acceptodds
Under review as a conference paper at ICLR 2027

Manifold-Projected Diffusion Model for Generating Physical Visual Fields

43 views

Abstract

Diffusion models are expressive generative priors for physical fields, but standard formulations typically treat coupled field variables as free outputs even when some are deterministically determined by known physical operators. As a result, generated parameter–state pairs can be individually plausible yet mutually inconsistent under the governing physics, particularly when forcing or boundary conditions shift beyond those seen during training. We propose the Manifold-Projected Diffusion Model (MPDM), which factorizes physical-field generation according to this dependency structure. Instead of directly denoising the complete field stack, MPDM predicts a compact latent representation of the independent physical parameter field, decodes it into a feasible parameter map, and uses a differentiable numerical solver to render the dependent state fields under prescribed conditions. This construction restricts each clean-state prediction to a solver-rendered, operator-consistent manifold while allowing reconstruction losses to propagate through the solver for end-to-end training without an auxiliary physics-residual objective. Across multiple PDE systems under controlled distribution shifts, MPDM consistently reduces state reconstruction error relative to ambient full-state and residual-based diffusion, while achieving strong parameter recovery and inverse consistency. These results demonstrate the benefit of aligning generative parameterization with the dependency structure imposed by the governing physics.

Primary area:
neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)
Keywords:
Diffusion models, generative models

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_bydVOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes a method for reconstructing physical fields consistent with a set of given conditions. The method factorises physical parameters from state variables that can be derived from the forward operator. It then performs diffusion-based generation over the joint field space, predicting the physical parameters and deriving the state variables from the known physics at each diffusion step. The method is evaluated on out-of-distribution settings with simulated data.

Critical strengths

The method ensures that predicted parameter-state pairs are consistent with the embedded numerical solver, without requiring an additional physics-residual loss. The experiments cover multiple PDE systems and include ablations examining the contributions of the different components.

Critical weaknesses
  1. Limited novelty. A primary contribution is the factorization into learned physical parameters and solver-generated dependent states, but this approach has substantial overlap with prior work combining latent generative models and differentiable PDE solvers (e.g., LD-DIM, L-DPS, GenPANIS). While embedding this factorization directly into a joint-space diffusion denoiser is a meaningful architectural distinction, it is unclear whether this constitutes a sufficiently substantial methodological advance.

  2. Insufficient justification of the proposed architecture. The experiments demonstrate the benefits of solver-based rendering, but it is unclear whether the proposed joint-space diffusion formulation offers significant advantages over simpler parameter-only generative approaches followed by the same solver. Indeed, the controlled ablation shows that simpler alternatives achieve comparable or better unconditional generation performance. This raises questions about whether the additional complexity of joint-space diffusion is justified.

Questions for authors
  1. [Clarification] Please can you clarify what joint-space diffusion offers over parameter-only generative models followed by the same numerical solver? In particular, how should we interpret the comparable or better unconditional generation results of simpler methods in Table 4? E.g., in Table 4, MPDMfield performs almost identically to MPDM on OOD denoising (0.5598 vs. 0.5576), suggesting that most of the improvement comes from solver-based rendering rather than compact latent parameterization.
  2. [Clarification/Future work] To what extent are the OOD improvements due to the proposed diffusion architecture, rather than access to the correct test-time physical conditions through the solver? Would simpler physics-aware baselines achieve similar results given the same information?
  3. [Future work] Please can you report wall-clock training and inference times relative to the baselines? How does the cost of repeated PDE solves compare with parameter-only generation followed by a single solve?
6 comments

We thank the reviewer for the constructive feedback.

[CW1]: Methodological Novelty Beyond Latent Generative Models and PDE Solvers

On the central methodological contribution. MPDM reformulates diffusion denoising around the deterministic dependency between independent physical parameters and their associated states. Specifically, MPDM performs diffusion over the joint parameter--state stack x0=[p1:m,K]x_0=[p_{1:m},K], while constructing each clean-state prediction through

x^0=[RE(Dϕ(hθ(xt,t))),Dϕ(hθ(xt,t))].\hat{x}_0=\left[\mathcal R_{\mathcal E}(\mathcal D_\phi(h_\theta(x_t,t))),\mathcal D_\phi(h_\theta(x_t,t))\right].

Crucially, the solver-rendered clean-state prediction participates directly in both the supervised denoising objective and the DDPM reverse transition. This formulation incorporates the physical dependency structure into the denoiser's admissible prediction space and learning process, rather than introducing physics solely through separate inverse optimization or sampling-time correction. This solver-rendered clean-state mapping and joint diffusion objective were already specified in the original Sections 3--4; the revision makes their distinction from closely related latent physics-aware methods more explicit.

 

On distinctions from related methods. The cited approaches integrate latent representations and physical models through different mechanisms. LD-DIM combines a pretrained latent diffusion prior with differentiable PDE-constrained inverse optimization. L-DPS employs latent diffusion posterior sampling with surrogate-based likelihood guidance. GenPANIS uses a latent-variable generative model with a normalizing-flow prior and a physics-aware coarse-grid PDE decoder. In contrast, MPDM integrates the numerical PDE solver directly into the joint-field denoiser, constructing solver-consistent parameter--state predictions at each reverse step without separate inverse optimization, likelihood guidance, or a learned surrogate renderer. We have clarified these methodological distinctions in revised Section 2 and Appendix I.

 

On the learning mechanism. MPDM jointly optimizes coefficient recovery and dependent-state reconstruction through Eqs. (14) and (16). The coefficient loss directly supervises physical parameter inference, while the state-reconstruction loss propagates gradients through the differentiable PDE solver and frozen pretrained decoder to the coordinate predictor. These complementary supervision signals train the denoiser to infer compact coefficient representations that account for both parameter fidelity and their associated physical responses. Thus, MPDM unifies coordinate-space parameter inference, solver-rendered state reconstruction, and joint-field supervision within a single diffusion objective, without requiring an auxiliary PDE-residual penalty.

 

On empirical evidence of the architectural contribution. The revised experiments provide controlled evidence beyond solver consistency. In revised Tables 1–3, MPDM consistently improves both state and coefficient MSE over MPDM$_{\mathrm{field}}$, which shares the same numerical solver, joint-field diffusion objective, and reverse DDPM updates but predicts coefficients directly in field space. Furthermore, the DM + Solver comparisons in Appendix H (Table 13) demonstrate that applying the same solver to DM predictions does not achieve comparable parameter recovery. On Darcy, MPDM reduces state MSE from 0.04300 to 0.00665 and coefficient MSE from 1.04826 to 0.55704 relative to DM + Solver. The additional three-dimensional Darcy evaluation further demonstrates improved volumetric coefficient recovery. Together, these controlled results substantiate MPDM's methodological contribution: embedding learned coordinate-space parameter inference and solver-rendered clean-state prediction into joint-field diffusion improves both physical-field reconstruction and coefficient recovery.

