acceptodds
Under review as a conference paper at ICLR 2027

Mechanism Learning through Prototype-Anchored Dynamics

Abstract

We study mechanism learning for scientific forecasting: constructing an explicit description of local evolution and learning to use it in prediction. Prototype anchoring builds this dynamics descriptor from a finite bank of reusable support patterns. We instantiate this approach using local predictor coefficients in Burgers, observed increment patterns in WeatherBench 2 (WB2), and learned support vectors in Lorenz96. Three sets of experiments assess predictive utility, the contribution of anchoring, and the descriptor's dynamical meaning. First, across 20 shared seeds on a two-variable WB2 task at h, the full model reduces temperature RMSE by K and Z500 RMSE by m s relative to an approximately parameter-matched direct predictor (paired 95% CIs and ). Both contrasts remain significant after Holm correction. Second, finite support improves temperature forecasts over a learned descriptor generator with 20% and 50% of the training data; its effect varies with target and sample size. Third, Burgers descriptors align with independently computed PDE quantities after history matching and viscosity-pair/time controls (within-stratum partial rank correlation ), and shuffling them increases one-step RMSE by (95% CI ). Matched swaps provide no clear evidence that descriptor changes reproduce the direction of dynamical change. These findings support an empirical form of mechanism learning: descriptors can organize local dynamics and contribute to prediction, while the value of prototype anchoring depends on the forecasting regime.

Primary area:
learning on time series and dynamical systems
Keywords:
mechanism learning, scientific forecasting, prototype anchoring, dynamical systems

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_c8EYOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a prototype-anchored forecasting framework. The main idea of the framework is to construct an explicit dynamics descriptor from retrieved support patterns and use it to condition future predictions. The authors evaluate the proposed framework on the Burgers equation, WeatherBench 2, and Lorenz96. The empirical results show forecasting performance improvements, descriptor–dynamics alignment (to some extent), and the effect of descriptor interventions.

Overall, the paper aims to propose an interesting forecasting architecture with useful initial assessment. However, its methodological novelty is limited. In particular, the mechanistic learning in its title is not sufficiently supported by the experiments. I would encourage the authors to clarify the technical contribution and strengthen the evidence connecting the learned descriptors to dynamical mechanisms.

Critical strengths

The paper has the following strengths:

  1. It attempts to make the representation of local dynamics explicit, rather than relying solely on implicit neural representations.
  2. The experiments include comparisons with different architectures and diagnostic tests of the descriptor's predictive utility and relationship to independently computed dynamics.
  3. The evaluation across three different systems (Burgers, WB2 and Lorentz96) demonstrates the flexibility of the framework.
Critical weaknesses

The paper suffers from the following weaknesses:

  1. It appears to me that the paper simply combines existing ideas, including prototype-based retrieval, descriptor aggregation, and conditional forecasting. However, it does not clearly establish what fundamental technical innovation makes it from these approaches or demonstrate that prototype anchoring consistently provides benefits beyond simpler alternatives.

  2. The experiments show that the constructed descriptors can contribute to forecasting and contain some dynamics-related information, particularly in Burgers. However, the probe results are weak, and descriptor interventions do not establish that changes in the descriptor produce physically consistent changes in predicted evolution. The experiments reveal a useful conditioning representation for prediction more strongly than a faithful representation of underlying mechanisms.

Questions for authors

Related to the two identified weaknesses:

  1. What is the precise technical contribution beyond combining prototype retrieval, dynamics-descriptor construction, and conditional forecasting? Can the authors isolate the benefit of prototype anchoring against other alternatives?

  2. What evidence to demonstrate that the learned descriptors faithfully capture underlying dynamics, rather than merely providing useful predictive features? Can you also show that controlled descriptor interventions produce changes consistent with independently known dynamics and generalise to unseen dynamical regimes?

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

The paper introduces a novel method for dynamic forecasting coined "Mechanism Learning" which seeks to improve prediction quality by incorporating a learned support dictionary into the forecasting framework. The authors introduce a prototype-anchored model "Full" and compare the performance of this model on a number of datasets (including a Burgers equation, WeatherBench 2 and the Lorenz96 system) against a variant that does not include the prototype bank and a "Directly-matched" method in terms of predictive performance measured on RMSE.

