arxiv: 2604.23060 · v1 · submitted 2026-04-24 · 💻 cs.CE · math.OC

Recognition: unknown

Learning to Trust AI and Data-driven models in Data Assimilation through a Multifidelity Ensemble Gaussian Mixture Filter Framework

Andrey A. Popov

Authors on Pith no claims yet

Pith reviewed 2026-05-08 09:02 UTC · model grok-4.3

classification 💻 cs.CE math.OC

keywords data assimilationparticle filtersGaussian mixture modelsmultifidelityAI trustensemble methodsLorenz '96

0 comments

The pith

A multifidelity ensemble Gaussian mixture filter can perform convergent high-dimensional inference by learning to trust AI models over theory-driven ones in undersampled settings.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

This paper seeks to demonstrate that AI and data-driven models, despite risks of producing nonphysical results, can be safely integrated into data assimilation by developing a relative trust measure against slower but reliable theory-driven models. The key is representing this trust through bandwidth scaling factors in kernel density estimates and adapting them with an expectation-maximization procedure. By building this into a multifidelity ensemble Gaussian mixture filter, the approach enables particle filters to achieve high-dimensional convergent inference even when the count of theory-driven samples falls below the system dimension, as tested on a banana distribution and the Lorenz '96 system. This matters for creating faster yet trustworthy forecasting methods in fields like meteorology and engineering where both speed and physical accuracy are essential.

Core claim

The authors establish that bandwidth scaling factors in the kernel density estimates can represent trust between theory-driven and data-driven models, and that these factors can be adaptively computed via expectation-maximization to create multifidelity ensemble Gaussian mixture filters capable of high-dimensional data assimilation in the undersampled regime.

What carries the argument

The multifidelity ensemble Gaussian mixture filter and its adaptive trust version, which blend ensembles from theory-driven and data-driven models using bandwidth scaling factors to encode trust.

Load-bearing premise

Bandwidth scaling factors in the kernel density estimates provide a meaningful, stable, and unbiased way to represent trust that the expectation-maximization procedure can compute without introducing instability or losing physical consistency.

What would settle it

A failure to achieve convergence in the sequential filtering of the Lorenz '96 equations under the undersampled condition, where theory-driven samples are less than the system dimension, would indicate the method does not work as claimed.

read the original abstract

AI and data-driven models have large potential for data assimilation applications by creating fast and accurate forecasts. Their tendency to produce spurious inaccurate, nonphysical results -- hallucination -- however, raises a serious question about their long-term use, and can be categorized as untrustworthy methods. Theory-driven methods on the other hand are slow, but are capable of staying physically realistic due to their mathematical underpinning, and can be categorized as trustworthy methods. We argue that by making use of these methods in tandem, it is possible to build a relative measure of trust between the theory-driven and data-driven methods that results in a combined trustworthy methodology. We argue, and then show, that the bandwidth scaling factors in the kernel density estimates can be used to represent our trust in the theory-driven and data-driven models. We provide for ways in which these measures of trust can be adaptively computed through an expectation-maximization approach. We combine all of these ideas to create the multifidelity ensemble Gaussian mixture filter and its adaptive trust version, which are particle filters capable of high-dimensional data assimilation. We validate our ideas on both a static banana problem and on a sequential filtering example with the Lorenz '96 equations, showing that it is possible to create a particle filter that is capable of high dimensional convergent inference in the undersampled regime -- when the number of theory-driven samples is less than the dimension of the system.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

The paper gives a practical way to encode trust between physics and AI ensembles as adaptive kernel bandwidths inside a multifidelity Gaussian mixture filter, with tests claiming convergence on banana and Lorenz-96 even when samples fall below dimension.

