Neural Networks Crack the Secret of Tracking Chaos Without an Ensemble

For decades, one of the central dogmas of weather and climate prediction has been that you cannot accurately track a chaotic system without an ensemble. To know where the errors are growing, forecasters run dozens of parallel forecasts, each nudged slightly differently, and use the spread among them to estimate the uncertainty of the day. Now a team of researchers in France and the United Kingdom has shown that a neural network can match the accuracy of the best ensemble-based methods using only a single forecast trajectory, and has gone a long way toward explaining why this seemingly impossible feat works. The study, published in Nonlinear Processes in Geophysics, could reshape how the next generation of data assimilation systems is designed.

The work, led by Marc Bocquet of CEREA, ENPC, EDF R&D and Institut Polytechnique de Paris, together with Tobias Sebastian Finn and Sibo Cheng of the same laboratory and Alban Farchi of the European Centre for Medium-Range Weather Forecasts, builds on the concept of data assimilation networks, or DANs. Data assimilation is the process by which observations of the real world are continually folded into a numerical model to keep its trajectory anchored to reality. In chaotic systems, where tiny errors amplify exponentially, this correction must happen frequently and intelligently. Classical methods such as the ensemble Kalman filter, or EnKF, rely on an ensemble of forecasts to represent the flow-dependent error statistics that tell the analysis step which parts of the state are most uncertain.

What makes the DAN approach conceptually radical is what the network is trained to do. In earlier machine learning applications to data assimilation, neural networks were trained to imitate the outputs of an existing scheme, essentially learning to copy a Kalman filter or a variational method. A DAN, by contrast, is embedded directly in the forecast–analysis cycle and trained against only two things: the true trajectory of the system and the noisy, sparse observations of it. The network must discover, from scratch, how to combine a forecast state with incoming observations to produce a stable and accurate sequential estimator. It is not taught how a good assimilation algorithm works; it must invent one.

The surprise, first reported by the group in 2024, was that a DAN trained this way could match a well-tuned EnKF even when its architecture was reduced to process a single forecast state rather than an ensemble. That result flew in the face of long-standing practice. Operational weather centres such as ECMWF run ensembles of variational analyses precisely because single-state methods like 3D-Var and 4D-Var, which use static background error covariances, systematically underperform ensemble methods on chaotic dynamics. In the Lorenz-96 model, a standard benchmark representing a mid-latitude atmospheric circle, a well-tuned EnKF achieves a time-averaged analysis error of roughly 0.18 to 0.20, while classical single-state methods stall near 0.40. The DAN, with no ensemble at all, reached the EnKF level.

The new paper dissects the mechanism behind this performance. The key insight is that the learned analysis operator implicitly reconstructs an effective analysis error covariance matrix from the forecast state alone. When the researchers linearised the network’s response to the projected innovation, the difference between the observations and what the forecast predicts, the resulting Jacobian behaved like a flow-dependent covariance matrix remarkably close to that of a well-tuned EnKF, particularly in its dominant eigenvectors, which carry most of the uncertainty. In other words, the network had learned, without ever being told, to estimate the errors of the day from a single state, something that was thought to require an ensemble.

The theoretical underpinning, the authors argue, lies in a generalised form of the multiplicative ergodic theorem, originally due to Oseledec. For an ergodic chaotic system, the unstable subspaces along which errors grow are measurable functions of the state itself. Viewing the entire sequential assimilation process as its own random dynamical system, there exists a mapping from the forecast state to the dominant error-growth directions, and hence to an effective forecast error covariance. Recent work by Sacco and colleagues and by Sakov has shown that such a state-to-covariance map can be exploited explicitly to build competitive filters from a single trajectory. The DAN, it turns out, learns this map implicitly.

Crucially, the network does not memorise global configurations of the state. The researchers showed through scaling experiments that the learned operator extracts local patterns with a limited spatial range, a property made natural by its residual convolutional architecture. When they trained a DAN on the standard 40-variable Lorenz-96 model and applied it unchanged to a doubled, 80-variable version, it performed on par with an EnKF tuned for the larger system. In the limit of large state dimensions, a fixed-size network could not possibly memorise ever more numerous global patterns, so it must be identifying a small set of local structures. The team made these patterns visible by computing a mean marginal gain tensor, a three-index sensitivity of the analysis to perturbations of the forecast state and the innovation, averaged over the invariant distribution of the dynamics. Exploiting the translational symmetry of Lorenz-96, they showed this tensor must be circulant, and its dominant entries indeed form a compact local pattern that persists across model sizes and across sparse and randomised observation configurations.

The second major advance is a stress test in stronger nonlinearity. By lengthening the interval between observation updates, the team pushed the system well beyond the mildly nonlinear regime, past the Lyapunov time of the model, equivalent to about three days of a real weather forecast, and up to twice that horizon. The EnKF degrades steadily under these conditions and becomes uninformative beyond a certain point. The hardest baseline, the iterative ensemble Kalman filter, which combines an ensemble-derived prior with a nonlinear Gauss-Newton optimisation, remains the most accurate scalable method in such regimes. Remarkably, a computationally boosted DAN matched the IEnKF’s accuracy across this extended range, and could even be trained where the IEnKF became difficult to stabilise, all without an ensemble and without any iterative minimisation at inference time.

The explanation for this second capability is subtle. A DAN restricted to a purely linear response to the innovation, which explicitly encodes the state-to-covariance map, matches the EnKF in the mild regime but degrades faster than the EnKF as nonlinearity increases, exactly as its implicit Gaussian prior would predict. The full DAN, however, learns an effective nonlinear analysis response consistent with a non-Gaussian prior inferred from the forecast state. A scaling diagnostic of the network’s response to amplified innovations shows that, in more nonlinear regimes, the impact of large innovations is systematically stronger than a linear analysis would produce. Both ingredients, the flow-dependent covariance information and the learned non-Gaussian response, are necessary; together they parallel the two pillars of the IEnKF, its ensemble and its iterative solver, compressed into a single deterministic network.

The implications reach beyond benchmark models. The authors point to newly developed end-to-end atmospheric processors, systems that ingest raw observations and predict future observations without any explicit background state, and argue that such systems must be implicitly learning analysis operators in their latent spaces, complete with internal representations analogous to error covariances. This reframes a seeming paradox in recent diagnostic studies of observation-driven processors. At the same time, the authors are careful about limits: ensembles remain essential for probabilistic forecasting, uncertainty quantification and strongly multi-modal posteriors, and the multiplicative ergodic theorem guarantees only that the state-to-covariance map exists, not that it is continuous, local or learnable in every setting. But for point estimation in chaotic geophysical flows, the message is striking. The ensemble, long considered indispensable, may be a computational shortcut rather than a fundamental requirement, and a neural network that has never seen a Kalman gain can rediscover the physics of error growth on its own.

Subject of Research: Learned data assimilation neural networks for chaotic dynamical systems

Article Title: Elucidating the performance of data assimilation neural networks for chaotic dynamics

Article References: Bocquet, M., Finn, T. S., Cheng, S., & Farchi, A. (2026). Elucidating the performance of data assimilation neural networks for chaotic dynamics. Nonlinear Processes in Geophysics, 33(3), 401-424. https://doi.org/10.5194/npg-33-401-2026

Image Credits: AI Generated

DOI: 10.5194/npg-33-401-2026

Keywords: data assimilation, neural networks, chaotic dynamics, ensemble Kalman filter, Lorenz-96, deep learning, error covariance, multiplicative ergodic theorem, nonlinear filtering, numerical weather prediction, machine learning, geophysics