Critical strengths

Overall, I think the premise of the authors as described on page 1, to learn "to forecast through an explicit representation of local evolution" is a worthwhile endeavour and I believe there is genuine merit in this approach. This is the case not least because the modelling of local evolution is precisely main idea of differential calculus and, most likely, one of the reasons this toolkit (calculus) has been so successful as a modelling framework in physics and beyond. I also appreciate the authors' efforts of very clearly listing parameter choices in models (Table 1) and describing the datasets used. I think the evaluation datasets are satisfactory for the premise of this work.

Critical weaknesses

Having said the above, I have very significant concerns with the scientific content of the manuscript:

  1. Above all, in my judgment, the paper is extremely difficult to read and it is very challenging to ascertain what exactly "mechanism learning" is in the context described and how it differs or improves upon the state-of-the-art. In particular, the paper does not refer to differential equations beyond using "PDEs" as a means for dataset production which is particularly worrying given the method suggests doing something very similar it would be paramount to compare this approach or at least discuss it.

  2. In line with the above the discussion of the actual method on page 3 is very sparse and it is not clear what function each of the components of the model takes. For example, equation (1) is followed by a note saying that "E supplies forecast features" without ever explaining what nature they take. The description also mentions lt,jl_{t,j} but it is not clear where these points come from.

  3. The authors use non-standard terminology in unusual or non-explanatory contexts, including "Holm correction" in the abstract, "PDE quantities" in the introduction, and "Burgers descriptors" (presumably refering to Burgers equation). I think it would be extremely beneficial if the authors could familiarise themselves with more literature on modelling of time-dependent systems, to help place this work in context and use terminology correctly.

  4. It appears from Table 1 that the way the proposed model was applied varies significantly from one dataset to the next, making it very difficult to extract a coherent approach. The variation includes differences in training objectives, encoder architectures and evaluation regime. As a result I would consider this a very highly parameter tuned approach that might be applicable very much only to the specific dataset consider without any evidence for generalisation.

  5. The numerical experiments in section 4 contain no baselines. Timeseries forecasting in the context of scientific computing is a very standard task and there are a plethora of methods (both classical and ML-based) that one might reasonably compare against. The only baselines used here are variations of the method developed by the authors. This means the method simply cannot be put in comparison against the state-of-the-art.

Questions for authors
  1. On p. 1 the authors refer to "PDE quantities" I wonder if they could clarify what that means?

  2. Given the current approach seems to be to take a PDE model, extract local information for a dictionary and then combine that information back into a forecast, why not use classical numerical methods for the same task? Is there a gain in speed? Is there a gain in accuracy?

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

This paper proposes a prototype-based approach for scientific forecasting. The model combines prototypes based on observed history to construct a local dynamics descriptor to condition predictions. Experiments on Burgers, WeatherBench 2, and Lorenz96 show improvements in prediction, but the value of the prototype bank and the physical meaning of the descriptors could be further clarified.

Critical strengths

This paper explicitly models local dynamics using a prototype bank and can be helpful for data-limited settings. Also this intermediate representation may help trace predictions to learned dynamics and analyze and inspect relationships between different dynamics, which is valuable for scientific applications.

Critical weaknesses

It is unclear how well the descriptor represents dynamics. In Burgers, it captures limited information about viscosity and the advection–diffusion and shows almost no correlation with the advection, diffusion, or tendency terms, performing worse than raw history. Swapping descriptors between matched histories also does not lead to changes consistent with the dynamics. What physical information the descriptor actually captures and how useful it is for helping and interpreting model predictions need further justification.

The benefit of the prototype bank is not sufficiently demonstrated. On the WB2 training set, Full does not significantly outperform NoBank. Neither Burgers nor Lorenz96 includes a NoBank comparison. Further evidence is needed to show the prediction advantage of the prototype bank.

Questions for authors

Does Full outperform a descriptor-free WB2 baseline using the same backbone? Are there NoBank comparisons for Burgers and Lorenz96? Would the WB2 Full–NoBank results change if both methods used the same history representation to construct the descriptor?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.