read the letter

The main takeaway is that this work turns the bandwidth scaling factors in the kernel density estimates into explicit, learnable trust weights between theory-driven and data-driven models, then tunes them on the fly with expectation-maximization inside an ensemble Gaussian mixture filter. That produces a particle filter meant for high-dimensional assimilation where the number of reliable physical samples is smaller than the state dimension. They demonstrate the idea on the static banana problem and on sequential filtering with Lorenz '96, reporting that the adaptive version maintains convergent inference in the undersampled regime. The construction is new in how it directly maps relative trust to those bandwidth parameters and embeds the adaptation inside the multifidelity filter loop. It addresses a genuine operational need: letting fast AI forecasts contribute without letting hallucinations dominate. The approach stays grounded in standard kernel mixture and EM machinery, which makes the implementation straightforward for people already working with ensemble DA methods. The validation stays at the level of those two classic low-dimensional tests. No quantitative error tables, convergence rates, or head-to-head comparisons against plain ensemble Kalman or particle filters appear in the description, so it is difficult to judge how large the practical gain actually is. The central risk is exactly the one the stress-test flags: in the N < d regime the mixture likelihood surface is ill-conditioned, and nothing in the abstract shows that the EM update is prevented from collapsing bandwidths on the physical component or from drifting away from the governing equations. Without an explicit safeguard that re-injects the physics into the bandwidth step, the trust weights could become purely statistical fits. This paper is for researchers in geosciences and engineering who are already trying to insert data-driven forecasts into operational assimilation pipelines and need a concrete mechanism for controlling their influence. A reader who wants to experiment with hybrid particle filters will find the construction useful as a starting point. It deserves a serious referee because the problem is timely and the proposed solution is specific enough to be tested and improved, even though the current evidence is still preliminary and the stability questions need direct answers.

Referee Report

3 major / 2 minor

Summary. The paper proposes a multifidelity ensemble Gaussian mixture filter and its adaptive-trust variant for data assimilation. Bandwidth scaling factors in the kernel density estimates are argued to serve as relative trust measures between theory-driven (physically consistent) and data-driven (potentially hallucinating) ensembles; these factors are adaptively computed via an expectation-maximization procedure. The resulting particle filters are claimed to support high-dimensional data assimilation with convergent inference even when the number of theory-driven samples is less than the system dimension. Validation is reported on a static banana problem and on sequential filtering with the Lorenz '96 equations.

Significance. If the EM-tuned bandwidth mechanism can be shown to avoid degeneracy and preserve physical consistency, the framework would offer a principled way to combine fast data-driven forecasts with trustworthy physical models in data assimilation, potentially enabling reliable particle-filter performance in high-dimensional undersampled regimes where standard methods fail.

major comments (3)

[Abstract] Abstract: the central claim that bandwidth scaling factors constitute a stable, interpretable trust measure and that EM adaptation computes them without bias or instability is load-bearing, yet no analysis is supplied of the conditioning of the likelihood surface or of the risk that one bandwidth collapses to zero (discarding the theory-driven ensemble) in the N < d regime.
[Abstract] Abstract: the validation statements assert 'convergent inference' on the banana problem and Lorenz '96 in the undersampled regime, but no quantitative error metrics, convergence rates, or sensitivity analysis to post-hoc parameter choices are referenced, leaving the empirical support for the high-dimensional claim unquantified.
[Abstract] Abstract (adaptive version): no mechanism is described for re-injecting the governing equations into the bandwidth-update step of EM; the procedure therefore remains purely statistical and offers no safeguard against loss of physical realism when the data-driven component dominates the mixture likelihood.

minor comments (2)

[Abstract] Abstract: the precise definition of the multifidelity ensemble Gaussian mixture filter (how the two ensembles are combined before the mixture step) should be stated explicitly rather than left implicit.
[Abstract] Abstract: the term 'hallucination' is used without a formal definition in the data-assimilation context; a short clarification would improve readability.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for their constructive and detailed comments. We address each major comment point by point below, acknowledging where the manuscript requires strengthening and outlining specific revisions.

read point-by-point responses

Referee: [Abstract] Abstract: the central claim that bandwidth scaling factors constitute a stable, interpretable trust measure and that EM adaptation computes them without bias or instability is load-bearing, yet no analysis is supplied of the conditioning of the likelihood surface or of the risk that one bandwidth collapses to zero (discarding the theory-driven ensemble) in the N < d regime.

