-Linear Interface Problems, ArXiv Preprint ArXiv:2602.05800 (2026). 35

work page arXiv 2026
[36]

S. Dong, Z. Li, Local extreme learning machines and domain decom- position for solving linear and nonlinear partial differential equations, Comput. Methods Appl. Mech. Eng. 387 (2021) 114129

work page 2021
[37]

F. Ren, P. Zhuang, X. Chen, H. Yu, H. Yang, Physics-informed extreme learningmachine(PIELM)forStefanproblems, Comput.MethodsAppl. Mech. Eng. 441 (2025) 118015

work page 2025
[38]

Chang, T

C. Chang, T. Lin, M. Lai, Physics-informed machine learning for two-phase moving-interface and Stefan problems, ArXiv Preprint ArXiv:2512.14010 (2025)

work page arXiv 2025
[39]

Methods Appl

W.Hu, Y.Shih, T.Lin, etal, Ashallowphysics-informedneuralnetwork for solving partial differential equations on static and evolving surfaces, Comput. Methods Appl. Mech. Eng. 418 (2024) 116486

work page 2024
[40]

A. Lin, Z. Zhang, W. Zhao, et al, Discontinuous extreme learning ma- chine for interface and free boundary problems, J. Comput. Phys. 2025 114329

work page 2025
[41]

S. Lang, Z. Zhang, A Predictor Corrector Convex Splitting Method for Stefan Problems Based on Extreme Learning Machines, ArXiv Preprint ArXiv:2601.16655 (2026)

work page arXiv 2026
[42]

Nocedal, S

J. Nocedal, S. Wright, Numerical optimization. New York, NY: Springer New York, 2006

work page 2006
[43]

R. M. Furzeland, A comparative study of numerical methods for moving boundary problems, IMA J. Appl. Math. 26 (1980) 411–429

work page 1980
[44]

B. T. Johansson, D. Lesnic, T. Reeve, A meshless method for an inverse two-phase one-dimensional linear Stefan problem, Inverse Probl. Sci. Eng. 21 (2013) 17–33

work page 2013
[45]

Colton, R

D. Colton, R. Reemtsen, The numerical solution of the inverse Ste- fan problem in two space variables, SIAM. J. Appl. Math. 44 (1984) 996–1013. 36

work page 1984