A team led by Haru Kuroki and corresponding author Kazumune Hashimoto of The University of Osaka, together with colleagues at The University of Tokyo, has developed a hybrid control algorithm they call ensemble-Kalman-guided model predictive path integral control, or EKG-MPPI. The method is designed to answer a deceptively simple question: if you could only nudge the atmosphere at one tiny location, with one small push, where should that push go and how big should it be to suppress an extreme event before it forms? The work sits within a growing research program known as control simulation experiments, in which scientists test hypothetical interventions entirely inside numerical models, using an uncontrolled simulation as a stand-in for the real atmosphere. This simulation-based approach allows researchers to probe the feasibility and side effects of intervention strategies without any real-world risk.

The architecture of EKG-MPPI is a marriage of two complementary control philosophies. The first ingredient is ensemble Kalman control, or EnKC, a technique introduced by co-author Yohei Sawada that cleverly repurposes the ensemble Kalman filter, the workhorse of modern weather data assimilation. In a conventional forecast, the filter blends observations with model predictions to estimate the current atmospheric state. EnKC flips the logic: it treats the desired future outcome, such as keeping wind speeds below a threshold, as a pseudo-observation to be assimilated. The ensemble Kalman smoother then works backward in time to compute the analysis increment, which is interpreted as the optimal perturbation to apply now to steer the system toward that target. Because the computation relies on ensemble statistics, it never requires an explicit covariance matrix, making it feasible for high-dimensional systems.

EnKC, however, has a fundamental weakness. Its mathematics assumes the dynamics are approximately linear over the prediction horizon, an assumption that becomes increasingly shaky in strongly nonlinear, chaotic flows. The second ingredient of the new method, model predictive path integral control, or MPPI, addresses exactly this gap. MPPI is a sampling-based technique with roots in robotics and autonomous driving, derived from a variational free-energy bound. Instead of linearizing the model or computing gradients, MPPI draws hundreds or thousands of candidate control sequences from a Gaussian distribution, rolls each one forward through the full nonlinear model, scores the resulting trajectories with a cost function, and combines them using exponential importance weights. The weighted average becomes the refined control input. Crucially, all the rollouts and weight computations can be run in parallel, which makes the method a natural fit for modern GPU hardware.

The problem with vanilla MPPI in geophysics is sample inefficiency. When the control space has thousands of dimensions, as it does for a 128-by-128 atmospheric grid, naive random sampling almost never discovers the rare, precisely placed perturbations that actually matter. This is where the ensemble Kalman guidance earns its place in the acronym. The researchers apply an adaptive thresholding procedure to the EnKC-derived perturbation, discarding small components as noise and retaining only the largest one, which identifies the single most promising actuator location and its nominal amplitude. That information is then embedded into the mean and covariance of the Gaussian proposal distributions from which MPPI samples. In effect, EnKC provides a physically informed educated guess about where the system is most sensitive, and MPPI polishes that guess using the full nonlinear dynamics. The final control is projected back onto a single-actuator form, applied to both the nature run and every ensemble member, and the forecast-assimilation-control loop repeats at the next assimilation time.

The team tested the method on two models of very different character. The first was the Lorenz-96 model, a 40-variable chaotic system that has become a standard proving ground for data assimilation research because it mimics the multiscale structure of the atmosphere while remaining computationally cheap. In control simulation experiments aimed at keeping state variables below a threshold of 12, EKG-MPPI reduced the number of extreme-event counts to 7,622 plus or minus 146, compared with 7,820 plus or minus 142 for EnKC alone, while simultaneously cutting the mean control-input magnitude from 0.191 to 0.166. The improvement was consistent across ten simulations with different random seeds. The researchers attribute the gain to the nonlinear evaluation introduced in the MPPI refinement stage, which captures effects that linearized EnKC simply cannot see.

The comparisons went further. Against a standalone Bayesian optimization baseline, which is the natural benchmark for expensive black-box optimization problems, EKG-MPPI delivered both fewer threshold exceedances and roughly half the control effort, with 7,622 events at an input magnitude of 0.166 versus 8,299 events at 0.340 for Bayesian optimization. The team also built an ablation experiment, BO-MPPI, in which Bayesian optimization rather than EnKC supplies the prior for the MPPI stage. Across multiple values of the MPPI temperature parameter, the EnKC-informed version consistently sat below the Pareto-like trade-off curve traced by the BO-informed variant, meaning it achieved comparable suppression with smaller interventions. A sensitivity analysis showed that smaller temperature parameters concentrate the weight updates on rare, highly effective samples, generally strengthening suppression but increasing input magnitude, with the smallest mean input occurring near a temperature of one. Notably, the entire EKG-MPPI computation took about 3.87 times ten to the minus four seconds per control decision, roughly 560 times faster than the BO-MPPI pipeline and comfortably shorter than a single model time step.

The second testbed was far more ambitious: the surface quasi-geostrophic model, a fluid-dynamical framework used to describe mesoscale atmospheric and oceanic dynamics, discretized on a 128-by-128 grid with surface buoyancy acting as an active tracer from which the full velocity field is diagnosed. Here the control objective was to suppress regional wind speed within eight different target areas, each roughly ten grid squares wide, over simulations spanning about ten days of model time. Across forty simulations combining eight scenarios and five random seeds, EKG-MPPI achieved lower maximum wind speeds in the target regions than EnKC in nearly every case while requiring smaller mean control magnitudes. Visualizations of one scenario showed strong winds persisting in the uncontrolled run, partial weakening under EnKC with localized hotspots remaining, and the strongest suppression under EKG-MPPI. The authors note that the sampling-based refinement introduces larger variability in intervention magnitude, a plausible consequence of exploration in a high-dimensional system, and that scenarios in which winds intensify late in the simulation showed more complex behavior as intervention effects accumulated.

The authors are careful to frame EKG-MPPI as a building block rather than a blueprint for weather modification. The current implementation assumes a single, spatially sparse actuator applying a one-step perturbation, and it does not yet enforce hard physical constraints on the control. Future work will extend the framework to multiple actuators acting over multiple time steps, constraint-aware formulations, and evaluation metrics that explicitly quantify unintended impacts on regions outside the target zone, along with robustness to model and observation uncertainty. Still, the significance of the result is hard to overstate. For the first time, researchers have a computationally tractable, derivative-free method that can identify where a tiny intervention will matter most in a chaotic geophysical flow, verify that judgment against the full nonlinear dynamics, and do so fast enough for online control. Whether the eventual actuators are dust plumes, cloud-brightening sprays, or turbine arrays, any serious assessment of their potential will need exactly this kind of mathematics. The era of guessing at weather intervention may finally be giving way to an era of computing it.

Subject of Research: Hybrid ensemble Kalman and model predictive path integral control for suppressing extremes in chaotic geophysical flows

Article Title: Ensemble Kalman-guided model predictive path integral control for spatially localized suppression of extremes in chaotic geophysical flows

Article References: Ensemble Kalman-guided model predictive path integral control for spatially localized suppression of extremes in chaotic geophysical flows. (n.d.). https://doi.org/10.5194/npg-33-473-2026

Image Credits: AI Generated

DOI: 10.5194/npg-33-473-2026

Keywords: chaos theory, weather modification, ensemble Kalman filter, model predictive control, data assimilation, extreme events, geophysical fluid dynamics, Lorenz-96 model, quasi-geostrophic model, control simulation experiments, tropical cyclones, computational optimization