acceptodds
Under review as a conference paper at ICLR 2027

Training Neural PDE Solvers with Partial Observations

Abstract

In many scientific settings, acquiring complete observations of PDE coefficients and solutions can be expensive, hazardous, or impossible. Recent diffusion-based methods can reconstruct fields given partial observations, but require complete observations for training. We introduce Ambient Physics, a framework for learning the joint distribution of coefficient-solution pairs directly from partial observations, without requiring a single complete observation. Unlike self-supervised methods, which create supervision by masking complete observations and then predicting the held-out content; Ambient Physics randomly masks a subset of already-observed measurements and supervises only where measurements exist, so the model cannot distinguish "truly unobserved" from "artificially unobserved", and must produce plausible predictions everywhere. Compared with prior diffusion-based methods, Ambient Physics achieves state-of-the-art reconstruction performance: a 61.82% reduction in average overall error while using 500x fewer function evaluations. We provide theoretical justification for learning directly from partial observations and identify a "one-measurement transition": masking a single already-observed measurement enables learning from partial observations across architectures and measurement patterns. We also extend our empirical evaluation to super-resolution tasks and real-world ERA5 weather fields. Ambient Physics thus enables scientific progress in settings where complete observations are unavailable.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Neural Operators, Neural Simualtors, Partial Differential Equations

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_YsCEOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Ambient Physics, a training framework for neural PDE solvers that reconstructs complete coefficient and solution fields from partial observations of both fields. The authors introduce three models: Ambient Flow, Ambient FNO, and Ambient U-Net.

For the main model, Ambient Flow, the network learns a mapping from partial observations (Aaa,Auu)(A_a a, A_u u) to the complete fields (a,u)(a,u). To prevent the network from simply copying the observed values, an additional random mask BB is applied, and noise is added to the retained observations (BaAaa,BuAuu)(B_a A_a a, B_u A_u u). The resulting network input is (BaAaat,BuAuut,BaAaa,BuAuu,BaAa,BuAu,t)(B_a A_a a_t,B_u A_u u_t,B_a A_a a,B_u A_u u,B_a A_a,B_u A_u,t). Training follows the flow-matching framework, with the loss evaluated on the originally observed entries (Aaa,Auu)(A_a a, A_u u). At inference time, the model reconstructs (a,u)(a,u) using the same masking and conditioning procedure as during training.

The main contributions are:

  1. The paper proposes the Ambient Physics framework for joint reconstruction of coefficient and solution fields from incomplete training observations, using additional masking and clean observed values for reconstruction. Three implementations are provided: Ambient Flow, Ambient FNO, and Ambient U-Net.
  2. The paper studies the effect of additional masking on reconstruction performance across four PDE problems and finds that withholding even a single observed measurement can substantially improve reconstruction compared with training without additional masking.
Critical strengths
  1. The paper provides a training approach for joint reconstruction using only partial training observations. By combining additional masking with explicit conditioning on clean observations, the method achieves impressive results on four types of PDEs.
  2. The Ambient Physics framework also achieves strong results with different backbones, including FNO and U-Net, supporting the value of the framework across architectures.
Critical weaknesses
  1. The algorithmic advances over prior work are limited. Ambient Diffusion and IDCD already include related designs for training with incomplete observations and further masking observed entries, while methods such as DiffusionPDE support joint reconstruction using observations of both fields. Appendix I claims that IDCD does not study the relevant ratios, yet IDCD already includes context–query ratio experiments.
  2. The corresponding baseline values in Table 1 match those in FunDPS, where forward and inverse tasks are evaluated separately using observations from only one field, whereas the proposed method uses observations of both fields. Given that both FunDPS and DiffusionPDE support observations of both fields, the paper does not clearly establish whether these results were obtained under matched observation conditions.
Questions for authors
  1. Given that IDCD already uses context–query masking and studies the corresponding ratios, is the main algorithmic addition over Ambient Diffusion and IDCD essentially the explicit use of clean observations as conditioning inputs?
  2. The baseline results in Table 1 match those in Table 1 of FunDPS, where the corresponding experiments use observations from only one field. Please clarify whether these results were directly quoted. If the original single-field results were used, given that both FunDPS and DiffusionPDE support observations of both fields, could you provide matched comparisons using the same observations of both fields?
  3. Although the paper compares with IDCD in the appendix, it does not provide detailed experimental settings. Could you provide the full experimental details, such as the evaluation metrics and numerical error values?
2 comments
AuthorsOpenReview ↗

Response to Reviewer YsCE

We thank the reviewer for their review and for recognizing that Ambient Physics enables learning directly from partial observations and achieves strong results across PDEs and architectures. We address each concern below.

