State estimation of Rayleigh-B'enard convection with reduced-order models

arXiv:2606.20511 · physics.flu-dyn · Submitted 2026-06-18 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Physics. The day in general physics: instruments, methods and results across the field.

Kai: I'm Kai.

Mira: And I'm Mira, here with Lev.

Kai: We're looking at the paper: "State estimation of Rayleigh-B'enard convection with reduced-order models".

Mira: Here's the gist of the paper. State estimation of Rayleigh-B´enard convection with reduced-order models develops a state estimation framework for two-dimensional Rayleigh–B´enard (RB) convection that combines a stable Galerkin reduced-order model (ROM) with an extended Kalman filter (EKF).

Kai: The filter tracks the most energetic modes with high fidelity and achieves time-averaged reconstruction errors below 14% for velocity and 9% for temperature.

Mira: Model Formulation The ROM is constructed from controllability modes of the linearised Boussinesq equations, which provides the nonlinear dynamical model for the filter prediction step.

Kai: This approach reduces the problem dimension from O(106) unknowns to O(10)–O(102).

Title and authors: Kai: So, what they found is that this combination of the ROM and EKF actually works across three very different types of flow—periodic, quasiperiodic, and chaotic regimes <ref:2606.20511#pg4>. This shows it's quite robust.

Mira: They demonstrate that this filter successfully tracks the most energetic modes with high fidelity, achieving time-averaged reconstruction errors below fourteen percent for velocity and nine percent for temperature across all those regimes <ref:2606.20511#pg1>.

Lev: That level of error is pretty respectable when you're dealing with sparse measurements, but we need to remember that the model accuracy itself is a variable here <ref:2606.20511#pg4>.

Kai: They also looked at how they could improve things. One big suggestion is a greedy sensor placement strategy based on the Kalman gain matrix <ref:2606.20511#pg3>.

Mira: That strategy lets them progressively remove the least informative sensors, which helps minimize the number of sensors needed for state estimation <ref:2606.20511#pg3>.

Lev: That’s smart from a hardware standpoint because it tells you exactly which measurements are most valuable for correcting the state estimate in real time <ref:2606.20511#pg3>.

The paper's summary: Kai: They also tackled the noise estimation. Instead of assuming the noise is isotropic and uncorrelated—which is a common, simpler assumption—they estimated the noise covariances Q and R directly from direct numerical simulations or DNS data <ref:2606.20511#pg3>.

Mira: That adjustment significantly improves performance across all three dynamical regimes, making filter errors drop by about four point five percent for velocity and three percent for temperature in the chaotic case <ref:2606.20511#pg3>.

Lev: Estimating those covariances from DNS data adds complexity to the setup, but if that reduction in error is what you need to get a reliable state estimate on real hardware, then it’s worth the computational overhead <ref:2606.20511#pg4>.

Kai: They also compared their results against using fixed low-dimensional approximations versus improving the model fidelity itself, and they found that for the temperature field, increasing just the ROM dimension didn't give as much accuracy gain as making a better dynamical model <ref:2606.20511#pg4>.

Mira: So it shows that sometimes focusing on getting the physics right in your equations is more important than just having a bigger approximation space to cover the dynamics <ref:2606.20511#pg4>.

The paper's improvements: Kai: To wrap up, this paper on "State estimation of Rayleigh-B´enard convection with reduced-order models" shows that combining the Galerkin ROM with the EKF gives you a robust state estimation framework across periodic, quasiperiodic, and chaotic flows <ref:2606.20511#pg4>.

Mira: The main implication is that you can get fairly good estimates of velocity and temperature even when measurements are sparse, provided you use data-estimated noise covariances from DNS data <ref:2606.20511#pg3>.

Lev: And the greedy sensor placement strategy means you can design a much more efficient sensor setup by intelligently pruning what's least informative for the state correction <ref:2606.20511#pg3>.

