acceptodds
Under review as a conference paper at ICLR 2027

ForcingDAS: Unified and Robust Data Assimilation via Diffusion Forcing

Abstract

Data assimilation (DA) estimates the state of an evolving dynamical system from noisy, partial observations, and is widely used in scientific simulation as well as weather and climate science. In practice, filtering methods rely on frame-to-frame transition models. However, these models accumulate errors over long horizons, and they are further fragile when observations are non-Markovian. At the same time, learned DA methods typically commit to a single regime, either filtering (nowcasting, real-time forecasting) or smoothing (retrospective reanalysis), which splits what should be a shared prior across application-specific pipelines. To address both issues, we introduce **ForcingDAS**, a *unified* and *robust* DA framework. Built on Diffusion Forcing with an independent noise level assigned to each frame, ForcingDAS learns a joint-trajectory prior instead of frame-to-frame transitions. This allows it to capture long-horizon temporal dependencies and reduce error accumulation. In addition, the same trained model spans the full filtering to smoothing spectrum at inference time. Specifically, nowcasting, fixed-lag smoothing, and batch reanalysis are selected through the inference schedule alone, without retraining. We evaluate ForcingDAS on 2D Navier–Stokes vorticity, precipitation nowcasting, and global atmospheric state estimation. Across all settings, a single model is competitive with or outperforms both learned and classical baselines that are specialized for individual regimes.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Data Assimilation, Inverse Problem, Diffusion Forcing

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_U7xbOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes ForcingDAS, a unified data assimilation framework built on Diffusion Forcing. By introducing observation guidance and different denoising schedules, the method supports filtering, fixed-lag smoothing, and full-sequence smoothing using a single trained model. The authors also propose causality-aware training (CAT) and noise-level-aware observation guidance to improve performance. Experiments on Navier–Stokes, SEVIR, and ERA5 show competitive results against existing DA methods.

Critical strengths

The motivation of unifying filtering and smoothing within a single model is meaningful. Adapting Diffusion Forcing to DA is a natural and promising direction, and the proposed framework makes sense. The experiments across three datasets also demonstrate its potential for different DA tasks.

Critical weaknesses

CAT biases training toward temporally ordered noise levels, which seems more suitable for autoregressive filtering and pyramid schedules. However, full-sequence smoothing uses synchronous denoising, and it is unclear whether CAT is equally beneficial in this setting. This raises a concern about how well CAT supports the paper's central claim of a unified DA framework. Moreover, the CAT ablation only evaluates three-step autoregressive forecasting without observation guidance, rather than the actual DA tasks. Even in these forecasting experiments, CAT does not consistently improve performance: the gains are substantial on ERA5, relatively small on SEVIR, and almost absent on Navier–Stokes.

Noise-level-aware observation guidance is another key component of ForcingDAS. However, the paper does not provide a direct ablation comparing it with standard constant-weight guidance. Therefore, it is unclear how much this design contributes to the final DA performance.

Questions for authors
  1. How does CAT benefit different DA regimes? Since full-sequence smoothing uses synchronous denoising rather than a staggered schedule, why should sorted-noise training also benefit this regime?

  2. Could the authors provide CAT ablations on the actual DA tasks? Comparisons with and without CAT across filtering, fixed-lag smoothing, and full-sequence smoothing would help clarify its contribution. It would also be useful to explain why CAT improves ERA5 substantially but has little effect on Navier–Stokes.

  3. How much does noise-level-aware observation guidance contribute? Could the authors compare it with standard constant-weight guidance under the same model and inference schedules?

My current rating is 2 (Weak Reject), although I lean toward 2.5. I find the overall framework promising, and the experimental results are encouraging. However, the contributions of CAT and noise-level-aware observation guidance are not sufficiently supported by the current ablations. If the authors can convincingly address these concerns, I would consider raising my rating.

Reviewer_q8xUOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

The paper identifies that modern data assimilation pipelines (driven by machine learning models) are made of plenty of downstream tasks (reanalysis, forecasting, filtering, ...) which all share a common dynamical system but condition on different observation segments and let the observation information flow from the past and/or the future. They derive a clean common backbone that cleanly specify to each task by changing a single parameter: the scheduling matrix. The pipeline works on diffusion forcing for the prior model learning the dynamics, and the conditioning part (to assimilate observations) is done with a signal-to-noise aware guidance (variant of DPS). They show good results against baselines, consistently outperforming them.

Critical strengths

Clarity

The paper is quite neat about the framework description, the unified view of different tasks, the description and motivation of each components and the details of algorithmic parts (generation, conditioning, task selection, ...).

Motivation

The explanation are concise but sufficient to follow. We understand why each choice has been made, and what are the practical advantages of each. The unified view for dynamical posterior tasks with the scheduling matrix is neat.

