acceptodds
Withdrawn from ICLR 2027

Certifying Simulation-Based Inference for Nonlinear Scientific Models

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Simulation-based inference (SBI) learns information for posterior inference from simulator outputs when likelihood evaluation is unavailable or expensive. In scientific applications, however, statistical accuracy of the learned posterior does not guarantee that every retained parameter value defines a mathematically well-posed continuous model, nor that an approximate solver is close to that model's solution. We introduce certified SBI, an estimator-agnostic layer that conditions a learned posterior on parameters passing a deterministic model-validity certificate. The framework covers both time-dependent models described by fixed-point maps and steady-state models described by residual equations. For certified parameters, we prove sample-wise existence and uniqueness, residual-to-error bounds for approximate solvers, stability of posterior conditioning, and end-to-end bounds that propagate certification and posterior-approximation errors to downstream posterior summaries and Bayes decisions. The certification interface is compatible with neural posterior, likelihood, or ratio estimation and with classical or learned PDE solvers. A controlled nonlinear inverse problem demonstrates the resulting validity–coverage–cost trade-off.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
Simulation-Based Inference, Scientific Machine Learning, Certified Inference, Bayesian Inverse Problems, Nonlinear Partial Differential Equations

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_ijrbOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper proposes a certification layer that conditions a posterior inferred with SBI to check whether the posterior draw is well-posed, stable, and whether solver-error control is satisfied.

Critical strengths

/

Critical weaknesses

Figures

  • The layout of Figure 1 strongly suggests that it was generated with AI: see typical artifacts in the blue figure. However, the AI use statement did not disclose that generative AI was used to create figures directly; please be more transparent if this is the case, and create the figures yourself.
  • Figure 1: “Learn from N simulations and observed data y”. This statement is technically ambiguous and incorrect. NPE, NLE, and NRE do not use observed data to train a conditional density estimator. Thereby, a conditional density estimator is trained on both the parameters and simulations, which is not clear from the mathematical formulation QN,y(θ)Q_N,y(\theta).
  • Figure 2:
    • Limited experimental evidence of the benefit of the work
    • Black/purple line is not properly visible, nor is it explained what it does.
    • Caption: “inferred from the synthetic observations”: The wording is unclear. Which synthetic observations/experiments were used?

Reproducability

  • No code has been shared
  • No baselines

Unclear language

  • Unclear language; not using jargon from the field.

  • Unclear math, not using standard notations.

  • The paper does not reference any of the three figures within the text.

  • e.g., Line 14-15: “statistical accuracy of the learned posterior does not guarantee that every retained parameter value defines a mathematically well-posed continuous model, nor that an approximate solver is close to that model’s solution.”: Please make this sentence clearer. It is also not very clear what you mean by a “continuous model”.

  • e.g., Line 22-23: “The certification interface is compatible with neural posterior, likelihood, or ratio estimation and with classical or learned PDE solvers.” It is unclear why the certificate is stated to be compatible with SBI methods (NPE, NLE, NRE) and PDE solvers.

  • e.g.: Line 36-37: “Most SBI guarantees and diagnostics focus on the statistical layer: whether the learned posterior approximates the intended Bayesian target, whether inference is calibrated under repeated simulation, and how efficiently simulations are allocated”. The phrasing and corresponding citations are ambiguous and incorrect.

Extremely weak experiments

  • Only 1 experiment has been performed, and it was not even performed on standard SBI methods (e.g., NPE, NLE, and NRE), although the abstract claims to work on those methods.
Questions for authors

/

Reviewer_E2aJOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes restricting posterior sampling in simulation based inference to an area in which posterior parameters correspond to models that have well-defined solutions. This is done using theoretical guarantees in the parameter space, and rejection sampling any posterior samples that are outside those theoretical guarantees. Various general results about such guarantees are presented, and the method is demonstrated empirically in a simple PDE example, although without neural posterior estimation methods.