Kai: So, despite relying on a fixed low-dimensional approximation and a sparse observation configuration, this framework is pretty solid for tracking the most energetic modes <ref:2606.20511#pg4>.

Mira: The paper lays out a path forward with extensions to three dimensional configurations and coupling the EKF with state feedback laws for closed-loop control of thermal convection <ref:2606.20511#pg4>.

Lev: I’d say the next step is testing this on actual hardware where you have those real-time measurement constraints, because getting that O(ten) to O(one hundred two) model running efficiently is where the real engineering challenge lies <ref:2606.20511#pg2>.

Kai: So we’ve seen how they build this estimation framework for Rayleigh-B´enard convection using a stable Galerkin reduced-order model and an extended Kalman filter, which really sets a baseline for how we can handle complex fluid dynamics with limited sensor data.

Mira: It's a solid piece of work that shows the power of combining model reduction with intelligent filtering to get high-fidelity results in challenging flow regimes.

Lev: It’s important to keep an eye on how these estimation techniques scale up when you move from two dimensions to three dimensions, because that’s where the complexity really starts climbing <ref:2606.20511#pg4>.

Conclusion: Kai: So we've been looking at how they used a stable Galerkin reduced-order model combined with an extended Kalman filter to track Rayleigh-Bénard convection across periodic, quasiperiodic, and chaotic regimes.

Mira: They managed to get reconstruction errors below fourteen percent for velocity and nine percent for temperature in all those different flow types.

Lev: That’s a big win because the model itself is only an approximation of the full Boussinesq equations with about ten to one hundred unknowns, so tracking that reliably is hard.

Kai: Exactly, but they did it by using controllability modes from the linearised Boussinesq equations to build that reduced-order model for the filter prediction step.

Mira: And they used the extended Kalman filter to estimate those modal amplitudes and get those physical fields from sparse measurements obtained from direct numerical simulations.

Lev: The trick there is that they relaxed the assumption of uncorrelated noise—the standard way people usually do this—by actually estimating the noise covariances Q and R directly from that DNS data.

Kai: Which they say really helps pull down all those error metrics across the three dynamical regimes, even in the chaotic case where filter errors drop from about thirteen point eight percent down to nine point two percent.

Mira: And they also developed this greedy sensor placement strategy based on the Kalman gain matrix that lets you progressively remove the least informative measurement locations.

Lev: That ranking process showed that horizontal velocity measurements actually carry the biggest corrective weight in the Kalman update, which is useful for designing real hardware setups.

Kai: So, to wrap up, this paper on "State estimation of Rayleigh-B´enard convection with reduced-order models" shows that even with a fixed low-dimensional approximation and sparse sensors, you can get robust state estimation across all three flow regimes.

Mira: It’s a solid framework that proves the combination of the Galerkin ROM and the EKF works well when you use data-estimated noise covariances.

Lev: It definitely sets up a clear direction for future work, like extending this to three dimensional configurations or coupling it with state feedback laws for actual closed-loop control of thermal convection.

Kai: Right, so the next thing we'll look at is how they tackle those three dimensional setups and what kind of real-world hardware would actually be needed to run this on.

Enrique Flores-Montoya, André F. C. da Silva, André V. G. Cavalieri

Aerothermals department, ITP Aero · Divisóao de Engenharia Aeroespacial, Instituto TecnolƓgico de Aeronática

physics.flu-dyn

Submitted: 2026-06-18

Updated: 2026-11-09

License: http://creativecommons.org/licenses/by-nc-nd/4.0/

Importance score: 82/100

The gist: State estimation of Rayleigh-B´enard convection with reduced-order models develops a state estimation framework for two-dimensional Rayleigh–B´enard (RB) convection that combines a stable…

