acceptodds
Withdrawn from ICLR 2027

Residual Geometry Shapes Downstream Correction Work for Learned Warm Starts in Nonsmooth PDE Solvers

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

10 views

Abstract

A learned warm start hands a solver a proposal that needs correction. We study how the residual metric used in training changes that downstream correction work in nonsmooth PDEs. In two-phase enthalpy diffusion, a network proposes a window of implicit steps, a linear decoder enforces balance and semismooth Newton corrects them; only the residual weighting changes. Branch structure selects the candidate metric: for phase-change diffusion the inverse stiffness K⁻¹ is, up to scale, the only quadratic metric under which no fixed-branch step amplifies differences, whereas a reaction term admits the identity but not K⁻¹. K⁻¹ weighting cuts Newton outer iterations by 37% at 4,000 updates, and the advantage persists at the longest tested budget. All four prospective placement tests where both metrics accept every request favor the predicted metric, and a preregistered obstacle test favors Euclidean weighting. Residual size cannot stand in for correction work: starts with smaller residuals in every vector norm can require more exact Newton steps, and reference-assisted repair of wrong-phase sites closes most of the outer-work gap. K⁻¹-trained starts help Newton but not a fixed-point corrector; continued on solver-accepted states, the learned route beats tuned FISTA and a coarse-grid classical route, while a training-free temperature-form FISTA is faster still. These results support selecting candidate metrics from branch structure and judging them by the deployed solver's work.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
learned warm starts, residual metrics, semismooth Newton, nonsmooth PDEs, free boundary problems, PDE solvers, scientific machine learning

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_aPu1OpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The paper studies two-phase enthalpy diffusion and asks how the weighting of a residual-based training loss affects the work required by a numerical solver after a neural network provides an initial guess. Weighting the residual by the inverse stiffness matrix reduces the number of Newton iterations compared with Euclidean weighting in the main experiments. The authors motivate this choice through an analysis of fixed phase branches and test it against several alternative weightings. They also show that the preferred weighting depends on where the nonlinearity appears: for the reaction-term problems they study, Euclidean weighting performs better.

Critical strengths

The fixed-branch analysis provides a mathematical motivation for using the inverse stiffness matrix to weight the training residual.

Critical weaknesses

The presentation substantially impedes my ability to evaluate the paper’s claims. The title suggests a broadly applicable method, but the main analysis and experiments center on two-phase enthalpy diffusion and closely related problems. I also found it difficult to understand how each experiment supports the central claim. The motivation for individual comparisons is often brief, and the organization of the subsections makes it difficult to follow the authors' argument.

Questions for authors
  1. Which features of two-phase enthalpy diffusion make inverse-stiffness weighting a promising choice?
Reviewer_GEGjOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

The paper studies learned warm starts for semismooth Newton in two-phase enthalpy diffusion, a network proposes a window of implicit steps, a linear decoder enforces balance, and semismooth Newton corrects them; only the residual weighting changes. The learned route beats tuned FISTA and a coarse-grid classical route, while a training-free temperature-form FISTA is faster still.

Critical strengths

Across all the experiments, the comparisons are preregistered, and the controls are well chosen. Moreover, classical baselines that beat the proposed method are reported, and failed predictions are also reported (see Section 4.1 and Section 6.2, respectively).

The choice of the training loss from the branch structure is a nice idea. It is supported by a mathematical proof, including the converse and the incompatibility of K⁻¹ under reaction placement.

Critical weaknesses

The advantage mostly reflects faster training convergence, and a training-free temperature-form FISTA beats the full learned pipeline. Therefore, it’s unclear when a practitioner should actually use the proposed method.

The method is tested only on synthetic data at low resolution with a 163k-parameter network.

The presentation is really hard to follow; the introduction makes three separate claims, and the paper keeps jumping between theory and experiments. Moreover, the text itself is very dense.

Questions for authors

See weaknesses

Reviewer_yfE7OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The paper studies how the residual metric used in training affects learned warm starts for a nonsmooth PDE solver (two-phase enthalpy diffusion). A network predicts a window of time steps, and semismooth Newton corrects them. Only the loss metric is changed: Euclidean r⊤rr^\top r versus inverse-stiffness r⊤K−1rr^\top K^{-1} r. The authors prove that for nonlinear diffusion K−1K^{-1} is the only compatible quadratic metric. For a reaction term the identity is the better choice. They also show that a smaller residual does not always mean less solver work; the number of sites in the wrong phase matters more. In experiments, K−1K^{-1} cuts Newton iterations by 3737% at 44k training updates and by 7.57.5% at 128128k. The preregistered placement tests mostly confirm the predictions.

Critical strengths
  1. Only the loss metric changes; the network, decoder, solver and data stay fixed. The controls rule out simple explanations: scalar and diagonal rescaling give almost no gain, and sign-masked metrics with the same spectrum lose the whole effect. Many tests were preregistered, and negative results are reported too, including that a classical method is faster than the learned route.
  2. The paper explains why K−1K^{-1} helps for diffusion but not for reaction. It also proves that residual size does not order Newton work for nonsmooth problems. The wrong-phase repair experiment supports this, recovering 97.697.6% of the gap.
Critical weaknesses
  1. A training-free temperature-form FISTA is 21–22% faster than the learned route. There is no comparison with standard solvers for this problem class, such as monotone multigrid [1] or truncated nonsmooth Newton multigrid [2], whose iteration counts barely grow with grid size. Without them it is unclear whether the method has any practical advantage.

    [1] Kornhuber, R. (1994). Monotone multigrid methods for elliptic variational inequalities I. Numerische Mathematik, 69(2), 167-184.

    [2] Gräser, C., & Sander, O. (2019). Truncated nonsmooth Newton multigrid methods for block-separable minimization problems. IMA Journal of Numerical Analysis, 39(1), 454-481.

  2. The gain drops from 5555% at 1k updates to 7.57.5% at 128k, and with solution labels it is only 6.86.8%. Much of the benefit may come from faster optimization rather than a better warm start.

  3. All experiments are 2D, on synthetic data and small grids. Training in H−1H^{-1} or energy norms is already known. The theory only proposes a candidate metric: in mixed placement the preference switches at θ≈0.94–0.99\theta \approx 0.94–0.99, not at the predicted 0.6680.668.

Questions for authors

See the weaknesses

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.