[CW2]: Justification of Joint-Space Diffusion

On the motivation for joint-space diffusion. The distinction between MPDM and parameter-only generative models lies in the inference objective, rather than the ability to produce solver-consistent states. A parameter-only generator learns the coefficient distribution and subsequently renders the dependent states through a numerical solver. MPDM instead learns to recover physical parameters from noisy joint parameter--state observations, using both coefficient reconstruction and solver-rendered state reconstruction to train its denoiser. Consequently, the dependent physical states contribute directly to learning the underlying coefficient representation, rather than being generated only after coefficient prediction. This provides an integrated formulation for joint-field denoising and physical parameter inference under changing physical conditions.

 

On the unconditional generation results. The original submission already included coefficient-only baselines (K-DM + Solver, K-LDM + Solver, and VAE-K + Solver) and explicitly separated paired OOD denoising from unpaired unconditional generation in Table 4 and Appendix A. VAE-K + Solver achieves the lowest unconditional Joint SW1\mathrm{SW}_1 (0.0783 versus 0.3389 for MPDM), as already acknowledged in Section 5.3 and Appendix A. This result demonstrates the effectiveness of directly modeling the coefficient distribution for unconditional generation. MPDM, however, is designed to exploit noisy joint-field observations for coefficient recovery and dependent-state reconstruction. A coefficient-only generator followed by a solver does not, by itself, provide the corresponding observation-conditioned inference mechanism. Thus, the unconditional generation results and MPDM's joint-field denoising advantages characterize complementary modeling capabilities rather than contradictory findings.

 

On controlled evidence for the architecture. We have strengthened the experimental evidence through the additional DM + Solver comparisons in Appendix H and MPDM$_{\mathrm{field}}$ results in revised Tables 1–3. All solver-based methods use the same numerical solvers and OOD physical conditions. Relative to DM + Solver, MPDM reduces coefficient MSE from 1.04826 to 0.55704 on Darcy, from 3.188 to 1.092 on Helmholtz, and from 0.06434 to 0.03755 on Navier--Stokes, while substantially improving state reconstruction. Furthermore, MPDM consistently outperforms MPDM$_{\mathrm{field}}$ in state and coefficient MSE across all three PDEs, isolating the benefit of learned coordinate parameterization within the same joint-space diffusion framework. The additional intrinsic-dimensionality experiments in Appendix A further show that the relative OOD denoising improvement over MPDM$_{\mathrm{field}}$ increases from 0.4% at six dimensions to 18.4% at fifteen dimensions. These results establish measurable benefits from both solver-integrated learning and compact parameter prediction.

 

On the computational trade-off. The additional computation serves the joint-field inference objective. In the reported paired OOD denoising evaluations, MPDM uses a single denoising prediction and one batched PDE rendering step, while full multi-step DDPM generation invokes the renderer at each reverse step. We have additionally reported wall-clock training and inference comparisons in revised Appendix B.3, allowing the reconstruction benefits to be assessed alongside computational costs. Taken together, the results justify joint-space diffusion for physically structured denoising and parameter recovery, where the learned predictor exploits coupled physical observations to achieve substantially more accurate reconstructions than solver post-processing alone.

[Q1]:

On the benefit of joint-space diffusion. Parameter-only generators learn a distribution over coefficients and subsequently render the dependent states, whereas MPDM learns to recover coefficients from noisy joint parameter--state observations. Its joint-field reconstruction objective combines direct coefficient supervision with state-reconstruction gradients propagated through the numerical solver. Thus, the physical states contribute to learning the coefficient predictor rather than being generated solely after coefficient sampling. This distinction is particularly relevant to paired OOD denoising and inverse parameter recovery.

[Q1]:

On the controlled ablation. The original submission already included parameter-only solver baselines and explicitly separated paired OOD denoising from unpaired generation in Table 4 and Appendix A. The stronger unconditional Joint SW1\mathrm{SW}_1 of VAE-K + Solver (0.0783 versus 0.3389 for MPDM) therefore reflects a task-specific advantage of coefficient-distribution modeling rather than an unexplained failure of MPDM. Regarding the small MPDM--MPDM$_{\mathrm{field}}$ gap in the original six-dimensional setting, the additional experiments in Appendix A show that MPDM's relative OOD denoising improvement increases from 0.4% at 6 dimensions to 5.9% at 10 dimensions and 18.4% at 15 dimensions. Moreover, revised Tables 1–3 demonstrate consistently lower state and coefficient MSE for MPDM than MPDM$_{\mathrm{field}}$ across Darcy, Helmholtz, and Navier--Stokes. These results establish the additional value of compact coordinate parameterization within joint-field denoising, beyond the shared use of a numerical solver.

[Q2]:

On disentangling solver and learning contributions. We have introduced DM + Solver and strengthened the MPDM$_{\mathrm{field}}$ comparisons to distinguish numerical rendering from learned coefficient inference. Importantly, all solver-based methods use the same numerical solvers and OOD physical conditions, so MPDM's advantages cannot be attributed solely to exclusive access to the correct test-time conditions. DM + Solver renders physically consistent states from DM-predicted coefficients but does not improve the coefficient estimates themselves. MPDM$_{\mathrm{field}}$ integrates solver-based state reconstruction into diffusion training, while MPDM further introduces learned coordinate-space parameterization.

 

On the experimental evidence. In revised Appendix H, MPDM reduces coefficient MSE relative to DM + Solver from 1.04826 to 0.55704 on Darcy, 3.188 to 1.092 on Helmholtz, and 0.06434 to 0.03755 on Navier--Stokes, alongside consistent improvements in state reconstruction and inverse consistency. The additional MPDM$_{\mathrm{field}}$ comparisons in revised Tables 1–3 further demonstrate the benefit of compact coefficient prediction under the same solver-integrated training framework. Together, these results provide direct evidence that MPDM's OOD improvements reflect learned parameter inference and structured diffusion parameterization, in addition to the physical consistency supplied by the numerical solver.

[Q3]:

On wall-clock efficiency. We report relative wall-clock training and inference measurements for the three main PDE benchmarks, normalized to the standard DM baseline (1.0). The following comparison includes solver-based and parameter-only alternatives, while the MPDM versus MPDM$_{\mathrm{field}}$ comparison is also documented in revised Appendix B.3 (Table 8):

PDEMethodRelative Training Time ↓\downarrowRelative Inference Time ↓\downarrow
DarcyDM + Solver1.0001.481
MPDM$_{\mathrm{field}}$0.7711.254
VAE-K + Solver0.0221.009
MPDM0.1811.007
HelmholtzDM + Solver1.0001.922
MPDM$_{\mathrm{field}}$1.3221.621
VAE-K + Solver0.0571.889
MPDM0.4312.035
Navier--StokesDM + Solver1.0005.486
MPDM$_{\mathrm{field}}$1.2984.327
VAE-K + Solver0.0193.117
MPDM2.0803.957

 

On repeated PDE solves. In the paired OOD denoising evaluation, MPDM performs a single clean-state prediction followed by one batched PDE rendering call, comparable to parameter-only prediction followed by the same solver. Full multi-step DDPM sampling instead evaluates the solver at each reverse step, whereas parameter-only generation requires only a final PDE solve. This distinction is important when comparing the computational requirements of denoising and unconditional generation.

