Autonomous regularized bias-aware ensemble Kalman filtering: adaptive tuning of the bias-regularization weight, validity of the linearized bias update, and robust learned bias prediction for time-dependent model errors

The regularized bias-aware ensemble Kalman filter (r-EnKF) of Nóvoa, Racca and Magri (2024) performs real-time combined state, parameter, and model-bias estimation by adding a Tikhonov-type penalty on the inferred bias norm to the assimilation cost functional and by modelling the bias with an echo state network whose Jacobian closes the update equations. As published, the method rests on manual choices and unverified approximations: the bias-regularization weight $\gamma$ is user-defined and fixed although the recovery of small-norm biases is sensitive to its value; the closed-form analysis step is derived under a linearization of the bias about the forecast whose validity limits are unknown; the learned bias model's closed-loop prediction degrades for explicitly time-dependent (quasi-periodic) biases; and the synthetic training dataset for the bias model is designed heuristically. The later literature contains no scheme that adaptively tunes $\gamma$ within bias-aware sequential assimilation, no characterization of the linearization's validity, and only partial progress on the bias-model issues. The problem is to remove these manual choices and unverified approximations so that bias-aware sequential estimation becomes autonomous and theoretically characterized.

Background

Data assimilation (DA) is the Bayesian machinery that combines uncertain observations with an imperfect predictive model to estimate the state and parameters of a physical system; in filtering jargon, the model prediction is the forecast and the corrected estimate after using the data is the analysis. The ensemble Kalman filter (EnKF) represents the state distribution by an ensemble of members and corrects each member with a Kalman-type linear update derived from the sample covariance. Classical DA assumes unbiased errors, but real low-order models — obtained from high-fidelity physics via physical assumptions and numerical approximations — carry systematic errors (model bias, or model error), whose functional form is not known a priori, hence the phrase 'unknown unknowns'. Assimilating unbiased observations into a biased model with a bias-unaware filter yields suboptimal analyses.

Bias handling has a long history. Friedland (1969) introduced the separate-bias Kalman filter, which augments the state with the parameters of an assumed bias model and runs two coupled filters; it requires the functional form of the bias to be specified in advance. In numerical weather prediction, variational schemes correct bias either in model space or observation space (Dee and da Silva, 1998; Dee, 2005), and the operational systems at weather centres still largely rely on parametric or off-line bias models. For sequential estimation, da Silva and Colonius (2020) augmented an EnKF with a low-rank representation of discretization error, again with a pre-defined bias form. Nóvoa and Magri (2022) made the first real-time thermoacoustic bias-aware estimation by using an echo state network (ESN) — a reservoir-computing recurrent network with fixed random internal weights and a trained linear readout, which is a universal approximator of dynamical systems — to estimate the forecast bias on the fly.

The source paper (Nóvoa, Racca and Magri, 2024) unifies these threads into the r-EnKF. Each ensemble member minimizes a cost functional with three weighted norms: the departure from the forecast, the misfit to the observations, and $\gamma$ times the squared norm of the inferred bias $\boldsymbol{b}$ (the bias-regularization weight $\gamma\geq 0$). The regularization favours small-norm bias analyses and provably yields a unique bias, which removes the non-uniqueness ill-posedness of earlier bias-aware schemes. To keep the minimization quadratic, the analysis bias is linearized about the forecast, $\boldsymbol{b}^{\mathrm{a}}\approx\boldsymbol{b}^{\mathrm{f}}+\mathbf{J}^{\mathrm{f}}\mathbf{M}(\boldsymbol{\psi}^{\mathrm{a}}-\boldsymbol{\psi}^{\mathrm{f}})$, where $\boldsymbol{\psi}$ is the augmented state-parameter vector, $\mathbf{M}$ maps it to the bias model input, and $\mathbf{J}^{\mathrm{f}}$ is the Jacobian of the bias model — here an ESN, for which the authors derive the Jacobian analytically. The ESN is trained on synthetic bias sequences generated with a multi-parameter strategy and data augmentation (a number $L$ of training trajectories plus amplitude-scaled copies), which requires an educated initial guess of the plausible parameter range. The method is validated on a van der Pol oscillator and a Rijke-tube thermoacoustic time-delay model (a laboratory-scale duct-acoustics model of self-excited oscillations) affected by linear, nonlinear, and explicitly time-dependent bias forms, and is scored by biased and unbiased root-mean-square errors (the RMS of the biased model output versus the truth and of the bias-corrected output versus the truth, respectively).

The weight $\gamma$ controls the trade-off between bias-norm suppression and the magnitude of instantaneous state/parameter updates: with $\gamma\ll 1$ the analysis evolves slowly while tolerating a large bias, whereas large $\gamma$ forces small-norm biases and allows large instantaneous updates to the ensemble state and parameters, so that the analysis can jump far from the forecast in a single step. The published results obtain good analyses only by sweeping $\gamma$ (and $L$) offline for each test case. The paper's conclusions explicitly leave the adaptive tuning of $\gamma$ to future work, and the remaining three design choices — the linearization's validity, the ESN's closed-loop prediction horizon for time-dependent biases, and the systematic construction of the training dataset — are left as demonstrated limitations without solutions.

