acceptodds
Under review as a conference paper at ICLR 2027

UniMaNO: Neural Operator-Based Unified Material Representations for Generalizable Thermal Modeling

Abstract

Identifying material thermophysical properties from limited temperature observations and predicting subsequent thermal evolution are key goals of thermal modeling. Neural networks offer new approaches, but methods based on per-instance reconstruction still require independent optimization for each material and lack a unified representation that can be shared and transferred across materials. To address this limitation, we propose UniMaNO: Neural Operator-Based Unified Material Representations for Generalizable Thermal Modeling, a neural operator framework centered on a unified material representation. Specifically, UniMaNO maps temperature observations to material thermophysical property functions through a two-stage framework: first constructing a shared thermophysical function basis from training conductivity curves, then using temperature reconstruction errors from a differentiable heat conduction simulator to learn the observation-to-material representation mapping and refine material representations. After training, UniMaNO identifies the thermophysical properties of unseen materials without per-instance optimization, reuses the inferred representations for thermal evolution prediction under new geometries and heat-source conditions, and further supports few-shot refinement using the input observations. Extensive experiments demonstrate that UniMaNO enables rapid property identification and physically consistent thermal evolution prediction across materials and operating conditions.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Scientific Machine Learning, Thermophysical Property Identification, Neural Operators

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_eiqpOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

UniMaNO learns to infer a temperature-dependent conductivity function and an unknown heat source from transient temperature observations. It constructs a fixed rank basis in log-conductivity space, learns the material coefficients and source using supervised labels, simulation feedback, and consistency across experiments of the same material, and uses a conventional heat solver for subsequent temperature prediction.

Critical strengths
  1. The separation between material identification and subsequent physical simulation is useful. In the cross-condition protocol, the recovered conductivity is held fixed and combined with a prescribed target source and target geometry. This tests predictive usefulness beyond reconstruction with the original estimated conductivity–source pair.

  2. The supplementary analysis reports important failure modes rather than only favorable averages. It separates errors over observed and full temperature ranges, evaluates changes in identification conditions, and reports difficult-case statistics. These analyses materially improve the interpretation of the results.

Critical weaknesses
  1. The future heat source is not specified sufficiently to reproduce forecasting. Appendix C.1 explicitly states that the reference source is not used at inference under the original operating condition. Appendix C.2 then states that frames 1–40 identify conductivity and heat source, and that the solver predicts frames 41–80 using the inferred conductivity and model-predicted source. However, the manuscript does not specify how the source values required at those future times are obtained. This is central to the problem definition. Identical temperature histories can be followed by different future heating schedules, producing different future temperatures. Accurate prediction from the history therefore requires restrictions or statistical predictability in the future forcing. Although the paper acknowledges learned heating priors, it does not identify the temporal assumptions supporting this forecast or explain whether the network predicts future samples, estimates a source law, or extends an estimate from the observation window.
Questions for authors
  1. For original condition prediction, what exactly produces the heat source values used over frames 41–80? Since I find this a core flaw(or misunderstanding) on the problem setting. As I explained in the weakness, identical temperature history can lead to totally different future temperature curve. What constraints or learned temporal priors make the future heat source predictable from the observed history?

  2. To what extent can the reported improvements be attributed to the proposed material representation, as distinct from the use of a PDE solver?

  3. How accurately does the source branch recover the ground-truth heat source in space and time? Is there direct quantitative or qualitative evidence of source recovery accuracy beyond downstream temperature prediction?

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

This paper proposes UniMaNO, a neural operator framework for identifying temperature-dependent thermal conductivity from limited temperature observations and reusing the inferred material properties for thermal prediction. The method constructs a shared log-POD basis and learns an observation-to-material mapping through differentiable simulation feedback. It supports zero-shot identification without per-instance optimization and further allows few-shot adaptation. The authors introduce MatTherm, a synthetic benchmark covering diverse materials, geometries, and heat sources in 2D and 3D. Experiments show strong results in conductivity identification and thermal prediction, including generalization to unseen materials and operating conditions.

