A framework for realisable data-driven active flow control using model predictive control applied to a simplified truck wake

TL;DR

This work addresses real-time active flow control by coupling a data-driven latent dynamics model with Model Predictive Control, using only nonintrusive surface-pressure sensors. The authors develop an end-to-end, offline-trained framework comprising a temporal encoder, an action-aware latent dynamics model, and a force decoder, trained with a VICReg-based regularization and unrolled multi-step losses. They then apply SHAP analysis to identify a minimal, physically meaningful sensor set and use knowledge distillation to create a lightweight 'slim' encoder; control is performed in the latent space with a differentiable MPC over a horizon of $H=25$, achieving a $12.8\%$ drag reduction in a $2$D truck wake at $Re=500$ using just four sensors. The results demonstrate robust, real-time capable AFC with interpretable sensor reduction, suggesting a practical path toward hardware deployment for drag reduction in ground vehicles.

Abstract

We present an efficient and realisable active flow control framework with few non-intrusive sensors. The method builds upon data-driven, reduced-order predictive models based on Long-Short-Term Memory (LSTM) networks and efficient gradient-based Model Predictive Control (MPC). The model uses only surface-mounted pressure probes to infer the wake state, and is trained entirely offline on a dataset built with open-loop actuations, thus avoiding the complexities of online learning. Sparsification of the sensors needed for control from an initially large set is achieved using SHapley Additive exPlanations. A parsimonious set of sensors is then deployed in closed-loop control with MPC. The framework is tested in numerical simulations of a 2D truck model at Reynolds number 500, with pulsed-jet actuators placed in the rear of the truck to control the wake. The parsimonious LSTM-MPC achieved a drag reduction of 12.8%.

Figures (15)

• Figure 1: Schematic of the flow configuration. Sensor locations depicted in blue and zero-net-mass flow jets schematised in red.
• Figure 2: DNS mesh and detail of the flow near the rear of the truck, with the jets actuating at maximum suction/blowing intensity.
• Figure 3: Sample portion of the modulated control signal applied during training data generation. Top: normalized signal. Bottom: Wavelet spectrogram of the same signal to show the frequency content over time.
• Figure 4: Schematic of the model architecture. A history of sensor observations $\mathbf{S}_t$ is mapped by the temporal encoder to a latent state $\mathbf{z}_t$. The action-aware dynamics model propagates this state forward to $\hat{\mathbf{z}}_{t+1}$ using the current control action $\mathbf{a}_t$. The force decoder maps the latent state $\mathbf{z}_t$ to the predicted aerodynamic force coefficients $\hat{\mathbf{C}}_t$.
• Figure 5: Comparison between the aerodynamic force coefficients estimated by the force decoder and the ground truth values from the test dataset using a chirp actuation signal with instantaneous frequency $f_{chirp}$. The plot shows a range where the wake briefly stabilises and the highest frequency region. The dashed blue line represents the prediction and the solid black line represents the reference values.