acceptodds
Under review as a conference paper at ICLR 2027

JENO: Full-Field Latent Prediction for Sparse Inverse PDE Inference

Abstract

Whole-field PDE recovery under partial observation requires inferring complete physical states from sparse and incomplete evidence. Deterministic neural operators such as FNO provide efficient one-pass inference, but often struggle under sparse observations because they primarily learn direct input-output regression rather than representations of the underlying physical state manifold. Probabilistic diffusion-based approaches learn stronger generative priors for ill-posed PDE inference, but require expensive iterative sampling during inference. In this work, we propose JENO: Joint-Embedding Neural Operators, a deterministic framework that learns latent operators from partial observations to representations of complete physical fields through joint-embedding predictive learning. Instead of directly regressing solutions in pixel space, JENO learns structured latent representations of the full state. As a result, JENO is not rewarded for just matching observed outputs, but explicitly encouraged to learn a globally consistent representation of the underlying physical state. Importantly, JENO combines the strengths of deterministic neural operators and probabilistic diffusion models. Extensive experiments show that JENO is nearly as efficient as standard neural operators while being more than 2× more accurate. At the same time, JENO is competitive with or surpasses diffusion models in accuracy while being four orders of magnitude faster than diffusion-based approaches. Overall, our results suggest that joint-embedding predictive learning offers a strong non-probabilistic alternative, establishing a new paradigm for PDE inference under partial observations.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
joint-embedding predictive architectures, neural operators, inverse PDE problems, partial observations, self-supervised learning, latent prediction, sparse-to-full reconstruction, scientific machine learning

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_pj23OpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This paper develops an adaptation of the JEPA approach to learn an embedding of both the inputs and targets for PDE field reconstruction. Once a latent representation is learned, the authors use a decoder to make predictions of full fields from sparse observation locations. The proposed method performs well against the (specific) chosen benchmarks.

Critical strengths

The paper proposes a substantively new way of performing PDE inference -- attemtping to adapt JEPA-style approaches in a way I have not encountered before.

Critical weaknesses
  • My biggest concern is in Figure 3 and 4. The sparse observations which serve as the input are clearly numerous enough that a simple unstructured interpolation of the fields would be very good predictions for the full-field, yet the baseline comparisons shown completely fail this task. This leads me to believe these baselines are not appropriate comparisons. Furthermore, standard inverse problem methodology can be used for the inverse task this method is said to tackle, yet this is not included. Unless this is addressed, it is difficult to asses wheather the method is truly beneficial.

  • There is limited discussion of the experimental setup beyond discussion of the architectures themselves, can the experimental setup, the inference task itself, be more clearly discussed?

  • The lack of figures showing predictions of the full-field measurements makes it difficult to trust the output. The inferred fields are physically interpretable and should be included to help the reader asses the quality of inference.

Questions for authors

How was the decoder for prediction trained?

What would be the practical difference if one were to use an autoencoder (either VAE, standard AE, or sliced-Wasserstein/MMD AE)? Would these perform similarly?

Why was deterministic interpolation, or krigging, not compared against?

Reviewer_8EvYOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper introduces JENO, a method for sparse-to-dense PDE recovery (both forward and inverse) using paired training fields. First, a student FNO encoder receives zero-filled sparse inputs, masks, and coordinates, learning to match channel-normalized latents from an EMA teacher that sees complete field pairs. Training uses a multi-level latent loss without explicit field reconstruction. Next, the student encoder is fine-tuned with a task-specific readout under dense L1 supervision for single-pass deterministic inference. On the 128x128 resolution benchmark (Darcy, Poisson, Helmholtz, NS, NS-BCs), JENO achieves the lowest inverse errors on four PDEs and runs in ~5.5 ms, compared to ~300 to 450 s for DiffusionPDE and FunDPS.

Critical strengths
  1. Well-motivated problem and a clean idea. Using the complete field as a privileged teacher signal in latent space, rather than only as a regression target, is natural and simple to implement. Inference stays a single forward pass.
  2. Strong headline numbers at very low cost.
Critical weaknesses

The most dire issue is that there is no amortized conditional diffusion baseline under matched supervision:

Every generative baseline is a zero-shot guided sampler with an unconditional prior. JENO, by contrast, is trained directly on the sparse-to-dense mapping, for each task and at the test density. A conditional diffusion or flow baseline model trained on the same pairs, ideally with the same FNO backbone, is the natural comparison and is indispensable for this paper for the following reasons.

