MPC-Flow: Integrated Model Predictive Flow Control

• MPC-Flow is a framework that combines model predictive control with data-driven reduced-order modeling to predict and manipulate complex flow dynamics in real time.
• It employs a pipeline of input encoding, latent state dynamics, and SHAP-based sensor selection to achieve notable drag reduction and efficient control with minimal sensors.
• The approach uses a constrained finite-horizon MPC solved via autodiff-based optimization to ensure high prediction accuracy (R² > 0.94) and real-time performance.

Model Predictive Control–Flow (MPC-Flow) refers to a class of frameworks that combine model predictive control (MPC) with efficient flow modeling, typically for real-time decision making in complex dynamical systems. In the context of active flow control, as developed in “A framework for realisable data-driven active flow control using model predictive control applied to a simplified truck wake,” MPC-Flow integrates data-driven reduced-order modeling, interpretable sensor selection, and differentiable online optimization using learned models, enabling real-time flow manipulation with a modest sensor/actuation footprint (Solera-Rico et al., 13 Oct 2025).

1. Data-Driven Reduced-Order Modeling Architecture

The MPC-Flow pipeline initiates with offline data-driven modeling of the flow, leveraging high-fidelity direct numerical simulation (DNS) data. The central elements are:

• Input Encoding: The system ingests a temporal history of surface pressure probe measurements. For the truck wake application, a window of $L=32$ time steps from $N_s=90$ probes yields $(32\times 90)$ input sequences processed at each control step.
• Latent State Construction: A single-layer LSTM processes each temporal window, producing a final hidden state that is projected via a small MLP to an $N_z=8$-dimensional latent state, $z_t$:

$$z_t = f_{\text{enc}}(\{s_{t-31},...,s_t\};\theta_{\text{enc}})$$

• Latent Dynamics: The system's evolution in latent space is modeled as a residual MLP-driven update:

$$\Delta z_t = f_{\text{dyn}}(z_t,a_t;\theta_{\text{dyn}}),\ z_{t+1} = z_t + \Delta z_t$$

where $a_t$ denotes the present actuation (e.g., jet command).

• Output Decoding: A compact ResNet-style MLP decodes $z_t$ to the predicted drag and lift coefficients, $[\hat{C}_{d,t},\hat{C}_{l,t}]$.

Training Regimen: The model is trained end-to-end with a composite loss comprising multi-step latent prediction error ($N_s=90$0), smoothed $N_s=90$1 force loss ($N_s=90$2), and VICReg-based batch statistics regularization ($N_s=90$3) for stable representation learning. DNS data are generated via an open-loop, spectrally rich actuation protocol to ensure coverage of the flow’s natural timescales and modes [(Solera-Rico et al., 13 Oct 2025), Sec. 2.2].

2. Model Predictive Control Formulation and Optimization

At each control interval, MPC-Flow solves a finite-horizon constrained optimization to select a sequence of control inputs that minimizes a differentiated cost functional:

• Horizon: $N_s=90$4 steps (~1 vortex shedding cycle), $N_s=90$5 (non-dimensionalized).
• Decision Variables: $N_s=90$6, subject to amplitude bounds $N_s=90$7.
• Cost Function:

$N_s=90$8

with

$N_s=90$9

• Solver: The joint pipeline (encoder, latent unroll, decoder) is deployed in an autodiff framework (PyTorch), ensuring explicit gradient access $(32\times 90)$0. The optimizer (typically Adam, $(32\times 90)$1 steps per $(32\times 90)$2) leverages warm starting and box projection to maintain operational feasibility.

Only $(32\times 90)$3 is applied at each step; the process recedes one step forward, re-solving as new data are available.

3. Sensor Selection via SHAP and Knowledge Distillation

High-dimensional sensing is reduced through a principled feature selection strategy:

• SHAP Analysis: SHapley Additive exPlanations (SHAP) quantify the marginal importance of each probe over a 32-step window. For each sensor and lag, absolute SHAP values are summed and averaged, yielding single scalar importance scores per sensor.
• Results: The vast majority of control-relevant information is concentrated at four base probes in the truck geometry.
• Slim Encoder: Knowledge distillation is deployed: a student encoder (identical LSTM+MLP structure, but with only top-$(32\times 90)$4 sensors) is trained with the teacher output as soft targets ($(32\times 90)$5). With $(32\times 90)$6, drag/lift decoding error is virtually unchanged compared to the original $(32\times 90)$7 architecture [(Solera-Rico et al., 13 Oct 2025), Fig. 10].

4. Closed-Loop Control Performance and Computational Efficiency

The MPC-Flow framework yields notable performance gains under severe sensor and actuation constraints:

• Drag Reduction: Average drag coefficient $(32\times 90)$8 versus $(32\times 90)$9 uncontrolled—a $N_z=8$0 reduction.
• Wake Dynamics: Modification in base pressure ($N_z=8$1 shift from $N_z=8$2 to $N_z=8$3, $N_z=8$4 increase), elongation of the recirculation bubble, and $N_z=8$5 decrease in wake fluctuation amplitude [(Solera-Rico et al., 13 Oct 2025), Figs. 15–18].
• Prediction Accuracy: Multi-step latent and force prediction $N_z=8$6, $N_z=8$7 over test data and minimal error gain over the horizon during MPC operation.
• Real-Time Feasibility: Full MPC-Flow loop (encoding, $N_z=8$8-step latent rollouts, inner optimization) executes in $N_z=8$9 ms for $z_t$0, thus matching or exceeding real-time constraints for high-speed vehicle flows on standard computing hardware.

5. Scientific Context and Generalization Potential

MPC-Flow exemplifies a paradigm shift in real-time flow control toward scalable, data-driven, sensor-efficient architectures:

• Model–Control Decoupling: Offline data-efficient model learning avoids burdensome online adaptation, shifting computational and data requirements to pre-deployment.
• Interpretable Feature Pruning: SHAP-based sensor selection produces minimal instrumentation requirements, facilitating practical deployment on real vehicles.
• Application Scope: The methodology is directly transferable to airfoil separation control, hydrodynamic drag reduction in water tunnels, closed-loop maneuvering of underwater vehicles, and distributed actuator arrays in turbomachinery environments.

MPC-Flow offers a modular, practical, and scalable solution for real-time flow control by synthesizing latent dynamics modeling, automatic sensor sparsification, and differentiable closed-loop optimization into a tightly integrated feedback architecture (Solera-Rico et al., 13 Oct 2025).

6. Related Methodologies and Position in the Literature

MPC-Flow sits at the intersection of multiple strands of contemporary flow control research:

• Deep Learning MPC: Architectures embedding learned surrogate models (RNNs or latent dynamics) in the MPC loop (as in "Deep Model Predictive Control with Online Learning for Complex Physical Systems" (Bieker et al., 2019)).
• Reduced-Order Modeling: Approaches using Galerkin POD, Koopman operator theory, or DMD as compact flow predictors for high-dimensional MPC (Waterman et al., 27 Nov 2025, Uchytil et al., 10 Jul 2025).
• Sensor-Efficient Control: Emphasis on interpretable, sparse measurement strategies to maximize controllability with minimal sensor footprint.
• Differentiable Programming for Real-Time Control: End-to-end autodiff-based optimization pipelines replacing black-box solvers for explicit and efficient gradient computation.

A key distinguishing feature of the approach in (Solera-Rico et al., 13 Oct 2025) is the seamless integration of interpretable sensor pruning, offline-trained latent dynamics, and end-to-end differentiable real-time MPC, producing both high control efficacy and operational practicality.