acceptodds
Withdrawn from ICLR 2027

Decompose-and-Superpose: A Superposition-Informed Neural Network for Steady-State Temperature Field Prediction

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

The finite element method (FEM) is the standard tool for steady-state temperature field simulation, but every change of a design parameter – thermal conductivity, convective heat transfer coefficient, or domain size – requires a new mesh and a new matrix solve, which is prohibitive in parametric thermal design and in real-time thermal management. This paper proposes a superposition-informed neural network (SINN) for fast steady-state temperature field prediction. Starting from the linear operator structure of the steady-state heat conduction equation with Robin boundary conditions, the control parameters are split into operator parameters, which alter the operator itself (thermal conductivity, convective heat transfer coefficient, geometric dimensions), and load parameters, which act linearly on the solution once the operator is fixed (heat source strength, ambient reference temperature, imposed boundary values). The temperature field is accordingly decomposed into one source-driven base subfield and several boundary-response subfields; neural networks learn only the non-superposable mapping from operator parameters to these subfields, and the field under an arbitrary load is reconstructed analytically from them without further network evaluation. On a two-dimensional steady-state heat conduction dataset covering 9 boundary condition combinations and 52 rectangular geometries, the method attains an RMSE of 1.94 mK and is 229–980 times faster than COMSOL FEM at inference. Because the field under an arbitrary load is assembled from the learned subfields rather than re-predicted, load conditions absent from the training data are reconstructed at the same accuracy, without retraining and without any further network evaluation. Because the operator/load split relies only on the linear structure of the governing equation, the framework extends to other steady-state linear diffusion problems; the conditions for this transfer are analyzed in Section 6. The source code is provided as supplementary material.

Primary area:
neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)
Keywords:
Decompose-and-Superpose: A Superposition-Informed Neural Network for Steady-State Temperature Field Prediction

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 3 reviews, above 0% of the others.
Reviewer_6UK5OpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes a Superposition-Informed Neural Network (SINN) for predicting two-dimensional steady-state temperature fields. Its central idea is to distinguish parameters that change the differential or boundary operator—such as thermal conductivity, convective heat-transfer coefficients, and geometry—from load parameters that enter linearly when the operator is fixed. The method trains SourceNet to predict the response to a volumetric heat source and BoundaryNet to predict individual boundary responses. Complete temperature fields are reconstructed through analytical linear combinations of these predicted components. Experiments consider rectangular domains with two adjacent controllable boundaries, each assigned Dirichlet, Neumann, or Robin conditions, while the remaining boundaries are held at a reference temperature. The dataset covers nine boundary-condition combinations and 52 rectangular geometries. The authors report an overall RMSE of 1.944 mK, approximately 31% lower than DirectNet; an RMSE of 1.46 mK on an unseen combination obtained by removing one boundary excitation; and a 229-fold single-sample GPU speedup over their COMSOL workflow. In my assessment, the experiments demonstrate the feasibility of learning and recombining basis responses within a restricted load space and operator-parameter range. They do not yet establish a sufficiently distinct methodological contribution or adequately substantiate the broader accuracy and efficiency claims.

Critical strengths
  1. Explicitly exploiting the known linear load structure has practical value.
  2. The restricted experimental study includes informative diagnostics and acknowledges an important baseline limitation.
Critical weaknesses
  1. The methodological novelty is insufficiently established. Separating operator parameters from linearly entering loads follows directly from the structure of linear PDEs. The manuscript does not adequately distinguish its contribution from classical basis-response methods, Green's-function learning, or parametric reduced-order modeling. Without a substantive comparison, the proposed framework appears primarily to combine established superposition principles with a conventional neural surrogate.

  2. The SINN--DirectNet comparison does not isolate the benefit of decomposition. SINN uses approximately 4040 million parameters across two networks, compared with 20.0520.05 million for DirectNet, and receives substantially more component-level supervision. Consequently, the reported 31%31\% RMSE improvement cannot be attributed specifically to decomposition. Comparisons controlling total capacity, supervision, and training budget are needed to substantiate this central claim.

  3. The unseen-load experiment evaluates only a restricted recombination, obtained by switching one boundary excitation off. DirectNet's corresponding input state receives no training signal, limiting the comparison's informativeness. Moreover, superposition does not guarantee unchanged absolute error: ∥u^−u∥≤∑j∣aj∣∥ej∥\|\widehat{u}-u\|\leq\sum_j |a_j|\|e_j\|. The claim of maintaining the same accuracy under arbitrary loads is therefore stronger than the presented evidence supports.