 

On the accuracy--efficiency trade-off. MPDM achieves substantially lower state reconstruction and coefficient errors than the solver-matched alternatives, with solver-dependent computational costs. Compared with MPDM$_{\mathrm{field}},MPDMreducestrainingtimeonDarcyandHelmholtzandinferencetimeonDarcyandNavier−−Stokes,whileimprovingreconstructionaccuracyacrossallthreebenchmarks.Theseresultsdemonstratetheeffectivenessofsolver−integratedcoordinatepredictionandcharacterizeitsaccuracy−−efficiencytrade−off.TheMPDMversusMPDM, MPDM reduces training time on Darcy and Helmholtz and inference time on Darcy and Navier--Stokes, while improving reconstruction accuracy across all three benchmarks. These results demonstrate the effectiveness of solver-integrated coordinate prediction and characterize its accuracy--efficiency trade-off. The MPDM versus MPDM_{\mathrm{field}}$ timing comparison is also reported in revised Appendix B.3 (Table 8).

Reviewer_cwNKOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The Manifold-Projection Diffusion Model (MPDM) addresses a well-motivated problem in generating physical fields with diffusion models. Standard diffusion models treat physically coupled variables as independent outputs, producing parameter and state pairs that may each appear plausible but jointly violate physical laws. This inconsistency is particularly severe outside the training distribution. MPDM introduces a meaningful structural design: some variables in the joint output can be determined by other variables and known physical laws. Exploiting this relationship can also reduce the learning burden on the network. The authors evaluate MPDM on Darcy flow, a Helmholtz-type equation, and steady Navier–Stokes equations with variable viscosity, focusing on OOD reconstruction under changes in boundary conditions and external forcing. The results show substantial reductions in state reconstruction error across multiple benchmarks, along with improvements in some parameter recovery and inverse consistency metrics. Controlled ablations also examine the contributions of the numerical solver, the low-dimensional coordinate parameterization, and different approaches to generating parameter fields.

Critical strengths

The paper addresses a practically meaningful problem and effectively combines physical consistency with trainable parameter estimation through structured parameterization.

  1. MPDM exploits the deterministic dependency between parameter and state fields by having the network predict the independent parameters and using a differentiable numerical solver to construct the dependent states. This design ensures that clean predictions are consistent with the embedded discrete physical model. It also allows gradients from state reconstruction supervision to propagate through the solver to the parameter predictor, without an additional PDE residual penalty.

  2. Under the evaluation protocols used in the paper, improvements in both state and parameter recovery errors are reported for all three main PDEs, including the nonlinear steady Navier–Stokes equations. In Table 5, MPDM reduces parameter MSE from 0.1817 for DM to 0.1699 and inverse consistency error from 0.2771 to 0.2362. These results provide stronger evidence of practical inference value than a near-zero self-residual.

Critical weaknesses
  1. The relationship to Diffusion with Forward Models (DFM) is insufficiently clarified. The related work section describes DFM as using differentiable forward models for diffusion-based inverse problems. However, DFM already adopts a denoising structure that predicts underlying variables and then uses a forward model to construct a clean observation estimate. The difference in problem settings alone is insufficient to establish a new diffusion mechanism. DFM primarily studies inverse problems without direct latent supervision, whereas MPDM performs supervised denoising in the full space of PDE parameter and state fields, using a pretrained parameter decoder and a numerical PDE solver. The authors should clarify how MPDM’s training and prediction objectives relate to, and differ from, those of such diffusion methods with forward models.

  2. The OOD experiments in Tables 1, 2, and 3 do not sufficiently distinguish the benefits of the known numerical solver from the contribution of MPDM’s learning mechanism. All three main experiments preserve the distribution of the parameter field (K) and change only the external physical conditions: the Dirichlet boundary patterns for Darcy flow, the spatial frequency of the forcing for Helmholtz, and the phase of the forcing for Navier–Stokes. MPDM can pass the test conditions directly to the known numerical solver. Therefore, for any feasible predicted parameter field, the constructed state field will satisfy the corresponding discrete PDE as long as the numerical solve succeeds, with the consistency error limited by the solver’s numerical accuracy. This is a structured physical modeling strategy. The property is a natural consequence of using a known forward solver to generate dependent states and is not unique to MPDM’s parameter predictor or low-dimensional latent representation. The current results do not sufficiently establish how much of the improvement in state reconstruction comes from the known solver and how much comes from MPDM’s specific parameter estimation and diffusion design. More importantly, all three main experiments involve two-dimensional, steady PDEs on regular domains with explicitly available numerical solvers. The paper does not evaluate three-dimensional problems, where numerical solution is often substantially more expensive and challenging. For example, if the existing Darcy benchmark were extended to flow through three-dimensional porous media, would the physical field generation model still provide meaningful benefits beyond those demonstrated in the current two-dimensional setting?

Minor Comments

  1. Please check the citation formatting and use \citet{} for citations that form part of a sentence and \citep{} for parenthetical citations, as specified in the ICLR 2027 template. Page 5 also contains an unresolved “Appendix ??” reference.

  2. Some terminology overstates what the method does.

    (1) Manifold Projection. The method directly constructs a constrained clean-state prediction through a solver-rendered parameterization, rather than explicitly computing a metric projection onto a physical manifold. The distinction should be made clear.

    (2) Coordinate-space denoising. The model predicts low-dimensional parameter coordinates, but the forward diffusion process and noisy states remain in the full ambient field space. The terminology should not lead readers to believe that the entire diffusion process operates in a low-dimensional latent space.

Questions for authors
  1. Existing PDE solution methods based on physics-informed neural networks (PINNs) and random feature methods (RFMs) incorporate known physical laws through equation residuals or loss functions, grid-based states, or boundary condition handling. Some methods with hard constraints also satisfy specified constraints directly through output parameterization. Could the authors clarify the connections and differences between these ideas and MPDM’s solver-rendered decomposition of parameter and state fields? In particular, beyond inheriting the numerical solver’s discrete physical consistency, what independently verifiable modeling or inference benefits arise from embedding this structure in diffusion denoising?
  2. The paper defines OOD as a difference between the test and training distributions. This does not require more difficult PDEs, higher dimensionality, or distribution shifts in every variable. In the reported experiments, the shift is confined to external physical conditions. Which results under these shifts demonstrate the generalization ability of the learned modules, and which primarily reflect the known solver’s ability to handle new conditions?
  3. To address Critical Weakness 1, could the authors directly adapt DFM to this paper’s experimental setting and compare it with MPDM using the same observations, supervision, and physical conditions, together with comparable networks and training budgets?
  4. To address Critical Weakness 2, could the authors report DM + Solver results on all three main PDE benchmarks? Specifically, the parameter fields predicted by DM would be passed through the same numerical solver used by MPDM, under the same test conditions, to reconstruct the state fields. These results would help determine how much of the reported improvement comes from the known solver and how much comes from MPDM’s specific learning mechanism.
7 comments

We thank the reviewer for the constructive feedback.