Key concepts

Reduced-Order Model (ROM)
The ROM simplifies the complex physics of Rayleigh-Bénard convection by using only the most important modes (controllability modes) from the linearised Boussinesq equations. This reduces a massive system of unknowns down to a manageable, low-dimensional set, making it computationally feasible for real-time state estimation.
Extended Kalman Filter (EKF)
The EKF is a statistical tool used to estimate the true state of the system by combining predictions from the ROM dynamics with sparse measurements. It predicts where the system will be next and then corrects this prediction using actual sensor data, effectively filtering out noise and tracking complex dynamics.
Greedy Sensor Placement Strategy
This strategy intelligently selects which sensors to keep based on how much they help correct the state estimate. It uses the Kalman gain matrix to score measurements; sensors that provide the most corrective information are prioritized, allowing for a sparser, more efficient sensor network without losing accuracy.

Terminology

Summary

State estimation of Rayleigh-B´enard convection with reduced-order models develops a state estimation framework for two-dimensional Rayleigh–B´enard (RB) convection that combines a stable Galerkin reduced-order model (ROM) with an extended Kalman filter (EKF). p. 0

Where the paper says it

da Silva† and Andr´e V. Cavalieri‡ Divis˜ao de Engenharia Aeroespacial, Instituto Tecnol´ogico de Aeron´autica, S˜ao Jos´e dos Campos, SP, Brazil (Dated: June 23, 2026) In this work, we develop a state estimation framework for two-dimensional Rayleigh–B´enard (RB) convection that combines a stable Galerkin reduced-order model (ROM) with an extended Kalman filter (EKF).

The filter tracks the most energetic modes with high fidelity and achieves time-averaged reconstruction errors below 14% for velocity and 9% for temperature. p. 0
Where the paper says it

Direct numerical simulations (DNS) are used to generate synthetic measurements for data assimilation. We assess filter performance across periodic, quasiperiodic, and chaotic regimes, demonstrating that the filter tracks the most energetic modes with high fidelity and achieves time-averaged reconstruction errors below 14% for velocity and 9% for temperature.

Model Formulation

The ROM is constructed from controllability modes of the linearised Boussinesq equations, which provides the nonlinear dynamical model for the filter prediction step. This approach reduces the problem dimension from O(106) unknowns to O(10)–O(102). p. 1

Where the paper says it

2 with some loss in model accuracy but achieves a large reduction of the problem dimension, from O(106 ) unknowns to O(10)–O(102 ). At this scale, classical control and estimation methods become tractable, with several successful applications reported in the literature [5–7, 17].

The extended state vector X is decomposed into a baseline steady solution and a time-dependent perturbation expressed as a sum of orthonormal modes. p. 2
Where the paper says it

For that, we define an extended state vector that groups the velocity and temperature fields: X = u v θT . (4) The extended state is decomposed into a baseline steady solution, θ0(y), and a time-dependent perturbation expressed as a sum of orthonormal modes: X (t, x) = 0 0 θ0(y) T + Xn j=1 cj (t) χj (x), (5) where, θ0(y) = 1 − y, is the conductive equilibrium temperature profile and χj (x) are coupled modes containing velocity and temperature components.

The physical fields are reconstructed from this decomposition using the modal amplitudes cj(t). p. 2
Where the paper says it

Each mode χj can be written as: χj (x) = Uj (x) Vj (x) Tj (x) T , (6) where Uj , Vj and Tj are the horizontal velocity, vertical velocity and temperature components of mode j, respectively. From this decomposition, the physical fields are reconstructed as: u(t, x) = Xn j=1 cj (t) Vj (x), θ(t, x) = θ0(y) +Xn j=1 cj (t) Tj (x), (7) where Vj = Uj Vj T groups the velocity components of mode j.