  4. The physical justification for aspect-ratio-based geometry deduplication is incomplete. Under Robin conditions, normalized boundary equations depend on quantities such as hW/khW/k and hH/khH/k, which change when geometrically similar rectangles are rescaled. Prescribed nonzero fluxes also introduce size dependence. Equal aspect ratios therefore do not generally imply equivalent temperature fields, and the consequences for dataset construction remain insufficiently explained.

  5. The reported millikelvin errors measure agreement with interpolated COMSOL labels, but the manuscript provides insufficient mesh-convergence, solver-tolerance, and interpolation-error evidence to establish comparable physical accuracy. Boundary interpolation is particularly relevant because the largest errors occur near corners. The negative temperature rise of −0.26 K-0.26\,\mathrm{K} in Table 14 also requires clarification regarding flux conventions, numerical artifacts, or preprocessing.

  6. The efficiency comparison does not adequately consider computational reuse. Fixed-geometry parameter changes do not inherently require remeshing, and fixed-operator load changes can exploit cached factorizations or precomputed basis responses. The reported speedups therefore characterize the selected COMSOL workflow rather than establish an advantage over appropriately reusable numerical baselines. Missing hardware specifications, mesh sizes, and offline costs further limit interpretation.

  7. The independence of the evaluation data is insufficiently documented. The manuscript does not clearly establish synchronized splits across SourceNet, both BoundaryNet directions, and DirectNet at the physical-configuration level. This leaves component-level overlap between training and final validation unverified. Furthermore, random splits over a discrete parameter grid primarily demonstrate interpolation, without establishing generalization to held-out geometry families or operator-parameter extrapolation.

Questions for authors
  1. Does SINN retain its accuracy advantage when total parameter count, training budget, and available supervision are matched? Please include a DirectNet baseline trained with all input states represented and with load combinations generated from the same component fields. Which remaining improvement can then be attributed specifically to the decomposition strategy?

  2. Were geometries sharing an aspect ratio simulated independently, or were their fields reused? How does the deduplication procedure account for hW/khW/k, hH/khH/k, and prescribed-flux scaling? Please also clarify whether the boundary operator implemented in the simulations follows Equation (2) or the inconsistent sign convention stated immediately afterward in Section 2.1.

  3. Were all datasets partitioned using a common grouping by physical parameter configuration? For each final validation case, were its source response and both boundary responses excluded from the relevant training sets? Please provide the grouping rule and clarify whether the reported results use an independent test set or the validation set used during training.

  4. What mesh-convergence, solver-tolerance, and interpolation checks establish reference accuracy below the reported millikelvin errors? What causes the negative temperature rise in Table 14? Please also report how absolute and relative prediction errors change as load coefficients vary continuously, exceed their reference amplitudes, or produce cancellation between the component responses.

  5. What speedups remain when FEM reuses meshes, matrix factorizations, or precomputed basis responses where applicable? Please distinguish new-operator queries from repeated-load queries under a fixed operator, and report hardware specifications, mesh degrees of freedom, and timing boundaries. What are the total data-generation and training costs for the complete SINN system?

Reviewer_9RB9OpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes a decomposition approach towards modelling PDE heat transfer through homogenous rectangular solids. The key proposed innovation is to separate the problem into operator (with linear scaling) and load/forcing/boundary (learnt) properties that are then superimposed. The results are shown to have good predictive power for unseen configurations e.g. different boundary types. This is shown to outperform a naive parameter-equivalent approach which does not perform decomposition/superposition.

Critical strengths
  • The results show good predictive power for different load configurations.
  • The work could show how neural surrogates could be used for superposition - the evidence is not strong enough however, nor has it been placed properly into context.
Critical weaknesses
  • There seems to be a significant lack of comparison against related techniques e.g. PINNs (and all their derivatives like FB-PINNS, XPINNS, PINO), FNOs, reduced basis methods etc. Whilst other networks such as DeepONet are mentioned in the discussion, it would have been useful to provide a direct comparison against existing methods. The naive baseline of DirectNet is not really physics governed or informed, so it is not surprising that this approach is superior.
  • The broad principle of decomposition into operator and load/forcing function is well established classically. For example, Ghnatios et al. (2012) "Proper Generalized Decomposition based dynamic data-driven control of thermal processes" tackles this very problem and demonstrates rapid simulation across parameter variations. A fundamental question really is what this neural approach gains over a more classical surrogate. At the very least, there needs to be an objective comparison.
  • The experimental setup is far too limited to make claims of generality. Whilst this is mentioned in Sec 6.1 (with a heading that sounds more like a marketing strap-line rather than a scientific paper), it would greatly help the strength of this paper if you were able to demonstrate that it does actually aid other types of networks. The comparison against both a naive simulation and COMSOL is not objective, given the point above about the massive literature on PINNs and their friends.
  • Similarly, the only experiment is a rectangular geometry for heat transfer using a homogenous media at steady state. For this to be truly demonstrated as a step change in accelerating the solution space of FEM, it is necessary to show that it yields benefits across many more varied PDE problems. Even with thermal diffusion, looking into more varied geometry and problems would show that the amortized cost of generating and training the decomposition yields transferrable benefits to out-of-domain settings.
Questions for authors
  • Can you show that your method does indeed improve the accuracy/generalization of competing PINN/FNO methods?
Reviewer_xF1qOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