[CW1]: Relationship to DFM

On the methodological distinction from DFM. Although DFM (Tewari et al., NeurIPS 2023) and MPDM both integrate forward models into diffusion denoising, they differ in three key methodological aspects: diffusion variables, clean-state parameterization, and training objectives. DFM performs observation-space diffusion, inferring unobserved signals through target and novel-observation reconstruction losses. In contrast, MPDM performs diffusion over the joint parameter--state stack x0=[p1:m,K]x_0=[p_{1:m},K] and constructs a solver-rendered clean estimate x^0=[RE(K^),K^]\hat{x}_0=[\mathcal R_{\mathcal E}(\hat K),\hat K], explicitly supervising both coefficient recovery and state reconstruction. This formulation integrates direct physical-parameter supervision with solver-constrained joint-field denoising, rather than relying solely on observation reconstruction. The solver-rendered clean-state formulation itself was already specified in the original Sections 3--4. In response to the reviewer, we have made its relationship to DFM explicit in revised Section 2 and provided a formulation-level comparison in Appendix G.

 

On diffusion and prediction spaces. DFM performs diffusion over target observations and reconstructs them through a forward model. MPDM instead applies diffusion to the complete parameter--state stack x0=[p1:m,K]x_0=[p_{1:m},K] (Eq. 12) and constructs its clean-state prediction as

x^0=[RE(K^),K^],K^=Dϕ(hθ(xt,t)).\hat{x}_0=\left[\mathcal R_{\mathcal E}(\hat K),\hat K\right],\qquad \hat K=\mathcal D_\phi(h_\theta(x_t,t)).

This formulation explicitly retains the physical parameter field within the denoised sample while generating dependent states through the numerical PDE solver. Consequently, the learned coefficient representation and solver-rendered state fields jointly define the admissible clean-state prediction space, establishing a structured denoising formulation beyond observation reconstruction alone.

 

On training objectives. DFM learns underlying signals through reconstruction losses on rendered target and novel observations, without direct supervision of those signals. In contrast, MPDM jointly supervises parameter recovery and state reconstruction through Eqs. (14) and (16):

LMPDM=E[wp∑j=1m∥RE,j(K^)−pj∥22+wK∥K^−K∥22].\mathcal L_{\mathrm{MPDM}}=\mathbb E\left[w_p\sum_{j=1}^{m}\|\mathcal R_{\mathcal E,j}(\hat K)-p_j\|_2^2+w_K\|\hat K-K\|_2^2\right].

The coefficient loss directly supervises parameter recovery, while the state-reconstruction loss propagates gradients through the PDE solver and frozen pretrained decoder to the coordinate predictor. These complementary objectives couple physical parameter inference with the reconstruction of dependent states, while the learned coefficient chart provides a compact, data-supported representation for denoising.

 

On architectural evidence. We have strengthened the controlled comparison with MPDM$_{\mathrm{field}}$ in revised Tables 1–3. Both methods share the same diffusion process, reconstruction objective, numerical solver, and reverse DDPM updates, differing in their coefficient parameterization. MPDM reduces state MSE from 0.03497 to 0.00665 on Darcy, from 5.64×10−85.64\times10^{-8} to 7.70×10−97.70\times10^{-9} on Helmholtz, and from 0.00937 to 0.00327 on Navier--Stokes, while consistently improving coefficient recovery. These results demonstrate that learned coordinate-space parameterization provides measurable benefits beyond solver-based rendering alone. Together, these methodological distinctions and controlled results demonstrate the value of MPDM's joint-field diffusion formulation, which integrates compact coefficient inference, direct parameter supervision, and solver-rendered state reconstruction into a unified denoising framework.

[CW2]: Disentangling Solver Contributions and Three-Dimensional Applicability

On the contribution of the numerical solver. The original submission already distinguished solver-imposed physical consistency from accurate parameter recovery. In Section 5, we report coefficient MSE and inverse consistency (ICSK\mathrm{ICS}_K) in addition to physical residuals, and explicitly note that near-zero residuals are expected from solver-based rendering rather than evidence of successful coefficient identification. We agree, however, that these metrics alone do not fully isolate how much of the reconstruction gain is attributable to solver access. We therefore introduced a DM + Solver baseline, which reconstructs dependent states from DM-predicted coefficients using the same numerical solver and OOD physical conditions as MPDM. Importantly, all solver-based comparisons have access to the same test-time physical conditions, allowing us to isolate the additional contribution of learned coefficient inference.

 

On disentangling the learning contributions. We compare three complementary formulations under matched OOD conditions: DM + Solver applies physical rendering after standard diffusion denoising; MPDM$_{\mathrm{field}}$ integrates the numerical solver into the supervised clean-state predictor while estimating coefficients directly in field space; and MPDM further introduces learned coordinate-space parameterization. The supplementary results are:

PDEMethodState MSE ↓\downarrowMSEK↓\mathrm{MSE}_K\downarrowICSK↓\mathrm{ICS}_K\downarrow
DarcyDM + Solver0.043001.0482657.7571
MPDM$_{\mathrm{field}}$0.034971.0020056.2300
MPDM0.006650.5570418.3720
HelmholtzDM + Solver5.722×10−85.722\times10^{-8}3.1880.01137
MPDM$_{\mathrm{field}}$5.640×10−85.640\times10^{-8}3.3510.01159
MPDM7.704×10−97.704\times10^{-9}1.0920.00553
Navier--StokesDM + Solver0.011860.064340.43916
MPDM$_{\mathrm{field}}$0.009370.057970.33840
MPDM0.003270.037550.18828

 

On the independent benefits of learning. Solver post-processing substantially reduces state reconstruction errors but leaves the predicted coefficients unchanged. Among the solver-matched methods, MPDM achieves the lowest state MSE, coefficient MSE, and inverse-consistency error across all three PDEs, demonstrating improvements in physical parameter recovery rather than merely satisfying the numerical solver. Moreover, the comparison with MPDM$_{\mathrm{field}}$ isolates the benefit of learned coordinate parameterization within the same solver-integrated diffusion framework. Together, these comparisons distinguish the roles of physical rendering, solver-integrated learning, and compact parameter prediction. We have added the complete solver-matched evaluation in revised Appendix H (Table 13), with the coordinate-space comparisons also included in revised Tables 1–3.

 

On three-dimensional applicability. We have further extended MPDM to three-dimensional Darcy flow, evaluating volumetric coefficient fields and their corresponding PDE states. The additional results in Appendix H (Table 14) are:

MethodState MSE ↓\downarrowMSEK↓\mathrm{MSE}_K\downarrowICSK↓\mathrm{ICS}_K\downarrow
DM + Solver4.039×10−34.039\times10^{-3}1.366×10−21.366\times10^{-2}2.85195
MPDM$_{\mathrm{field}}$1.935×10−31.935\times10^{-3}1.279×10−21.279\times10^{-2}2.83839
VAE-K + Solver2.084×10−42.084\times10^{-4}1.605×10−31.605\times10^{-3}0.48158
MPDM5.884×10−55.884\times10^{-5}4.006×10−44.006\times10^{-4}0.27938

MPDM achieves the best state reconstruction, coefficient recovery, and inverse-consistency results among these methods. Relative to DM + Solver, MPDM reduces state MSE by approximately 68.6$\times$ and coefficient MSE by 34.1$\times$, demonstrating substantial improvements under the more challenging volumetric formulation. Together, the solver-matched and 3D experiments demonstrate that MPDM's performance gains arise not merely from access to a numerical solver, but also from its learned coefficient inference and structured diffusion parameterization.