Ambient Diffusion and IDCD (Weakness 1, Q1):

... IDCD already includes context–query ratio experiments.

Given that IDCD uses context-query masking and studies the corresponding ratios ...

Ambient Physics does not use what IDCD refers to as "context-query masking," and our withholding ratio is a different quantity from IDCD's "context and query mask ratio."

In IDCD, the context and query subsets are resampled at every training step (see their public code: https://github.com/LOGO-CUHKSZ/incomplete-data-complete-dynamics/blob/720c701967f031f317a55d5ee2705d6143aa6f21/ns_uniform_missing/models/sde.py#L105). This means that, given enough training steps, no observation is withheld throughout training; every observation is eventually given to the model as input. IDCD's ratio is this per-step subsampling rate.

In Ambient Physics, each training example has a fixed additional mask throughout training. Withheld observations are never given to the model as input. Our ratio is between withheld and not withheld observations, and studying this, we discover the one-measurement transition. We will highlight these fundamental distinctions in Appendix I.

To show that these distinctions matter empirically, we perform an additional experiment with identical training budgets, architectures, data, and observation masks, only replacing the Ambient Physics objective with IDCD's objective (see Q3 below), and observe a 1.5-2x higher error.

Beyond this, Ambient Physics extends ambient training to widely used (deterministic) neural solvers like FNO and UNet (Table 2), and super-resolution tasks (Table 6).

Baseline: FunDPS (Weakness 2, Q2):

Table 1 includes FunDPS's published results, where they formulate the forward and inverse problems separately. In our work, we formulate the generalized reconstruction problem in which both fields may be partially (or completely) observed.

As an additional experiment, we apply FunDPS (NFE=500) to the reconstruction problem in which both fields are partially observed (3%) and passed in as input simultaneously. We report the results in the table below.

MethodNFEDarcy a / uHelmholtz a / uPoisson a / uNS a / u
FunDPS (Published)5005.18 / 2.4917.16 / 2.1320.47 / 1.998.48 / 3.32
FunDPS + Reconstruction Problem5002.97 / 0.828.06 / 0.5011.63 / 0.955.26 / 0.28
Ambient Flow, both fields12.03 / 0.767.75 / 0.478.00 / 0.496.33 / 1.07

Applying our reconstruction formulation to FunDPS reduces its errors, confirming the benefit of both fields being passed as input simultaneously (shown in Table 2). Still, Ambient Flow remains better on Darcy Flow, Helmholtz, and Poisson (coefficients and solutions), without a single complete training example. We will include these results in the Appendix.

Empirical comparison with IDCD (Q3):

For ERA5 experiments, we use nine variables at 103x120 spatial resolution, but unlike IDCD, which reconstructs 3-hour windows, we reconstruct a single time-step (giving the model strictly less information). Each field is 20% observed. Architecture and optimization match our main experiments. Ambient Physics achieves an MSE of 0.038, while IDCD reports an MSE of 0.250. Figure 18 shows IDCD's published prediction (their Figure 2) with visible speckles consistent with their sampler copying observed points into the output. Ambient Physics' prediction is sharper and without these visible artifacts.

To provide a more rigorous quantitative comparison, we performed an additional experiment by implementing IDCD's objective and sampler in our pipeline. We use the Darcy Flow dataset, the reconstruction problem framing, same training budget (25 epochs), same observation masks (3%), and architecture (UNet). We report the results in the table below.

MethodNFECoefficientSolution
IDCD Objective103.13%1.53%
Ambient Flow12.08%0.76%

Ambient Flow achieves 33.5% lower misclassification rate on the coefficient field and 49.9% lower relative L2 error on the solution field. IDCD has higher error on more than 99% of test samples. Note that IDCD's default sampler uses an ensemble of 10 predictions. We will include these results in Appendix I.