Critical strengths
  1. A coherent framework for material identification. The shared log-POD basis provides a compact representation of conductivity functions, while differentiable simulation feedback connects material identification with thermal prediction. This design avoids repeated optimization for each new material.
  2. Comprehensive evaluation within the synthetic benchmark. The experiments cover 2D and 3D settings, different geometries and heat sources, long-horizon prediction, and cross-condition generalization. The results demonstrate clear improvements over the compared methods.
Critical weaknesses
  1. Limited validation beyond synthetic data. All experiments rely on simulated trajectories generated under a fixed physical model. The method has not been evaluated on measured data, where uncertain boundary conditions, measurement noise, and model mismatch may affect identification accuracy.
  2. Limited generalization to unseen material families. Although the test materials are distinct from training samples, they appear to follow the same response archetypes and functional subfamilies. The results therefore mainly demonstrate generalization within known material families, leaving performance on genuinely unseen conductivity behaviors unclear.
  3. Insufficient evaluation of conductivity recovery. Table 3 reports identification results for only three seen and three unseen materials. This is a rather small sample for supporting the claimed identification accuracy and efficiency, particularly compared with the larger-scale thermal prediction experiments.
  4. Unclear interpretation of full-range conductivity recovery. The conductivity functions span 200–1000 K, while the training trajectories remain within 200–600 K. Full-range recovery therefore partly depends on the learned material prior, and its distinction from observation-supported recovery deserves more discussion.
Questions for authors
  1. How does UniMaNO perform on held-out conductivity families or under physical model mismatch? These settings would help establish the limits of its generalization.
  2. Can the conductivity identification results in Table 3 be extended to more materials, with variability across test instances reported?
  3. How much of the full-range conductivity recovery beyond the observed temperature range comes from the learned material prior?
  4. In the cross-condition experiments, UniMaNO uses an inferred conductivity function with a PDE solver, while the neural-operator baselines directly predict temperature evolution. How should this difference in prediction pipelines be considered when interpreting the comparison?
Reviewer_fKiuOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes UniMaNO to identify the temperature-dependent thermal conductivity function k(T) from temperature observations and predict future temperature evolution. In Stage I, a 32-dimensional shared basis is constructed by Log-POD and used as a fixed trunk in POD-DeepONet. In Stage II, the model jointly learns material representations and heat-source predictions by combining material-coefficient loss, heat-source loss, cross-condition consistency loss for the same material, and temperature reconstruction loss through a differentiable finite-volume solver. The authors also build a 2D and 3D synthetic benchmark called MatTherm, and evaluate the method on long-horizon prediction, cross-condition prediction, and conductivity inversion tasks.

Critical strengths

Decoupling material identification and physical simulation for cross-condition reuse. By recovering k(T) through a low-dimensional representation and combining it with a numerical solver (given target geometry, initial field, and known heat source), the identified material property can be directly reused under different operating conditions. The experimental results in Table 1 and Table 2 show good prediction accuracy and reusability across the tested materials and conditions.

Critical weaknesses

Practicality of the experimental setup and observation assumptions. In the 3D experiments, the model takes almost complete volumetric temperature fields as input, where each temperature entry is only randomly dropped with a very small probability of 5×10^{−4} . However, in real-world scenarios, we usually only have surface temperature measurements or sparse sensor observations, which are not verified in this paper. Besides, the experiments fix the constants as a=10^{−4} and β=0.1, and assume zero normal heat-flux boundaries (Appendix A). In addition, while the paper highlights the fast per-instance identification speed, the offline data generation time and training costs are not reported. It is also unclear what the real-time latency requirements are in target applications and how many test instances are needed to justify the offline cost. Sensitivity to heat-source shifts during identification and limited inversion baselines. As shown in Table 14, when there is a heat-source shift in the identification stage, the mean relative L_2 error of conductivity over the full temperature range across 500 seen materials increases significantly from 0.363 to 2.74. This suggests that the property recovery is quite sensitive to the training heat-source distribution. Meanwhile, the error within the observed temperature range (Obs. L2) is missing for this setting. Furthermore, the direct inversion comparison in Table 3 only covers 3 seen and 3 unseen materials, and lacks comparisons with standalone amortized neural inversion baselines. Entangled comparisons in forward prediction tasks. UniMaNO combines learned material identification with a numerical solver for rollout, whereas the forward baselines predict temperatures autoregressively. Also, in the cross-condition setting, UniMaNO identifies the material from Experiment A, while the forward baselines use the target temperature history from Experiment B as input (Appendix C.2). Because of these different setups, it is hard to tell how much of the performance gain comes from the learned material representation itself versus the numerical solver pipeline. In particular, the near-zero energy balance error is largely due to the conservative numerical scheme, no-flux boundary conditions, and given target heat sources in cross-condition settings. It would be more convincing to include controlled comparisons against other inversion baselines that share the exact same backend solver.

Questions for authors
  1. Application scenarios and observation setups: What are the actual latency and throughput requirements for the main target applications? Could the authors report the time costs for offline data generation, model training, and downstream numerical integration, and discuss when this approach becomes more cost-effective than traditional methods at a comparable accuracy level? Also, how does the model perform under surface-only or sparse sensor observations, and how can we obtain the full initial temperature field in practice? Finally, is the model robust when the assumed physical constants are inaccurate?
  2. Robustness to heat-source shifts: Could the authors provide the Obs. L2 conductivity error and the heat-source estimation error for Table 14? When the true effective heat source during identification is known, will directly feeding q⋆q^\star into the material branch or applying UniMaNO-FS help reduce the performance drop in full-range property recovery?
  3. Ablation studies and inversion baselines: In Table 17, removing Lsim\mathcal{L}_{\mathrm{sim}} only increases the RMSE by about 1.4%1.4\% under the seen-material heat-source change setting, compared to 20.8%20.8\% for long-horizon Tail RMSE. Could the authors report the Full and Obs. L2L_2 conductivity errors across all 500 unseen materials for both the full model and the variant without Lsim\mathcal{L}_{\mathrm{sim}}? Also, what is the impact of removing the direct coefficient supervision Lz\mathcal{L}_z in Stage II while keeping Lq\mathcal{L}_q and Linv\mathcal{L}_{\mathrm{inv}}? Finally, please consider expanding the test set in Table 3 and adding an amortized inversion baseline that uses the same backend solver.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.