[MC1]:

We have corrected the citation formatting throughout the revised manuscript, using \citet{} for textual citations and \citep{} for parenthetical citations. We have also fixed the unresolved Appendix reference on page 5.

[MC2]: Terminology Clarification

(1) Manifold Projection. In MPDM, manifold projection refers to an architectural constraint on the denoiser's clean-state prediction, rather than an explicit metric projection. Specifically, the predicted coordinates are mapped through the learned decoder and physical renderer to construct x^0=ΦE(z^)=[RE(Dϕ(z^)),Dϕ(z^)]∈M~E\hat{x}_0=\Phi_{\mathcal E}(\hat z)=[\mathcal R_{\mathcal E}(\mathcal D_\phi(\hat z)),\mathcal D_\phi(\hat z)]\in\widetilde{\mathcal M}_{\mathcal E}. This construction guarantees that the predicted clean state belongs to the learned solver-rendered set, ensuring parameter--state consistency up to numerical solver accuracy without requiring a separate projection optimization. We have clarified this distinction in revised Section 3 and refined the terminology to emphasize solver-rendered constraints.

 

(2) Coordinate-Space Denoising. MPDM retains the standard DDPM forward and reverse processes in the ambient parameter--state field space. Coordinate-space denoising specifically refers to the parameterization of the clean-state predictor: the network estimates compact coefficient coordinates z^=hθ(xt,t)\hat z=h_\theta(x_t,t), which are decoded into K^\hat K and used to render the dependent states. Thus, the dimensionality reduction applies to the predicted coefficient representation, while diffusion remains in the ambient field space. We have clarified this distinction in revised Sections 3 and 4.

[Q1]:

On the relationship to PINNs, RFMs, and hard constraints. PINNs and RFMs commonly incorporate governing equations through residual-based optimization or physics-informed solution approximation, while hard-constrained methods enforce selected physical constraints through output parameterization. MPDM instead integrates the known PDE solver into a joint parameter--state diffusion denoiser: it infers compact coordinates of the independent coefficient field from noisy joint observations and deterministically reconstructs the dependent states. This combines learned parameter inference with solver-rendered clean-state prediction, rather than treating physical constraints solely as optimization penalties or output feasibility requirements.

 

On independently verifiable benefits. The original submission already evaluates state reconstruction MSE, coefficient MSE, and inverse consistency (ICSK\mathrm{ICS}_K), rather than relying on physical residuals alone, precisely because solver consistency does not guarantee accurate parameter identification. To further isolate the learning contribution from solver-imposed consistency, we additionally introduced DM + Solver and strengthened the MPDM$_{\mathrm{field}}$ comparison under the same numerical solvers and test-time physical conditions. Relative to DM + Solver, MPDM reduces coefficient MSE from 1.048 to 0.557 on Darcy, 3.188 to 1.092 on Helmholtz, and 0.0643 to 0.0376 on Navier--Stokes, while also improving state reconstruction and inverse consistency. These results demonstrate measurable improvements in physical parameter inference beyond solver-imposed consistency, as detailed in revised Tables 1–3 and Appendix H.

[Q2]:

On OOD generalization of the learned modules. Our original OOD evaluation already separates parameter recovery from solver consistency by reporting coefficient MSE and inverse consistency (ICSK\mathrm{ICS}_K) in addition to physical residuals. The OOD shift specifically changes prescribed physical conditions while maintaining the coefficient distribution, as defined in Section 5.2. The learned denoiser must infer coefficients from noisy parameter--state stacks whose dependent states are generated under previously unseen boundary or forcing conditions. Its generalization is evaluated through coefficient MSE and inverse consistency (ICSK\mathrm{ICS}_K), rather than physical residuals alone.

 

On distinguishing learned inference from solver generalization. The numerical solver enables state rendering under new physical conditions and provides discrete physical consistency, but does not guarantee accurate coefficient recovery. In the additional solver-matched comparisons in revised Appendix H, DM + Solver, MPDM$_{\mathrm{field}}$, and MPDM all use the same OOD conditions and numerical solvers. Nevertheless, MPDM achieves the lowest coefficient MSE and ICSK\mathrm{ICS}_K among these solver-matched methods across Darcy, Helmholtz, and Navier--Stokes. These results distinguish the solver's ability to handle unseen conditions from the learned model's ability to recover physically meaningful coefficients under those conditions.

Reviewer_P4VhOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2(+1)
Presentation
2
Soundness
2
Read review
Paper summary

This work performs physical-field generation using 1) a differentiable renderer and 2) a latent representation. Physical consistency is enhanced by predicting physical parameters and reconstructing state fields through the differentiable renderer.

Critical strengths

This work maintains physical consistency by factorizing independent physical parameters and state variables and introducing dependencies through a differentiable renderer. Furthermore, this paper experimentally demonstrates the effectiveness of the proposed framework in out-of-distribution evaluation.

Critical weaknesses

Insufficient Empirical Evidence

  • Benefits of coordinate-space parameterization. The paper provides insufficient empirical evidence for the efficiency and effectiveness of introducing a latent representation. In Appendix A, KK is generated from an exactly 6-dimensional latent variable (Eq. (21)), and MPDM also employs a 6-dimensional coordinate chart. This represents a particularly favorable setting in which the intrinsic dimensionality is known in advance. Nevertheless, the improvements over MPDMfield\mathrm{MPDM}_{\mathrm{field}} remain marginal (Tables 5 and 6). Furthermore, the paper does not investigate how performance changes when the coordinate dimension dzd_z differs from the intrinsic dimensionality, making it difficult to assess the practical benefits of coordinate-space parameterization on real-world datasets.
  • Limitations of differentiable solvers. The paper evaluates MPDM on Darcy flow, Helmholtz-type reaction-diffusion, and high-viscosity steady-state Navier-Stokes datasets, which are relatively favorable settings for differentiable solvers because they can be handled by simple numerical methods such as conjugate gradient (CG) or fixed-point iterations. Since MPDM relies on differentiable solvers, its applicability to problems requiring non-differentiable solvers remains unclear. Evaluating MPDM on more challenging datasets, such as OpenFWI [1], which involves indefinite wave equations, and time-dependent low-viscosity Navier-Stokes datasets (e.g., FNO NS and Kolmogorov flow) [2], would provide a more comprehensive understanding of the strengths and limitations of the proposed framework.

[1] OpenFWI: Large-Scale Multi-Structural Benchmark Datasets for Full Waveform Inversion, NeurIPS 2022.

[2] Fourier Neural Operator for Parametric Partial Differential Equations, ICLR 2021.