We believe these additional clarifications can directly address the reviewer's concerns, and we welcome any further discussion.

Thank you for the detailed rebuttal and additional experiments.The authors added a comparison experiment between FunDPS and Ambient Flow under matched observation of both fields which address my concerns about fairness in the main experiment. The author provides an additional comparison between IDCD and Ambient Flow on the Darcy problem, addressing my concern about the difference between Ambient Flow and IDCD. But the author's statement about the innovation of Ambient Physics over Ambient Diffusion didn't completely convince me. Considering the author's detailed experiments and results, I will raise the rating from 2 to 3.

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

The paper presents Ambient Physics, a framework for learning to reconstruct coefficient-solution pairs from partial observations. This framework runs opposed to current state of the art methods that require complete observations to train prior models and use posterior processing to get full field predictions from partial observations. The methodology relies upon masking of partial observations during training to jointly learn how to predict in the case of unknown observations, this methodology is dervied from another work Ambient Diffusion.

This method is tested by training on various datasets that contain PDE solutions for various coefficients with PDEs that model several different phenomenona. These experiments are compared against several different baselines that are learned on complete observations, but are evaluated in the case where only partial observations are given. Most of the baselines are very weak as they are evaluated in a manner much different than how they were trained. The DiffusionPDE and FunDPS baselines are more competitive and show the value of the training on partial observations that Ambient Physics provides even when trained on less information per sample, however they don't isolate whether improvement is coming from the architecture, one step inference, or the ambient training. Curiously there is a key baseline missing in Table 1 which is rectified flow trained on complete observations which would give us the best comparison and solidify exactly where the error is coming from.

Beyond these initial baseline experiments, the ambient physics framework is show to extendable to other neural network architectures. These results show that it is indeed a generalizable framework.

Critical strengths

The framework addresses a main practical bottleneck of existing methods, the requirement of complete observations for training, which would be problematic for real world data.

Critical weaknesses

The novelty is lacking. Even though the conditional training is new the core objective that is used is from Ambient Diffusion.

Questions for authors

Do you run an experiments where the partial data is always in the same place, mimicking sensors in the real world?

1 comment
AuthorsOpenReview ↗

Response to Reviewer aMcA

We thank the reviewer for their positive review and for recognizing that Ambient Physics addresses a main bottleneck towards making use of real-world data. We address each concern below.

Isolating the source of the improvement:

In Table 2, a deterministic UNet trained with the Ambient Physics objective outperforms FunDPS. In Table 9, the same architecture trained with the same objective, but used as an unconditional prior with DPS-style guidance (approach used by DiffusionPDE and FunDPS) performs comparably to FunDPS. We also find that increasing the NFE leaves the relative L2 error essentially unchanged. Therefore, the source of the improvement is the Ambient Physics training objective together with conditional training.

A rectified flow trained on complete observations would have access to strictly more information (and requires data our setting assumes is unavailable). In Table 13, we show the performance of an MAE trained with (and without) complete observations. As expected, performance collapses when trained directly on partial observations (up to 99% error), whereas Ambient Flow, trained only on partial observations, achieves 0.47-7.86%.

Novelty (Weakness 1):

As the reviewer notes, conditional training is new and directly drives performance (Table 9). We additionally demonstrate that the Ambient Physics objective extends beyond generative models to widely used neural solvers like FNO and UNet. Alongside these methodological contributions, our empirical findings introduce key phenomena, including the one-measurement transition (Figure 3) and robust extensions to super-resolution (Table 6).

Learning from fixed sensors (Q1):

Yes. In our super-resolution experiments (Appendix F, Figure 15 and Table 6), low-resolution measurements lie on a regular checkerboard-like lattice. Using the translation symmetry, each example is shifted to provide coverage everywhere, and Ambient Physics reconstructs 128x128 fields without a single 128x128 training measurement.

In general, if a point is never observed in any training example, there is no training signal there. Note that Corollary 1 assumes that every point is observed with some positive probability. However, using symmetries (where possible) can provide the required coverage.

We believe these additional clarifications can directly address the reviewer's concerns, and we welcome any further discussion.