Authors: We agree that a formal analysis of the likelihood surface conditioning and the risk of bandwidth collapse to zero is absent from the current manuscript. Our numerical experiments on the banana problem and Lorenz '96 demonstrate practical stability of the adaptive EM procedure in the undersampled regime, but this does not substitute for theoretical characterization. We will add a dedicated subsection in the methods or appendix that examines the EM update equations, discusses the conditioning of the mixture likelihood, and introduces a small regularization term on the bandwidths to mitigate collapse risk, with supporting analysis and additional high-dimensional tests. revision: yes
Referee: [Abstract] Abstract: the validation statements assert 'convergent inference' on the banana problem and Lorenz '96 in the undersampled regime, but no quantitative error metrics, convergence rates, or sensitivity analysis to post-hoc parameter choices are referenced, leaving the empirical support for the high-dimensional claim unquantified.

Authors: The full manuscript reports quantitative RMSE and other error metrics for both test problems, along with comparisons against standard ensemble filters. However, the abstract is concise and does not cite these metrics or convergence behavior. We will revise the abstract to include key quantitative results (e.g., error reduction rates in the N < d regime) and add an explicit sensitivity analysis subsection to the numerical experiments that varies post-hoc parameters such as initial bandwidths and EM iteration counts. revision: yes
Referee: [Abstract] Abstract (adaptive version): no mechanism is described for re-injecting the governing equations into the bandwidth-update step of EM; the procedure therefore remains purely statistical and offers no safeguard against loss of physical realism when the data-driven component dominates the mixture likelihood.

Authors: The theory-driven particles are always propagated using the governing equations, so physical consistency is enforced at the ensemble generation stage. Nevertheless, we acknowledge that the EM bandwidth update itself is data-driven and lacks an explicit physical constraint. We will revise the description of the adaptive algorithm to clarify this distinction and add a discussion of potential safeguards, including a proposed modification that augments the EM objective with a physical consistency penalty derived from the model residuals. revision: partial

Circularity Check

0 steps flagged

No circularity; trust measure is an explicit ansatz validated on independent test problems

full rationale

The paper explicitly argues that bandwidth scaling factors in the Gaussian mixture kernels can serve as a relative trust metric between theory-driven and data-driven ensembles, then applies the standard EM algorithm to adapt them. This is a modeling choice, not a derivation in which a claimed result (e.g., physical consistency or convergence) is forced by re-using the same fitted quantities. Validation proceeds via separate numerical experiments on the banana distribution and Lorenz '96, which are external to the definition of the trust weights. No self-citations, uniqueness theorems, or renamings of known results appear as load-bearing steps in the derivation chain.

Axiom & Free-Parameter Ledger

1 free parameters · 2 axioms · 1 invented entities

The central claim rests on interpreting bandwidth scaling factors as trust measures and on the applicability of EM for their adaptation; these elements are introduced in the paper rather than derived from external benchmarks.

free parameters (1)

bandwidth scaling factors
Serve as the explicit trust measures between model classes and are computed adaptively; their specific values are data- and model-dependent.

axioms (2)

domain assumption Gaussian kernel density estimates can represent the posterior in an ensemble filter
Standard modeling choice in Gaussian mixture particle filters.
standard math Expectation-maximization yields stable adaptive estimates of mixture parameters
Well-established algorithm invoked for trust adaptation.

invented entities (1)

trust measure realized as bandwidth scaling factor no independent evidence
purpose: To quantify and adaptively balance trust between theory-driven and data-driven models
Newly proposed mechanism that enables the multifidelity combination.

pith-pipeline@v0.9.0 · 5551 in / 1495 out tokens · 92468 ms · 2026-05-08T09:02:25.735797+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

10 extracted references · 6 canonical work pages

[1]

Afroogh, A

[1]S. Afroogh, A. Akbari, E. Malone, M. Kargar, and H. Alambeigi,Trust in AI: progress, challenges, and future directions, Humanities and Social Sciences Communications, 11 (2024), p