Critical weaknesses

As the review is asked to flag main weaknesses first, I will insist on the three main weaknesses I identified here. Referring to the Questions section for more concerns and interrogations + suggestions. To summarize, I do think this paper has a really interesting view and formulation of the problem of assimilation. I do think that a unified view of inference tasks on dynamical systems is natural and should be the way to design DA pipelines with recent deep-learning techniques. However, the authors claim too many things as novel, but they aren't, making the final contribution very small (mainly a good formulation of the problem).

Huge lack/gap concerning related work and existing efforts in machine learning and data assimilation, especially with diffusion models.

The paper is missing important references to previous work, having very close ideas or claims.

  • Appa [1] is cited but heavily overlooked. Additional details in App is not enough. It builds on SDA (that you compare to, in some settings). You are not the first to propose a unique model/framework that encompass all/plenty of assimilation/posterior tasks on dynamical systems. Appa and WIND (more details below) are a good example. And you work position more as "building on top of these" than being novel. I mention Appa as it applies to weather tasks and really frames and claims the flexibility of the prior mixed with guidance to solve any posterior task. But Appa is just an extension of SDA that you mention and compare with. I think you should change your position given the literature. On top of that, you do not mention it but your guidance loss (5) is extremely close to what SDA defines and uses in practice (their equation (15)). And they introduced it from the exact same observation: the need of adaptive guidance depending on the diffusion schedule.

  • ERDM [2] is not the only paper doing that, but it is a good and strong reference for the concept of diffusion forcing for dynamical systems. Mainly tailored to forecasting task. This work should be mentioned.

  • WIND [3] is a mix of Appa and ERDM. It is very close to Appa in its spirit as they build on top of their ideas (similar system, same guidance algorithm). But they use diffusion forcing to go beyond the blanket mechanism of SDA (Appa) and be able to unroll trajectories using a growing diffusion noise schedule with the trajectory time. So concretely, they are almost exactly using the same tricks as you. The noise schedule is not the same, they use a better guidance (MMPS) which uses the 2nd order Tweedie moment, and they claim the high flexibility of their method to adapt the inference task. I agree that they do not focus on filtering/smoothing/... but they provide, as Appa, a complete framework which works in some settings.

[1] Appa: Bending Weather Dynamics with Latent Diffusion Models for Global Data Assimilation

[2] Elucidated Rolling Diffusion Models for Probabilistic Weather Forecasting

[3] WIND: Weather Inverse Diffusion for Zero-Shot Atmospheric Modeling

As a summary for this point, for me the work is a good formulation and gathering of "best" ideas in the field. But it claims too much novelties which are not at all and miss a clear and fair literature acknowledgment.

The target (posterior distribution, no matter the task) is not correctly evaluated.

When solving posterior sampling problems, and any probabilistic problem, the surrogate should be evaluated accordingly. You have an estimator of the distribution, report other metrics (in the main text) than accuracy/point-wise evaluation. Especially in the assimilation field, there is a collection of metrics of interest that should be reported in this paper without a doubt. To mention a few: CRPS, Skill, Spread, Spread-Skill ratio, ... Which are reported in almost every recent efforts in generative modeling and assimilation (or even forecast focused models).

There exist work on trying to provide insightful evaluation of a posterior surrogate (I encourage the authors to take a look at them [4] [5] among others). I only recommend reporting these although I will not complain if they are not in an assimilation paper. The community needs to do effort towards evaluating correctly what we estimate, and this starts by using such diagnostics.

[4] Sampling-Based Accuracy Testing of Posterior Estimators for General Inference

[5] MIRA: A Score for Conditional Distribution Accuracy and Model Comparison

The baseline methods are very poor and inconsistent between experiments.

This point is linked to the previous ones. The reported baselines are not consistently used (why is SDA only reported in some settings? It can be applied in all of them, right?). You compare to flowDAS, but there are many other works solving these tasks or improving over flowDAS. To mention two of them: FA-APF [6] which is tailored to filtering as flowDAS but claim to be better, and DAISI [7].

[6] Training-Free Bayesian Filtering with Generative Emulators [7] DAISI: Data Assimilation with Inverse Sampling using Stochastic Interpolants

Questions for authors

I will list paper-oriented questions and remarks here that complete the previous section.

Questions

  1. lines 87-88 "SDA target offline reanalysis and cannot run as real-time filters". Can you elaborate? The window model can be sequentially rolled-out over its previous prediction using posterior sampling (see Appa) or masked training.