Filter Implementation

The EKF is employed to estimate the modal amplitudes c(t) from sparse measurements of the physical fields obtained from direct numerical simulations. The prediction step involves integrating the ROM dynamics, and the predicted covariance is computed using the linearized dynamics. p. 4

Where the paper says it

Prediction step Given the posterior estimate cˆk−1 and covariance Pk−1 at time tk−1, the predicted state at time tk is obtained by integrating the ROM: cˆ − k = Φ(cˆk−1; Ra, P r, ∆tk). (24) The predicted covariance is computed using the linearized dynamics: P − k = Fk Pk−1 F T k + Q, (25) where Q is the model error covariance matrix, which accounts for the uncertainty introduced at each prediction step by the unresolved dynamics in the ROM, and Fk is the state transition matrix obtained from the Jacobian of the ROM: Fk = I + J(cˆk−1) ∆tk, Jij = ∂fi ∂cj .

The update step computes the innovation and uses the Kalman gain to incorporate measurement information into the posterior state estimate. p. 5
Where the paper says it

The Kalman gain determines the optimal weighting between the model prediction and the measurement: Kk = P − k HT S −1 k . (30) The posterior state estimate incorporates the measurement: cˆk = cˆ − k + Kk zk.

Performance Across Regimes

The filter performance is evaluated across periodic, quasiperiodic, and chaotic regimes. In the chaotic regime (R = 120), the filter successfully tracks the chaotic trajectory throughout the simulation window. p. 8

Where the paper says it

Despite the chaotic dynamics, there is a very good agreement between the DNS projection and the filter estimates. The filter successfully tracks the chaotic trajectory throughout the simulation window, with the estimates closely following the

The mean filter errors remain moderate even at R = 120, with Ec ≈ 13. p. 9
Where the paper says it

The error metrics for the chaotic case confirm a progressive degradation of filter performance with increasing R. Nonetheless, the errors remain moderate even at R = 120, with Ec ≈ 13.

8% and Eθ ≈ 8. p. 3
Where the paper says it

(5) into the governing equations eqs. (2)–(3) and taking the inner product with χi yields the ROM ordinary differential equation: dci dt = P r  F 0 i + X j F 1 ij cj   + P r √ Ra X j D V ij cj + 1 √ Ra X j D T ij cj − X j Lij cj − X j,k Nijk cj ck, (10) where the projected operators are defined as: F 0 i = ⟨χi , θ0(y) ev⟩c, (11) F 1 ij = ⟨χi , Tj ev⟩c, (12) D V ij = ⟨Vi , ∇2Vj ⟩, (13) D T ij = γ 2 ⟨Ti , ∇2 Tj ⟩, (14) Lij = ⟨χi , (Vj · ∇θ0) eθ⟩c, (15) Nijk = ⟨χi , Vk · ∇χj ⟩c, (16) with ev = 0 1 0T and eθ = 0 0 1T being unit vectors in the wall-normal velocity and temperature components, respectively.

7%.

Sensor Optimization Strategy

A greedy sensor placement strategy is developed that exploits the Kalman gain matrix to progressively remove the least informative sensors. The relevance score for the j-th measurement channel is defined as Sj = sum k vuutX Nit k=1 Xn i=1 K2 k,ij. This algorithm reveals a clear hierarchy among sensor types, showing that horizontal velocity measurements carry the greatest corrective weight in the Kalman update. p. 0

Where the paper says it

Finally, we exploit the Kalman gain matrix to develop a greedy sensor placement strategy that progressively removes the least informative sensors. The algorithm reveals a clear hierarchy among sensor types and can be used to derive skeletal observation configurations.

Noise Covariance Estimation

The isotropic, uncorrelated assumption on the noise covariances is relaxed by estimating Q and R directly from DNS data. Using data-estimated covariances reduces all error metrics across the three dynamical regimes, with filter and velocity errors decreasing by roughly 4. p. 9

Where the paper says it

Overall, these results demonstrate that the Kalman filter, combined with the Galerkin ROM, provides robust state estimation across all three dynamical regimes (periodic, quasiperiodic, and chaotic), despite relying on a fixed low-dimensional approximation and a sparse, suboptimal sensor configuration. These results can be placed in the context of related works that combine POD-Galerkin ROMs with data assimilation methods for flow estimation, such as Kikuchi et al.