Theories

  • Structural inability to represent the optimal denoiser. The optimal solution to the MSE objectives in Eq. (14) and (16) is the posterior mean E[x0∣xt]\mathbb{E}[x_0 \mid x_t]. However, since the renderer is nonlinear with respect to KK (Eq. (17)), the posterior mean of the rendered field (L+diagK)−1F(L+\mathrm{diag}K)^{-1}F (defined immediately after Eq. (20)) generally lies outside the manifold M~E\widetilde{\mathcal M}_ {\mathcal E} permitted by Eq. (13). Consequently, MPDM cannot represent the optimal denoiser by construction. Nevertheless, lines 289–292 directly incorporate this x^0\hat{x}_ 0 into the standard DDPM reverse transition, potentially introducing bias that causes the reverse process to deviate from the true q(xt−1∣xt)q(x_{t-1}\mid x_t). The paper does not address this tradeoff between enforcing physical consistency and compromising the theoretical justification for distribution learning. The substantial advantage of VAE-K + Solver over MPDM in unconditional generation (Tables 4 and 6) is also consistent with this potential bias.
  • Inverse stability. Proposition 3 and Corollary 1 in Appendix C.3 (Eq. (41)–(45)) establish only an upper bound based on the forward Lipschitz assumption ∥R(K1)−R(K2)∥≤LR∥K1−K2∥\lVert\mathcal R(K_1)-\mathcal R(K_2)\rVert\le L_R\lVert K_1-K_2\rVert, showing that small coordinate errors imply small errors in the stacked representation. However, parameter recovery and inverse consistency, which are presented as key contributions of the paper, concern the inverse property of recovering KK from the state field. Therefore, this upper bound does not theoretically explain these results. In fact, when the state field is insensitive to KK (i.e., when LRL_R is small), as discussed in lines 373–375, the upper bound in Eq. (42) becomes tighter, seemingly suggesting a favorable condition. In practice, however, such insensitivity makes KK more difficult to identify. Therefore, theoretical support for parameter recovery would require an analysis of inverse (conditional) stability, such as ∥z1−z2∥≤C ∥R(Dϕ(z1))−R(Dϕ(z2))∥\lVert z_1-z_2\rVert\le C\thinspace\lVert\mathcal R(D_\phi (z_1))-\mathcal R(D_\phi (z_2))\rVert. Demonstrating that the low-dimensional coordinate representation improves the stability constant CC would also provide theoretical justification for the advantages of coordinate-space parameterization.
Questions for authors
  • Please use \citep to format citations in parentheses: Dhariwal & Nichol (2021) →\rightarrow (Dhariwal & Nichol, 2021).
  • There is a hyperref error in line 266 that should be corrected.
  • Is there a tradeoff between computational efficiency and performance when introducing a dzd_z-dimensional latent representation?
  • What is the motivation for using a chart autoencoder instead of a variational autoencoder (VAE), which is widely adopted in image and video domains?
  • What values of wpw_p, wKw_K, and dzd_z are used in the main benchmarks? Are these values consistent across all benchmarks?
  • How sensitive is the performance to dzd_z, wpw_p, wKw_K, solver tolerance, and the condition number mm? Is it necessary to set dzd_z exactly to six, matching the degrees of freedom of KK?
9 comments

We thank the reviewer for the constructive feedback.

[CW1]: Benefits of Coordinate-Space Parameterization

On the six-dimensional setting. Although the original controlled ablation uses a six-dimensional coefficient generator and a six-dimensional chart, MPDM is not provided with the generating coordinates or basis functions. As specified in Appendix A, the chart autoencoder learns its representation solely from training coefficient fields. More importantly, exact dimensionality matching is not required by MPDM. We have clarified this architectural distinction in revised Section 3 (Field-Space vs. Coordinate-Space Parameterization). In the three main PDE benchmarks, the coefficient generators use 12 variables, while MPDM employs a learned chart with dz=16d_z=16, demonstrating that the six-dimensional match in the controlled ablation is not an architectural requirement.

 

**On the marginal improvement over MPDM$_{\mathrm{field}}.∗∗Theoriginalsubmissionalreadymotivatedcoordinate−spacepredictioninSection3andincludedanindependentquantitativefield−versus−coordinatestudyinAppendixE,wherethetwoestimatorsusethesameregressionmodelanddifferonlyinwhethertheypredictthefullfieldoritsintrinsiccoordinates.Weagree,however,thattheoriginalsix−dimensionaldiffusionablationshowedonlyasmallgapbetweenMPDMandMPDM.** The original submission already motivated coordinate-space prediction in Section 3 and included an independent quantitative field-versus-coordinate study in Appendix E, where the two estimators use the same regression model and differ only in whether they predict the full field or its intrinsic coordinates. We agree, however, that the original six-dimensional diffusion ablation showed only a small gap between MPDM and MPDM_{\mathrm{field}}$. We therefore extended the controlled study to intrinsic dimensions of 10 and 15, with results added to Appendix A (new Table 7):

dintd_{\mathrm{int}}ModelEden↓\mathcal{E}_{\mathrm{den}}\downarrowJoint SW1↓\mathrm{SW}_1\downarrow
6MPDM$_{\mathrm{field}}$0.55980.3094
6MPDM0.55760.3389
10MPDM$_{\mathrm{field}}$0.72980.3738
10MPDM0.68710.3980
15MPDM$_{\mathrm{field}}$0.99850.3393
15MPDM0.81440.3539

The relative reduction in Eden\mathcal{E}_{\mathrm{den}} increases from 0.4% (6D) to 5.9% (10D) and 18.4% (15D). This shows that the near-parity observed in the original 6D setting does not persist as the intrinsic dimensionality increases, consistent with the larger MPDM--MPDM$_{\mathrm{field}}$ gaps observed in the full PDE benchmarks. This advantage remains task-specific: MPDM$_{\mathrm{field}}$ retains lower Joint SW1\mathrm{SW}_1 in these unconditional generation comparisons. We have added the corresponding discussion in revised Section 5.3.

 

On effectiveness in the main PDE benchmarks. Building on the original field-versus-coordinate study in Appendix E, we have further added MPDM$_{\mathrm{field}}$ results to the three full PDE benchmarks in revised Tables 1–3. Both variants employ the same numerical renderer:

BenchmarkMethodState MSE ↓\downarrowMSEK↓\mathrm{MSE}_K\downarrowICSK↓\mathrm{ICS}_K\downarrow
DarcyMPDM$_{\mathrm{field}}$0.034971.00256.23
MPDM0.0066510.557018.37
HelmholtzMPDM$_{\mathrm{field}}$5.64×10−85.64\times10^{-8}3.3510.01159
MPDM7.704×10−97.704\times10^{-9}1.0920.005532
Navier–StokesMPDM$_{\mathrm{field}}$0.009370.057970.3384
MPDM0.0032670.037550.1883

Here, State MSE denotes MSEp\mathrm{MSE}_p for Darcy and Helmholtz and MSEvorticity\mathrm{MSE}_{\mathrm{vorticity}} for Navier–Stokes. MPDM consistently reduces state reconstruction, coefficient, and inverse-consistency errors. In particular, the improvements in MSEK\mathrm{MSE}_K and ICSK\mathrm{ICS}_K provide evidence of more accurate parameter inference beyond the physical consistency inherited from the shared solver. The corresponding analysis has been added to Section 5.2.