  2. What about backward in time observation flow tasks? This is usually more like a toy task, or at least less common. This can be viewed as generating an ensemble of initial conditions who lead to the same state, or any other task that might require backward information flow. My questions and concerns are the following: eventually, what the model learns is a prior over the whole trajectory (designed with diffusion forcing) that can be conditioned on observation. The causal nature of the backbone make me doubt about the ability of flowing efficiently information backward in time. Thanks to the gradient of the denoiser in the guidance (which can be assimilated to the covariance via the 2nd order Tweedie formula, as in MMPS [1], and which will never be symmetric in time with a causal architecture), the state kk can be influenced by the way the states <k\lt k and the way they are themselves changed by their observation. Hence a past observation flows to future states. But how does the model adapt a previous state to a future observation? The gradient is 0 because of the causal structure, and the state adapts only to its own observation or its past history.

  3. Jumping on the previous point, why not a causal, anti-causal, and uniform schedule? And why not a non-causal model? Does your method work without a causal mask?

Remarks

  1. The author mentions the forecasting task but do not study it. I think this is the main task of interest. The whole point of assimilation is to estimate states and, practically use this background estimation as an initial condition of a forecast. The flexibility of diffusion prior works for assimilation is that they can perform both the assimilation of past observations and the forecast at the same time, avoiding the traditional multi-step pipeline. You should showcase more asymmetric scenarios such as forecasting. This should be a strength. This is quite linked to your cold-start protocol.

  2. The figures are really small and some parts not even relevant (at leas at this size). The spectrum in Fig.1 is impossible to read. The comparison plots are tiny. It would be nice to show where observations enter in each settings in Fig.2. Fig.3 brings nothing. It is impossible to distinguish difference by eye, I would remove it for better use of space of other figures, or replace it with metrics and comparison.

  3. line 97 "In summary, across all these methods, the regime choice is baked at design or training time". Again, I do not agree for all reasons previously mentioned.

  4. I mentioned it previously, but you should try/compare or comment the advantage of your guidance compared to what can be seen in other pipeline, such as MMPS [1]. And MMPS give an interpretation of the gradient that helps adapting a future state from previous states change, this is proportional to the denoising covariance.

  5. line 261 p(xL∣y1:L)p(x_L \mid y_{1:L}) is not the same as p(xL∣x1:L−1,y1:L)p(x_L \mid x_{1:L-1}, y_{1:L}) without assumptions.

  6. line 350 The reference to table S3 in Appendix is a bit unfortunate. It could be better to include discussed results in the main text, maybe in another form.

  7. Your observation setups are very far from realistic. Could you comment on that and include realistic settings. Similar to FA-APF (mentioned above) which shows (for ERA5 with Gencast) that realistic observation mask (station locations) impact badly (as expected) the generated states. Your down-sampled settings are still very informative for the network.

  8. It could help to have tables closer to the text they correspond to.

[1] Learning Diffusion Priors from Observations by Expectation Maximization

Reviewer_o1TWOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The authors present ForcingDAS, a DA framework that tackles long-horizon error accumulation in transition-based models while unifying different DA regimes (filtering and smoothing). The authors build directly on the Diffusion forcing framework and propose three main changes. Empirical results are provided on systems like PDEs (Navier-Stokes) and ERA5.

Critical strengths

The problem seems well motivated, and the presentation is quite clear. I was able to follow most details. Moreover, the presented method builds directly on Diffusion Forcing and is therefore easy to adopt.

Critical weaknesses

Methodological Contributions: The main methodological contributions of the method seem marginal. Causality-Aware training (CAT) in Section 3.1 defines a new distribution on the noise levels, but its effect on the two benchmarks is quite uncertain (see point 2 below). Observation-guided sampling has already been proposed in works like SDA [1]. Lastly, the scheduling matrix is the same as in Diffusion Forcing but re-interpreted in the context of data assimilation. Can the authors clarify more on this?

Experimental Results: I have a few comments on the experimental results presented in this work.

  1. First, for the results under Causality-Aware Training (CAT) in Section 3.1, the authors use only four trajectories to evaluate whether a sorted sample of noise levels across different frames improves performance. I think this sample size is quite small, and they should use a larger sample. Furthermore, the gains on 2d Navier Stokes and SEVIR-VIL by introducing CAT are marginal at best. Can the authors provide error bars on these results for a more accurate characterization of the claims? Moreover, it would be good to also clarify this in the main text.

  2. Missing baselines: Can the authors compare with additional baselines like [1,2,3] for the Navier-Stokes PDE system? More specifically, [2] uses a single score network for forecasting and assimilation (offline and online); therefore, it's a direct comparison. This also challenges the claim of a “first unified model” in line 120. [3] also tackles the problem of long-horizon modeling via a controller, which is used to stabilize long rollouts and is therefore a natural baseline too. Can the authors add more comparisons with these baselines and discuss key differences with these works in the main text?

References:

[1] DiffDA: a Diffusion Model for Weather-scale Data Assimilation, Huang et al. [2] On conditional diffusion models for PDE simulations, Shysheya et al. [3] Control-Augmented Autoregressive Diffusion for Data Assimilation, Srivastava et al.

Questions for authors

See weaknesses

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.