5% and 3% in the chaotic case. p. 8
Where the paper says it

The moderate increase in all error metrics relative to the R = 40 case – roughly 3% in Ec and Eu, and 2% in Eθ – is attributable to two concurrent effects: the increased dynamical complexity of the quasiperiodic attractor, and the larger truncation error of the ROM. When the Rayleigh number is further increased to R = 120, the system becomes chaotic.

The analysis shows that the best overall performance is obtained by combining a 192-mode ROM with m = 192 observations. p. 6
Where the paper says it

7 Basis type Nomenclature nα nβ N n nx ny Ra P r γ2 Coupled C96 6 16 96 96 22 64 1 1 1. 24 C192 8 24 192 192 30 64 TABLE I: Modal composition, generation parameters and mesh resolution of the different models evaluated.

Conclusion and Future Work

The present framework demonstrates that the Kalman filter, combined with the Galerkin ROM, provides robust state estimation across all three dynamical regimes (periodic, quasiperiodic, and chaotic) despite relying on a fixed low-dimensional approximation and a sparse, suboptimal sensor configuration. Future work includes extension to three-dimensional configurations and coupling with state feedback laws for closed-loop control of thermal convection. p. 0

Where the paper says it

INTRODUCTION Closed-loop flow control has been a long-standing goal in fluid mechanics, with applications ranging from drag reduction and noise suppression to thermal management and mixing enhancement [1–3]. Active feedback control requires knowledge of the system state, which is generally unavailable through direct measurement in fluid systems since the state corresponds to spatially distributed fields [4].

The EKF is general and can be applied to other flows for which a quadratic Galerkin ROM is available. p. 0
Where the paper says it

It also provides guidance on which measurement variables and spatial locations are most informative for state correction. The present framework is general, and may be applied to other quadratic Galerkin ROMs for state estimation.

Pruning redundant reasoning tokens

The ROM, constructed from controllability modes of the linearised Boussinesq equations, provides the nonlinear dynamical model for the filter prediction step. p. 0

Where the paper says it

Cavalieri‡ Divis˜ao de Engenharia Aeroespacial, Instituto Tecnol´ogico de Aeron´autica, S˜ao Jos´e dos Campos, SP, Brazil (Dated: June 23, 2026) In this work, we develop a state estimation framework for two-dimensional Rayleigh–B´enard (RB) convection that combines a stable Galerkin reduced-order model (ROM) with an extended Kalman filter (EKF). The ROM, constructed from controllability modes of the linearised Boussinesq equations, provides the nonlinear dynamical model for the filter prediction step.

The ROM, constructed from controllability modes of the linearised Boussinesq equations, provides the nonlinear dynamical model for the filter prediction step.

How the training signal is built

The observation matrix H is constructed from the modal basis evaluated at the probe locations. p. 4

Where the paper says it

Here, we retain the white noise assumption for simplicity and leave the extension to autoregressive measurement models for future work. The observation matrix H is constructed from the modal basis evaluated at the probe locations.

The observation vector at each time step is assembled from the DNS fields as described in §II. p. 4
Where the paper says it

The observation matrix has the block structure: H = Hu Hv Hθ T , (21) where the sub-matrices are: Hu ℓj = Uj (xℓ), Hv pj = Vj (xp), Hθ qj = Tj (xq). (22) The observation vector at each time step is assembled from the DNS fields as: yk = [u(x1, tk), .

p. 4
Where the paper says it

The observation matrix has the block structure: H = Hu Hv Hθ T , (21) where the sub-matrices are: Hu ℓj = Uj (xℓ), Hv pj = Vj (xp), Hθ qj = Tj (xq). (22) The observation vector at each time step is assembled from the DNS fields as: yk = [u(x1, tk), .