2024
[2]

[2]J. L. Anderson and S. L. Anderson,A Monte Carlo implementation of the nonlinear filtering problem to produce ensemble assimilations and forecasts, Monthly weather review, 127 (1999), pp. 2741–2758. [3]M. Asch, M. Bocquet, and M. Nodet,Data assimilation: methods, algorithms, and appli- cations, SIAM,

1999
[3]

Attia, R

[4]A. Attia, R. Stefanescu, and A. Sandu,The reduced-order hybrid Monte-Carlo sampling smoother, International Journal of Numerical Methods in Fluids, 83 (2016), pp. 28–51, https://doi.org/10.1002/fld.4255, http://dx.doi.org/10.1002/fld.4255. 24A. A. POPOV [5]C. M. Bishop and N. M. Nasrabadi,Pattern recognition and machine learning, vol. 4, Springer,

work page doi:10.1002/fld.4255 2016
[4]

Bocquet, J

[6]M. Bocquet, J. Brajard, A. Carrassi, and L. Bertino,Bayesian inference of chaotic dynamics by merging data assimilation, machine learning and expectation-maximization, arXiv preprint arXiv:2001.06270, (2020). [7]G. Burgers, P. Jan van Leeuwen, and G. Evensen,Analysis scheme in the ensemble Kalman filter, Monthly weather review, 126 (1998), pp. 1719–172...

work page arXiv 2001
[5]

[15]M. B. Giles,Multilevel Monte Carlo path simulation, Acta Numerica, 24 (2015). [16]A. Gregory and C. Cotter,A seamless multilevel ensemble transform particle filter, SIAM Journal on Scientific Computing, 39 (2017), pp. A2684–A2701, https://doi.org/ 10.1137/16M1102021, https://doi.org/10.1137/16M1102021, https://arxiv.org/abs/https: //doi.org/10.1137/16...

work page doi:10.1137/16m1102021 2015
[6]

[23]R. E. Kalman,A new approach to linear filtering and prediction problems, Journal of Basic Engineering, 82 (1960), pp. 35–45, https://doi.org/10.1115/1.3662552, https:// doi.org/10.1115/1.3662552, https://arxiv.org/abs/https://asmedigitalcollection.asme.org/ fluidsengineering/article-pdf/82/1/35/5518977/35 1.pdf. [24]E. Kalnay,Atmospheric modeling, dat...

work page doi:10.1115/1.3662552 1960
[7]

Kapoor and A

[25]S. Kapoor and A. Narayanan,Leakage and the reproducibility crisis in machine-learning- based science, Patterns, (2023), https://doi.org/10.1016/j.patter.2023.100804, https:// www.cell.com/patterns/abstract/S2666-3899(23)00159-9. Publisher: Elsevier. [26]A. Karpatne, G. Atluri, J. H. Faghmous, M. Steinbach, A. Banerjee, A. Ganguly, S. Shekhar, N. Samat...

work page doi:10.1016/j.patter.2023.100804 2023
[8]

[42]A. A. Popov and R. Zanetti,An adaptive covariance parameterization technique for the ensemble gaussian mixture filter, SIAM Journal on Scientific Computing, 46 (2024), pp. A1949–A1971. [43]A. A. Popov and R. Zanetti,The ensemble epanechnikov mixture filter, arXiv preprint arXiv:2408.11164, (2024). [44]S. Reich and C. Cotter,Probabilistic forecasting a...

work page arXiv 2024
[9]

[46]F. A. Silva, C. Pagliantini, and K. Veroy,An adaptive hierarchical ensemble Kalman filter with reduced basis models, SIAM/ASA Journal on Uncertainty Quantification, 13 (2025), pp. 140–170. [47]B. W. Silverman,Density estimation for statistics and data analysis, Routledge,

2025
[10]

[49]S. Yun, R. Zanetti, and B. A. Jones,Kernel-based ensemble Gaussian mixture filtering for orbit determination with sparse data, Advances in Space Research, 69 (2022), pp. 4179– 4197

2022