Reviewer_kwh9OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes Ambient Physics, a framework for learning neural PDE reconstruction models directly from partial observations. By additionally masking observed measurements, the method enables effective reconstruction without complete training samples. Experiments on four PDE benchmarks demonstrate substantial improvements over existing methods, with additional evaluations on super-resolution and ERA5 weather data.

Critical strengths

Practical significance: The proposed framework addresses an important limitation of existing PDE reconstruction methods by removing the requirement for complete training observations. Strong empirical performance: The method achieves impressive reconstruction accuracy with significantly fewer function evaluations. Extensive experiments across different PDEs, architectures, and observation patterns support its effectiveness.

Critical weaknesses

Limited theoretical guarantees: The theoretical analysis primarily establishes recovery of conditional mean predictions under specific assumptions. It does not fully justify the claimed learning of joint distributions or accurate posterior sampling. Experimental limitations: The main experiments use complete validation fields for checkpoint selection, which partially conflicts with the motivation of learning without complete observations. Additionally, computational efficiency is evaluated mainly through function evaluations rather than actual inference time.

Questions for authors

I have no further questions.

1 comment
AuthorsOpenReview ↗

Response to Reviewer kwh9

We thank the reviewer for their positive review and for recognizing the practical significance and strong empirical performance of our work. We address each concern below.

Theoretical guarantees (Weakness 1, Q1):

We agree with the Reviewer that there is a difference between learning the conditional mean and sampling from the posterior distribution given the observation. Conditional mean gives the best point recovery while posterior sampling is about distribution recovery.

What we learn.: Theorem 1 and Corollary 1 show that the population minimizer of the Ambient Physics objective is the conditional mean of the complete coefficient-solution fields, and that supervising only on partial observations yields the same minimizer as supervising on complete observations (under the assumptions of Corollary 1). This justifies training directly on partial observations.

Connection to posterior sampling: The Ambient Diffusion paper shows that if one has access to the conditional mean (what we learn), there exists an algorithm that can use it for exact posterior sampling (and vice versa). So in terms of learning difficulty, learning the conditional mean and learning the posterior distribution is (in theory) the same. That said, the practical algorithm for doing so is not provided in the Ambient Diffusion or our work. That said, the conditional mean is the reconstruction with the lowest expected squared error, which is the error measured in prior work.

We will highlight this discussion in Section 3. We thank the Reviewer for raising this.

Checkpoint selection (Weakness 2, Q3):

Using only partial observations for checkpoint selection does not change our conclusions. In Appendix H, under Training Seeds and Checkpoint Selection, we repeat the main experiments selecting checkpoints using only partial observations, and results (and conclusions) are essentially unchanged (see Table 8). For Darcy Flow, Helmholtz, and Navier-Stokes, performance on coefficients is equal; for Poisson, performance on coefficients is marginally better (-0.15%). For Helmholtz, Navier-Stokes, and Poisson, performance on solutions is equal; for Darcy Flow, performance on solutions is marginally worse (+0.22%). Still, compared to DiffusionPDE and FunDPS, Ambient Flow with checkpoint selection using partial observations remains the best on every equation and field.

Inference time (Weakness 2, Q2):

We measure inference time on one H100 (with batch size of 1 and FP32). We report the results in the table below.

MethodNFEInference Time
DiffusionPDE2000387.0 s
FunDPS20067.2 s
FunDPS500174.1 s
Ambient Flow (ours)10.02 s

Ambient Flow is roughly 3,000-19,000x faster.

Physical consistency (Q2):

To assess physical consistency, we solve the PDE using the same discretization as the data on the Helmholtz and Poisson reconstructed coefficients, and compare the results with the reconstructed solution. We report the results in the table below (relative L2 error averaged over 1000 test samples).

HelmholtzPoisson
Reconstruction error, ∥∥u^−u∥∥\|\|\hat{u} − u\|\|0.47%0.48%
Consistency ∥∥solver(a^)−u^∥∥\|\|\text{solver}(\hat{a}) − \hat{u}\|\|0.85%0.96%

The reconstructed fields are consistent with each other through the PDE to within 1%.

We believe these additional clarifications can directly address the reviewer's concerns, and we welcome any further discussion.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.