The paper proposes SINN, which splits steady-state heat-conduction parameters into “operator parameters”and “load parameters”. Two subnetworks learn a source-driven base field and boundary-response subfields; fields under arbitrary loads are reconstructed by analytical superposition. The authors claim millikelvin-level accuracy, substantial speedup over FEM, and recombination of learned subfields on an “unseen load-state” test set.

Critical strengths

None. The engineering observation that load dependence can be handled analytically for fixed operators is reasonable, but it is a classical property of linear PDEs rather than a methodological contribution.

Critical weaknesses

W1. The operator/load parameter split is not a valid theoretical contribution. The criterion—where a parameter appears in the composite operator equation—is writing-dependent. Rescaling the governing equation can move an operator parameter into the source term; shifting the field variable can absorb the ambient reference into active boundary values; and under Robin conditions the load term itself depends on an operator parameter. The authors’ own case states that the conductivity scaling reduces to an exact amplitude factor and is therefore not scanned, directly contradicting the claim that it is an operator parameter that “must be learned.” A formal, representation-invariant definition, a superposition theorem, and explicit conditions are missing.

W2. The paper does not close the three problems raised in the introduction. The introduction criticizes poor out-of-domain generalization, exponential sample growth in parameter dimension, and black-box networks that ignore physical structure. SINN is validated only on a stratified in-domain split; the “unseen load” case is merely a recombination of components each seen in training. There is no operator-parameter extrapolation, unseen geometry, sparse-training sample-efficiency study, or operator-learning baseline. Offline cost is hidden: the method requires a large FEM-generated dataset and two sizable networks, while the baseline uses a smaller dataset and a single network. The paper does not report the total data-generation and training cost of the proposed method.

W3. The DirectNet comparison is unfair, and the “unseen load” experiment is self-defeating. The proposed method is trained with substantially more samples and parameters than DirectNet; the “same split” claim covers only one subnetwork’s dataset. More seriously, the paper admits that a DirectNet input dimension is identically zero throughout its training set, so the corresponding weights receive no gradient and remain at initialization. Reporting that baseline’s catastrophic error therefore measures an untrained input dimension, not end-to-end learning, and cannot support the claimed superiority of decomposition.

W4. Superposition validation and efficiency comparison are incomplete. The paper claims a three-term decomposition, but the unseen-load experiment validates only a reduced two-term case with one boundary forced homogeneous. There is no experiment for source plus both active boundaries, different load amplitudes, all boundary-type combinations, or error accumulation. The speed comparison omits offline data generation and training cost, and does not compare against “offline FEM basis solves + analytical superposition” or operator-learning baselines under matched budgets. No multi-seed uncertainty, per-method failure counts, or grid-convergence check for the corner singularity is reported.

W5. Presentation and reproducibility are below the bar. Equations and tables contain inconsistent symbols. The CoordConv coordinate system is unspecified: physical or normalized coordinates? If normalized, the two geometric inputs are redundant beyond their ratio; if physical, the network must extrapolate across geometries. Code and data-generation scripts are only promised as supplementary material.

Questions for authors

Q1. Give a formal definition of operator and load parameters that is invariant to variable shifts and rescaling of the PDE. Prove the superposition theorem under Robin conditions, and state the exact conditions on homogeneous reference values and boundary-type consistency. Explain the b1b1 contradiction: why is the conductivity scaling an operator parameter that must be learned when the paper itself extracts it analytically?

Q2. Re-run DirectNet with the same total sample count, total parameter budget, and training time as SINN. Do not use an input dimension that is identically zero in DirectNet’s training set as evidence of failure. Report multi-seed mean/variance and per-method failure counts. Compare against classical solvers and simple warm starts at matched solution quality.

Q3. Validate the full three-term superposition: source plus both active boundaries, multiple load amplitudes, and all boundary-type combinations. Report total offline cost, including data generation and training, and compare against “offline FEM basis solves + analytical superposition” and at least one operator-learning baseline under matched budgets. Include operator-parameter extrapolation and unseen-geometry tests. Specify whether CoordConv uses physical or normalized coordinates and release code and data-generation scripts.

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

Related papers

Discussion (0)

Sign in to comment.