  1. It removes the confound between training paradigms. Is JENO winning because it is deterministic and JEPA-pretrained, or because it is supervised on the exact task while the samplers are not?
  2. It is the paper’s own argument. Sec. 2.3 and C.9 argue that DPS-style guidance introduces a likelihood approximation error. Conditional diffusion has no such approximation, so leaving it out weakens that motivation.
  3. The speed claim depends on it. The four-orders-of-magnitude gap is measured against 200–2000-step guided samplers run one sample at a time. Conditional models with few-step samplers, or consistency/distilled variants, would shrink that gap substantially.

Therefore, JENO needs to compare against conditional diffusion models such as: "Physical-Field Reconstruction from Sparse Observations: When Are Diffusion Models Preferable to Deterministic Regression" "Spatially-Aware Diffusion Models with Cross-Attention for Global Field Reconstruction with Sparse Observations" or "VideoPDE: Unified Generative PDE Solving via Video Inpainting Diffusion Models"

Moreover, the relative L2 loss should not be the only way to compare against generative methods. The regression loss of JENO explicitly encourages it to minimize the relative L2 loss. To be fair to the generative methods, you should report the posterior mean, report distribution metrics, or physical metrics that a regression model will blur out.

Questions for authors
  1. How does your model perform against conditional diffusion models?
  2. Why would a point estimate regression method of JENO is expected to produce valid, sharp output despite uncertainty?
  3. Can you isolate the contribution of the JEPA pretraining?
Reviewer_ykwVOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper proposes JENO (Joint-Embedding Neural Operators), a two-stage framework for recovering complete physical fields from sparse observations in PDE inference. During pretraining, a student network receives sparse measurements, observation masks, and coordinates, and predicts latent representations supplied by a teacher that processes complete paired fields. The teacher encoder is updated through an exponential moving average, while a spatially weighted objective supervises latent prediction over sampled unobserved regions and their complement. The pretrained student encoder is subsequently fine-tuned jointly with a task-specific readout using dense field supervision. At deployment, the model produces a complete field estimate in a single deterministic forward pass. The paper presents a conditional analysis relating latent alignment to downstream prediction error and evaluates the framework on several forward and inverse PDE tasks, alongside studies of inference efficiency, architectural choices, observation density, and measurement noise.

Critical strengths

Well-motivated problem setting. The paper addresses a practically relevant scenario in which complete physical fields are available from offline simulations during training, while deployment relies on sparse measurements. The focus on obtaining a dense point estimate with low inference latency provides a clear motivation for the proposed approach.

Coherent framework with efficient deployment. The combination of sparse-input representation learning, complete-field teacher targets, and supervised downstream adaptation offers a structured way to exploit complete training data. The teacher and pretraining predictor are discarded at inference, retaining a simple deployment path. The reported inference time of approximately 5.5 ms on an H200 for Poisson and Navier–Stokes demonstrates a substantial latency advantage over the particular iterative samplers evaluated (Table 2).

Broad empirical investigation within the benchmark setting. The experiments cover both forward and inverse tasks across multiple PDE families, including Darcy flow, Poisson, Helmholtz, and Navier–Stokes. Additional studies examine different backbones and readouts, varying observation densities, and measurement noise. These experiments provide useful information about the framework’s behavior beyond the main accuracy table and help characterize its practical limitations.

Explicit analysis of the connection between representation learning and field recovery. The theoretical treatment distinguishes normalized latent alignment from downstream prediction accuracy and makes the dependence on teacher decoding quality, latent scale, representation transfer, and fine-tuning explicit. This provides a useful account of the conditions required for improvements in the latent objective to translate into accurate physical-field predictions.

Critical weaknesses
  1. Limited methodological novelty and insufficient justification of the spatial loss. The paper appears to primarily transfer JEPA to neural operator learning, without clearly articulating the core insight specific to this setting. Moreover, separately averaging losses over two regions introduces weighting and spatial-scale choices whose robustness across PDEs and resolutions remains insufficiently established. The existing sensitivity study uses a different training configuration. Ablations under the main training setup should vary the relative loss weight and target-region size across PDEs and resolutions to determine whether the design requires careful task-specific tuning.

  2. Restricted demonstrated applicability to structured grids. The experiments rely on regular grids and grid-based sampling, leaving applicability to unstructured meshes unresolved. Many engineering problems involve irregular geometries and nonuniform meshes, where the current encoder, masking strategy, and spatial loss may not directly apply. The authors should demonstrate on an unstructured-mesh benchmark whether the framework can generalize through minor modifications or requires substantial redesign. Without such evidence, its demonstrated applicability to these practical settings remains limited.

  3. Figure 1 requires substantial revision. Some labels appear to contain irregular spacing, and the overall structure makes the framework difficult to follow. The figure needs corrected typography and clearer organization to communicate the method effectively and meet the presentation standard expected of a conference paper.

Questions for authors

See weakness.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.