Critical strengths
  • The idea of using theoretical guarantees for certifying posterior parameters in SBI using theoretical guarantees tailored for the specific cases seems like a good research direction, and could be useful.
Critical weaknesses
  • The paper does contain any neural SBI experiments or comparisons with baselines on posterior estimation performance. Instead, there is only one low-dimensional toy PDE example where the posterior is evaluated on a grid. The paper should have realistic experiments that clearly show the benefit of the method, comparing to baselines. It seems that the SBI literature already has approaches for dealing with invalid simulation outputs, see e.g., [1, 2].
  • As it is, the paper is mostly centered around very general mathematical claims. With the current writing, it is unclear if they have a particular relevance to machine learning or simulation based inference, or whether they are thinly applied to SBI as rejection sampling. If the key results are general mathematical facts, it is unclear if a machine learning venue is the most appropriate place for publication. Appendix F.7. also mentions that the bound assumption in Theorem 5 should be provided by a particular SBI method / benchmark, raising questions about how applicable the end-to-end guarantee provided here is.
  • For the main experiment, it is unclear whether removed posterior mass corresponds to data that is not well-posed. If, e.g., none of the parameter values have well-posedness issues, then the experiment should be changed to a setting where well-posedness issues are real.

[1] https://sbi.readthedocs.io/en/latest/advanced_tutorials/06_restriction_estimator.html

[2] Lueckmann, Jan-Matthis, et al. "Flexible statistical inference for mechanistic models of neural dynamics." Advances in neural information processing systems 30 (2017).

Questions for authors
  • Can the authors clarify which existing and known mathematical results are the Theorems in the paper relying on, for each theorem?
Reviewer_VbH4OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper identifies a potential issue related to the use of SBI with simulators that rely on scientific models. After estimating the posterior it is possible to sample parameters with high posterior probability that do not lead to well-posed models (i.e., models that exist, are unique, and depend stably on the data for that parameter set), or to simulator approximations that are guaranteed close to the models. To address this issue, the paper assumes that we have a deterministic model validity certificate (theorem) and proposes to use that certificate to filter the samples drawn from the posterior. Two alternative routes are offered for implementing the proposed filtering, which depend on the type of certificate available. Route I applies to contraction certificates while the second to residual stability certificates. One controlled Route I example is provided in the paper using a semilinear heat equation where a posterior shift of 0.289 is observed.

Critical strengths
  • The paper pinpoints a potential issue that can occur when applying SBI to non-linear scientific models. These issues can be due to sampled parameters leading to non-unique solutions, or to numerical solver approximations that are not guaranteed close to the model solution.
  • The paper offers a coherent and, as far as I understand, correct formulation offering a methodology allowing confirmation that parameters sampled via SBI are scientifically usable.
Critical weaknesses
  • The paper does not convincingly demonstrate the motivation behind the problem. Beyond stating that statistical calibration and mathematical validity are logically separate questions, it does not analyze how such parameters arise in practice, and it does not use an SBI estimator (in the example, the posterior is evaluated directly on a 21×21 grid).
  • In the demonstrated Route I experiment the certificate (Eq. (17)) is a sufficient condition, but not necessary. If unique solutions exist outside the condition, then this approach would essentially reject perfectly good solutions (see Question 1). No demonstration involving an actual solver exists for Route II (the verification in I.3 is analytic). Several other model types besides the one of Section 5.1 are referenced, but never actually applied.
  • It is not clear how Theorems 5 and 6 could be instantiated for a practical SBI pipeline: how is it possible to guarantee that the divergence is below εN\varepsilon_N? No numerical value for εN\varepsilon_N (or for the other quantities involved) is estimated anywhere. Additional experiments with numerical examples would help here.
  • The proposed method assumes that a usable certificate C is available for the model, but it is not discussed or demonstrated how certificates can be derived and evaluated for realistic scientific models.
Questions for authors
  • Can any parameter in the prior support of the controlled benchmark actually fail to be well-posed?
  • Is it possible to see an end-to-end run with an actual SBI estimator, including the calculation of all quantities mentioned in the paper and measurements related to computational cost?
  • Is it possible to see a Route II experiment involving an actual numerical solver (classical or learned), rather than an analytically evaluated residual?
  • Are theorems C available for scientific models, in practice? If not, can these be derived, easily?

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

Related papers

Discussion (0)

Sign in to comment.