Problem Statement

For the regularized bias-aware ensemble Kalman filter (r-EnKF) of Nóvoa, Racca and Magri (2024) — sequential assimilation in which each ensemble member minimizes $\|\boldsymbol{\psi}_j-\boldsymbol{\psi}_j^{\mathrm{f}}\|_{\mathbf{C}^{\mathrm{f}-1}_{\psi\psi}}^{2}+\|\boldsymbol{y}_j-\boldsymbol{d}_j\|_{\mathbf{C}^{\mathrm{d}-1}_{\mathrm{dd}}}^{2}+\gamma\|\boldsymbol{b}_j\|_{\mathbf{C}^{\mathrm{b}-1}_{\mathrm{bb}}}^{2}$ with a learned echo-state-network bias model whose Jacobian closes the update — resolve the four interlinked methodological issues the source leaves open, so that bias-aware sequential estimation becomes autonomous (free of manual, problem-dependent tuning) and theoretically characterized:

1. Design a principled scheme that selects or adapts the bias-regularization weight $\gamma$ during assimilation (online, from quantities available at analysis time), replacing the current user-defined fixed value to which the recovery of small-norm-bias analyses is known to be sensitive.
2. Characterize the validity limits of the linearized bias analysis $\boldsymbol{b}^{\mathrm{a}}\approx\boldsymbol{b}^{\mathrm{f}}+\mathbf{J}^{\mathrm{f}}\mathbf{M}(\boldsymbol{\psi}^{\mathrm{a}}-\boldsymbol{\psi}^{\mathrm{f}})$: establish when the analysis remains sufficiently close to the forecast for the closed-form update to be accurate, and determine the impact of the approximation in strongly nonlinear or large-update regimes.
3. Extend the learned bias model's closed-loop (autonomous) predictive horizon to explicitly time-dependent biases, such as the quasi-periodic bias for which the published results show degraded post-assimilation unbiased accuracy, without losing the small-norm-bias guarantee.
4. Replace the heuristic design of the synthetic training dataset (the educated initial guess of the parameter range, the number $L$ of training trajectories, and the augmentation factors) by a systematic, problem-independent methodology for constructing bias training data.

These four components are jointly required and serve one common objective: they are exactly the manual choices and unverified approximations that currently stand between the r-EnKF framework and a deployable autonomous estimator.

References

1. Andrea Nóvoa, Alberto Racca, Luca Magri. Inferring unknown unknowns: Regularized bias-aware ensemble Kalman filter. Computer Methods in Applied Mechanics and Engineering, 418:116502, 2024. DOI 10.1016/j.cma.2023.116502 (arXiv:2306.04315). https://doi.org/10.1016/j.cma.2023.116502
2. Andrea Nóvoa, Nicolas Noiray, James R. Dawson, Luca Magri. A real-time digital twin of azimuthal thermoacoustic instabilities. Journal of Fluid Mechanics, 1001:A49, 2024. DOI 10.1017/jfm.2024.1052. https://doi.org/10.1017/jfm.2024.1052
3. Andrea Nóvoa, Luca Magri. Real-time thermoacoustic data assimilation. Journal of Fluid Mechanics, 948:A35, 2022. DOI 10.1017/jfm.2022.653. https://doi.org/10.1017/jfm.2022.653
4. André F. C. da Silva, Tim Colonius. Flow state estimation in the presence of discretization errors. Journal of Fluid Mechanics, 890:A10, 2020. DOI 10.1017/jfm.2020.103. https://doi.org/10.1017/jfm.2020.103
5. Bernard Friedland. Treatment of bias in recursive filtering. IEEE Transactions on Automatic Control, 14(4):359–367, 1969. DOI 10.1109/TAC.1969.1099223. https://doi.org/10.1109/TAC.1969.1099223
6. Dick P. Dee, Arlindo M. da Silva. Data assimilation in the presence of forecast bias. Quarterly Journal of the Royal Meteorological Society, 124(545):269–295, 1998. DOI 10.1002/qj.49712454512. https://doi.org/10.1002/qj.49712454512
7. Dick P. Dee. Bias and data assimilation. Quarterly Journal of the Royal Meteorological Society, 131(613):3323–3343, 2005. DOI 10.1256/qj.05.137. https://doi.org/10.1256/qj.05.137
8. Alberto Racca, Luca Magri. Robust optimization and validation of echo state networks for learning chaotic dynamics. Neural Networks, 142:252–268, 2021. DOI 10.1016/j.neunet.2021.05.004. https://doi.org/10.1016/j.neunet.2021.05.004
9. Emanuele Bombardi, Andrea Nóvoa, Luca Magri, Alessandro Parente. Transferable inference of turbulence models for urban flows with the Parameter-Regularised Ensemble Kalman Filter. arXiv:2607.03571, 2026. https://arxiv.org/abs/2607.03571
10. Andrea Nóvoa, Luca Magri. Online model learning with data-assimilated reservoir computers. arXiv:2504.16767, 2025. https://arxiv.org/abs/2504.16767
11. Elise Özalp, Andrea Nóvoa, Luca Magri. Real-time forecasting of chaotic dynamics from sparse data and autoencoders. Computer Methods in Applied Mechanics and Engineering, 450:118600, 2026. DOI 10.1016/j.cma.2025.118600. https://doi.org/10.1016/j.cma.2025.118600
12. Simon Weissmann, Neil K. Chada, Claudia Schillings, Xin T. Tong. Adaptive Tikhonov strategies for stochastic ensemble Kalman inversion. Inverse Problems, 38(4):045009, 2022. DOI 10.1088/1361-6420/ac5729. https://doi.org/10.1088/1361-6420/ac5729
13. Igor Kavrakov, Yaswanth Sai Jetti, Ahmet Oguzhan Yuksel, Fehmi Cirak. Statistical finite elements for sequential data synthesis in solid dynamics. arXiv:2604.12688, 2026. https://arxiv.org/abs/2604.12688
14. Alban Farchi, Patrick Laloyaux, Massimo Bonavita, Marc Bocquet. Using machine learning to correct model error in data assimilation and forecast applications. Quarterly Journal of the Royal Meteorological Society, 147(739):3067–3084, 2021. DOI 10.1002/qj.4116. https://doi.org/10.1002/qj.4116