 

On computational efficiency. The original Tables 1–3 already report peak memory. We have additionally included relative wall-clock measurements in Appendix B.3 (new Table 8):

BenchmarkMethodRMtrain↓\mathrm{RM}_{\mathrm{train}}\downarrowRTtrain↓\mathrm{RT}_{\mathrm{train}}\downarrowRTinfer↓\mathrm{RT}_{\mathrm{infer}}\downarrow
DarcyMPDM$_{\mathrm{field}}$0.92920.77081.2545
MPDM0.59420.18151.0068
HelmholtzMPDM$_{\mathrm{field}}$1.74401.32151.6210
MPDM1.43800.43142.0349
Navier–StokesMPDM$_{\mathrm{field}}$0.83761.29814.3269
MPDM0.81552.08013.9565

All ratios are normalized to DM. Compared with MPDM$_{\mathrm{field}}$, MPDM reduces training memory on all three PDEs, training time on Darcy and Helmholtz, and inference time on Darcy and Navier--Stokes. These results show that coordinate-space parameterization improves reconstruction accuracy and memory efficiency, with runtime gains depending on the PDEs.

[CW2]: Limitations of Differentiable Solvers

On solver applicability. The original three PDE benchmarks employ differentiable numerical solvers, but solver-rendered clean-state construction does not inherently require solver differentiability. To investigate this distinction, we extend MPDM to seismic first-arrival travel-time modeling, where heterogeneous subsurface slowness fields determine travel-time fields under different source locations. The forward model uses Dijkstra's shortest-path algorithm, whose discrete path selection is nonsmooth when optimal paths change and does not provide a conventional automatic-differentiation backward pass.

 

On handling non-differentiable solvers. We replace solver backpropagation with zeroth-order finite-difference gradient estimates obtained from forward evaluations. MPDM retains its original parameterization: compact coordinates are decoded into slowness fields, and the black-box Dijkstra solver renders the corresponding travel-time fields. This preserves consistency with the embedded discrete solver while enabling state-reconstruction supervision without differentiating through its implementation. We have clarified this extension in revised Section 3 and detailed the approach in Appendix F.

 

Additional experimental evidence. We have added a source-location-shifted OOD evaluation in Appendix F (Table 12), with a summary in revised Section 5.3:

MethodMSEp↓\mathrm{MSE}_p\downarrowMSEK↓\mathrm{MSE}_K\downarrowMAR↓\mathrm{MAR}\downarrowICSK↓\mathrm{ICS}_K\downarrow
DM4.582×10−34.582\times10^{-3}5.672×10−35.672\times10^{-3}8.617×10−28.617\times10^{-2}9.041×10−49.041\times10^{-4}
PISD6.349×10−46.349\times10^{-4}2.988×10−42.988\times10^{-4}3.289×10−33.289\times10^{-3}2.535×10−42.535\times10^{-4}
MPDM$_{\mathrm{field}}$1.394×10−31.394\times10^{-3}6.029×10−36.029\times10^{-3}2.230×10−82.230\times10^{-8}8.783×10−48.783\times10^{-4}
VAE-K + Solver5.810×10−55.810\times10^{-5}3.298×10−43.298\times10^{-4}2.277×10−82.277\times10^{-8}2.579×10−42.579\times10^{-4}
MPDM2.025×10−52.025\times10^{-5}1.568×10−41.568\times10^{-4}2.288×10−82.288\times10^{-8}1.810×10−41.810\times10^{-4}

MPDM achieves the lowest state reconstruction, coefficient, and inverse-consistency errors among the compared methods, while maintaining numerical consistency with the black-box solver. These results demonstrate that MPDM can accommodate non-differentiable numerical renderers through gradient estimation, extending its applicability beyond the differentiable PDE solvers evaluated in the original submission.

[CW3]: Structural Inability to Represent the Optimal Denoiser

On the optimal denoiser. The reviewer's observation holds for unconstrained MSE minimization: the posterior mean of a nonlinear solver-rendered field need not lie on the admissible manifold. However, MPDM explicitly restricts clean-state predictions to the solver-rendered set, yielding a different, well-defined constrained estimation problem. Let μt=E[x0∣xt,E]\mu_t=\mathbb{E}[x_0\mid x_t,\mathcal E] denote the ambient posterior mean and ∥⋅∥W\|\cdot\|_W the channel-weighted norm in Eq. (14). For any admissible prediction y∈M~Ey\in\widetilde{\mathcal M}_{\mathcal E},
 
E[∥y−x0∥W2∣xt,E]=E[∥μt−x0∥W2∣xt,E]+∥y−μt∥W2.\mathbb{E}[\|y-x_0\|_W^2\mid x_t,\mathcal E]=\mathbb{E}[\|\mu_t-x_0\|_W^2\mid x_t,\mathcal E]+\|y-\mu_t\|_W^2.
 
Consequently, the ideal constrained estimator minimizes the distance to μt\mu_t over the solver-rendered set, with an additional minimum achievable risk of dW2(μt,M~E)d_W^2(\mu_t,\widetilde{\mathcal M}_{\mathcal E}). This establishes the trade-off between solver-consistent prediction and unconstrained Bayes optimality, rather than assuming that the posterior mean remains physically admissible. The result characterizes the constrained estimator class without assuming that MPDM explicitly performs metric projection or achieves its population optimum. We have formalized this analysis in revised Appendix C.4 (Proposition 4).

[CW3]: Structural Inability to Represent the Optimal Denoiser

On the DDPM reverse transition. We further quantify the effect of constrained prediction on the reverse transition. Under the standard DDPM posterior parameterization, the reverse mean is affine in the clean-state estimate. Let at=αˉt−1βt/(1−αˉt)a_t=\sqrt{\bar{\alpha}_{t-1}}\beta_t/(1-\bar{\alpha}_t). The deviation between the reverse mean constructed from MPDM's prediction x^0\hat{x}_0 and the true conditional reverse mean is

Δmt=at(x^0−μt).\Delta m_t=a_t(\hat{x}_0-\mu_t).

This identity explicitly quantifies how clean-state estimation error propagates to the reverse mean, including the potential contribution of the manifold constraint. Solver consistency does not imply an unbiased reverse transition or exact distribution learning. Our claim is that MPDM enforces parameter--state consistency in its clean predictions, while the reverse-mean deviation is explicitly characterized rather than assumed to vanish. We have clarified this distinction in revised Section 4 (Training Objective and Sampling) and formalized it in Appendix C.4 (Proposition 5).

 

On the VAE-K + Solver comparison. Section 5.3 and Appendix A already acknowledge that VAE-K + Solver outperforms MPDM in unconditional generation, achieving a lower Joint SW1\mathrm{SW}_1 (0.0783 vs. 0.3389 in Table 4). However, this difference cannot be attributed specifically to posterior-mean bias, since the methods employ different generative objectives and parameter-learning mechanisms. VAE-K + Solver learns a coefficient prior for unconditional generation, whereas MPDM targets joint physical-field denoising and parameter inference. These results reflect different task-specific strengths rather than establishing a causal link between constrained denoising and inferior distribution learning. We therefore distinguish MPDM's demonstrated OOD reconstruction benefits from claims of universal superiority in unconditional generation.