2. The observation matrix H is constructed from the modal basis evaluated at the probe locations.

Where the gains come from

The greedy sensor placement strategy exploits the Kalman gain matrix to progressively identify and remove the least informative measurement locations. The algorithm reveals a clear hierarchy among sensor types, showing that horizontal velocity measurements carry the greatest corrective weight in the Kalman update. p. 0

Where the paper says it

Finally, we exploit the Kalman gain matrix to develop a greedy sensor placement strategy that progressively removes the least informative sensors. The algorithm reveals a clear hierarchy among sensor types and can be used to derive skeletal observation configurations.

p. 0
Where the paper says it

Finally, we exploit the Kalman gain matrix to develop a greedy sensor placement strategy that progressively removes the least informative sensors. The algorithm reveals a clear hierarchy among sensor types and can be used to derive skeletal observation configurations.

The improvement in performance obtained with the greedy algorithm is better reflected in the global error metrics. p. 9
Where the paper says it

A summary of the filter errors as a function of the dynamical regime is presented in Tab. The error metrics for the chaotic case confirm a progressive degradation of filter performance with increasing R.

The observation matrix H is constructed from the modal basis evaluated at the probe locations. The observation vector at each time step is assembled from the DNS fields as described in §II.2.

The greedy sensor placement strategy exploits the Kalman gain matrix to progressively identify and remove the least informative measurement locations. The algorithm reveals a clear hierarchy among sensor types, showing that horizontal velocity measurements carry the greatest corrective weight in the Kalman update.

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Improved state estimation framework for Rayleigh-Bénard convection can track most energetic modes with high fidelity across periodic, quasiperiodic, and chaotic regimes by combining a stable Galerkin reduced-order model with an extended Kalman filter (EKF). This system can achieve time averaged reconstruction errors below 14% for velocity and 9% for temperature.

  2. The system can infer the temperature field from velocity measurements alone in hybrid simulation scenarios by employing a ROM-based EKF, as shown by results where velocity observations alone suffice to reconstruct the state, including the temperature field.

  3. A greedy sensor placement strategy based on the Kalman gain matrix allows for progressively sparser observation configurations, revealing a hierarchy among sensor types and providing guidance on which measurement variables and spatial locations are most informative for state correction.

  4. The improved system can identify dynamically active zones as the most informative measurement locations by exploiting the time-averaged norm of the Kalman gain to rank sensors, with horizontal velocity measurements being systematically retained the longest.

  5. The system can reduce all error metrics across dynamical regimes by using data-estimated noise covariances (Q and R estimated from DNS data) instead of assuming isotropic, uncorrelated noise, which improves performance in the chaotic case by reducing filter errors from 13.8% to 9.2%.

  6. The improved system can distinguish between the sensitivity of estimation to model fidelity versus observation density, demonstrating that for the temperature field, increasing the ROM dimension alone yields lower accuracy gains than improving the dynamical model.

Abstract

In this work, we develop a state estimation framework for two-dimensional Rayleigh-Bénard (RB) convection that combines a stable Galerkin reduced-order model (ROM) with an extended Kalman filter (EKF). The ROM, constructed from controllability modes of the linearized Boussinesq equations, provides the nonlinear dynamical model for the filter prediction step. Direct numerical simulations (DNS) are used to generate synthetic measurements for data assimilation. We assess filter performance across periodic, quasiperiodic, and chaotic regimes, demonstrating that the filter tracks the most energetic modes with high fidelity and achieves time-averaged reconstruction errors below 16% for velocity and 19% for temperature. We apply the ROM-based EKF to a hybrid simulation scenario where the system state is assimilated from velocity measurements. It is shown that velocity observations alone suffice to reconstruct the state, including the temperature field. Finally, we exploit the Kalman gain matrix to develop a greedy sensor placement strategy that progressively removes the least informative sensors. The algorithm reveals a clear hierarchy among sensor types and can be used to derive skeletal observation configurations. It also provides guidance on which measurement variables and spatial locations are most informative for state correction. The present framework is general, and may be applied to other quadratic Galerkin ROMs for state estimation.

Sources

Related papers