[CW4]: Inverse Stability

On forward versus inverse stability. Proposition 3 in Appendix C.3 characterizes forward error propagation, whereas inverse stability concerns the recovery of coefficients from observed states. These are distinct properties, and a smaller forward Lipschitz constant does not necessarily imply better inverse conditioning. To address the latter, we have extended the theoretical analysis in revised Appendix C.5 (Local Conditional Inverse Stability), establishing local inverse-stability guarantees for the physical renderer restricted to the learned coefficient manifold.

 

On theoretical inverse stability. Let JR=RE′(K0)J_R=\mathcal R_{\mathcal E}'(K_0) and QDQ_D denote an orthonormal basis of the decoder tangent space at K0=Dϕ(z0)K_0=\mathcal D_\phi(z_0). Under local smoothness, full decoder rank, and restricted observability βM=σmin⁡(JRQD/m)>0\beta_M=\sigma_{\min}(J_RQ_D/\sqrt m)>0, Proposition 6 establishes that sufficiently nearby coefficients on the learned manifold satisfy

≤(βM−1+ϵ)∥RE(K1)−RE(K2)∥p.\leq (\beta_M^{-1}+\epsilon) \|\mathcal R_{\mathcal E}(K_1)-\mathcal R_{\mathcal E}(K_2)\|_p.

Moreover, the corresponding local linearized inverse-stability constants satisfy

CM,linK=1σmin⁡(JRQD/m)≤1σmin⁡(JR/m)=CF,linK.C_{M,\mathrm{lin}}^K=\frac{1}{\sigma_{\min}(J_RQ_D/\sqrt{m})}\leq\frac{1}{\sigma_{\min}(J_R/\sqrt{m})}=C_{F,\mathrm{lin}}^K.

This result shows that restricting coefficient perturbations to the learned tangent space can exclude weakly observable directions and improve local inverse conditioning. Corollary 2 further gives a recovery bound accounting for observation noise, decoder representation error, and data-fitting error. Together, these results connect coordinate-space parameterization with inverse conditioning and parameter recovery.

 

On empirical evidence. We have additionally included controlled inverse-stability experiments in revised Appendix C.5, varying nonlinear interactions and spatial complexity through λ∈{0,0.3,0.6}\lambda\in\{0,0.3,0.6\} while keeping the underlying coefficient dimension fixed. We compare local inverse-conditioning constants and state-conditioned coefficient recovery:

λ\lambdaMedian CF,linK↓C_{F,\mathrm{lin}}^K\downarrowMedian CM,linK↓C_{M,\mathrm{lin}}^K\downarrowMedian CFK/CMK↑C_F^K/C_M^K\uparrowField MSEK↓\mathrm{MSE}_K\downarrowMPDM MSEK↓\mathrm{MSE}_K\downarrow
088.164.03621.630.0066780.006980
0.383.284.31819.170.0250550.016745
0.678.194.70816.200.0544400.042221

The learned manifold yields 16–22$\times$ smaller local inverse-stability constants across the tested settings. State-conditioned sampling also improves coefficient recovery in the two more complex cases, while remaining comparable in the simplest case. Together, these results show that learned coordinate restrictions can improve local inverse conditioning and parameter recovery.

Comments on the paper

2 comments

Dear reviewers,

This is a friendly reminder to read the authors’ rebuttals and post your responses by Nov 18. Please briefly indicate which concerns have been addressed and which remain unresolved. Even if your overall assessment remains unchanged, please provide a brief response to the authors. Please also update your reviews and scores as appropriate. Thank you for your time and effort!

AC.

AuthorsOpenReview ↗

Rebuttal Summary

We sincerely thank all reviewers for the thoughtful and constructive discussion. The rebuttal has resulted in a positive convergence on the main technical questions and has helped further clarify the evidence, scope, and positioning of MPDM without changing its central formulation.

Encouragingly, R1 (P4Vh) and R2 (cwNK) explicitly stated that their main concerns had been addressed and updated their ratings from Weak Reject to Weak Accept. R3 (bydV) also indicated that the rebuttal provided sufficient information to finalize the assessment. We briefly summarize the main points below.

1. Learned contribution beyond solver access — R2, R3

MPDM does not use solver consistency itself as evidence of successful inference: the original evaluation separately measures state reconstruction, coefficient recovery, physical residual, and inverse consistency. To further isolate the learning contribution, we added DM + Solver and solver-matched MPDM$_{\mathrm{field}}$ comparisons under identical numerical solvers and physical conditions. The results confirm that MPDM's gains arise from learned coefficient inference and structured denoising, rather than solver access alone (Tables 1–3; Appendix H, Table 14).

2. Coordinate-space parameterization — R1, R3

MPDM does not require prior knowledge of the intrinsic coefficient dimension; dzd_z is a representation-capacity hyperparameter rather than an assumed ground-truth dimension. We clarified this point and further supported it with intrinsic-dimensionality experiments and a fixed-generator sensitivity study over different dzd_z. MPDM remains advantageous over MPDM$_{\mathrm{field}}$ across the tested settings (Section 3; Appendix A, Tables 7–8).

3. Theoretical justification — R1

The discussion distinguished solver consistency, constrained denoising, and inverse recoverability as separate questions. We therefore formalized the constrained Bayes estimation gap and its effect on the DDPM reverse mean (Appendix C.4), and established local conditional inverse-stability results (Appendix C.5). The corresponding experiments show approximately 16–22× better local inverse conditioning (Table 10), with the scope of these guarantees stated explicitly.

4. Applicability and computational scope — R1, R2, R3

MPDM is not inherently restricted to differentiable 2D solvers. We demonstrate black-box solver integration using a Dijkstra travel-time renderer with finite-difference gradient estimation (Appendix F, Table 13) and extend the framework to 3D Darcy flow (Appendix H, Table 15). We also report wall-clock measurements to characterize the solver-dependent accuracy–efficiency trade-off (Appendix B.3, Table 9).

5. Novelty and methodological positioning — R2, R3

MPDM's contribution is not simply combining a latent representation with a numerical solver. Its key distinction is to embed compact coefficient inference and solver-rendered parameter–state clean prediction directly into joint-field diffusion training and reverse denoising. We make this distinction explicit relative to DFM, LD-DIM, L-DPS, and GenPANIS (Section 2; Appendices G and I) and clarify the precise meaning of manifold projection and coordinate-space denoising.

Overall, the discussion has reinforced the paper's central contribution: when part of a structured physical output is deterministically induced by known physics, that dependency need not be relearned as an unconstrained statistical relationship. MPDM instead learns the independent physical structure and delegates dependent-state construction to the known operator.

We are grateful for the reviewers' careful engagement and the positive resolution of the discussion. We believe the clarified formulation and strengthened empirical and theoretical evidence further establish MPDM as a meaningful direction for hybrid scientific generative modeling, and we respectfully hope they support a favorable final assessment.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.