Real-time control of multiphase processes with learned operators

This paper proposes a surrogate-assisted model predictive control framework that utilizes a trained Fourier Neural Operator to efficiently forecast spatiotemporal dynamics and optimize control inputs for multiphase processes, thereby overcoming the computational limitations of traditional numerical models.

The Gist

Imagine you are trying to keep the water level in a giant, bubbling soda fountain perfectly steady. But here's the catch: the bubbles are chaotic, the liquid sloshes unpredictably, and the sensors you have are slow and blurry. If you try to adjust the gas flow based on what you see right now, you'll be too late. The system will have already changed.

This is the problem scientists face with multiphase flows—systems where liquids, gases, and solids mix and interact, like in chemical reactors, inkjet printers, or even inside your body. Controlling them is like trying to steer a ship through a hurricane while wearing foggy glasses.

Here is how the authors of this paper solved it, using a mix of artificial intelligence and smart control theory.

1. The Problem: The "Black Box" is Too Slow

Usually, to control these bubbling systems, engineers use super-complex computer simulations (called CFD) to predict what will happen next. Think of these simulations as a high-definition, slow-motion movie of the physics. They are incredibly accurate, but they take hours to run a single second of the process.

If you want to control the system in real-time (like a self-driving car reacting to a pedestrian), you can't wait hours for the computer to tell you what to do. You need an answer in milliseconds.

2. The Solution: The "Crystal Ball" (The Neural Operator)

Instead of running the slow, heavy simulation every time, the authors trained a special type of AI called a Fourier Neural Operator (FNO).

• The Analogy: Imagine you have a master chef who has watched a million hours of soup cooking. If you show them a picture of the soup bubbling now, and tell them "I'm going to turn the heat up a little," they can instantly imagine what the soup will look like in 5 seconds. They don't need to re-simulate the physics of boiling water; they just "know" the pattern.
• How it works: The FNO is this master chef. It learns the "language" of the bubbling liquid. It looks at the recent history of the bubbles and the gas flow, and it instantly predicts the future shape of the liquid surface. It's not just guessing a number; it's predicting the entire map of where the bubbles and liquid will be.

3. The Strategy: The "Chess Player" (Model Predictive Control)

Once the AI can predict the future, the system needs to decide what to do. This is where Model Predictive Control (MPC) comes in.

• The Analogy: Think of a chess player. Before making a move, they look ahead several turns: "If I move my pawn here, my opponent might move their knight there, and then I can..."

• How it works: The controller asks the AI "Crystal Ball" to simulate the next few seconds for different possible gas flow settings.

  • "What happens if I increase the gas?"
  • "What happens if I decrease it?"
  • "What if I keep it the same?"

  The AI predicts the outcome for all these scenarios in a split second. The controller then picks the move that keeps the liquid level closest to the target, while avoiding dangerous spikes or crashes.

4. The Twist: The "Blind Spot" and the "Smart Search"

There was one tricky part. The goal was to control the height of the liquid. But the AI predicts the entire field of bubbles. To get the height, the computer has to look at the prediction and say, "Okay, where does the liquid stop and the air begin?" This involves a sharp "cut-off" (a threshold), which makes the math messy and jagged.

Because the math was jagged, standard gradient-based optimization (like sliding down a smooth hill to find the bottom) didn't work well. Instead, the authors used Bayesian Optimization.

• The Analogy: Imagine you are looking for the lowest point in a foggy valley, but the ground is full of sudden cliffs and holes. You can't just slide down. Instead, you take a few careful steps, check the ground, and use a smart map to guess where the next best step is. You don't need to check every single inch of the valley; you just need to check the most promising spots to find the bottom quickly.
• The Result: The system tested a few gas flow options, found the best one using this "smart search," and applied it.

The Outcome: A Smooth Ride

When they tested this on a virtual bubble column reactor:

• Speed: The AI could predict the future of the bubbles thousands of times faster than the traditional physics simulation.
• Accuracy: The system successfully kept the liquid level steady, even when the target level changed suddenly.
• Stability: It handled the chaotic bubbling without getting confused or making wild, jerky movements.

Why This Matters

This paper shows that we can finally control chaotic, fast-moving industrial processes in real-time without needing supercomputers running in the background. By teaching an AI to "see" the physics and using it as a fast crystal ball, we can make manufacturing safer, more efficient, and more precise. It's like giving a self-driving car the ability to see around corners and predict traffic patterns instantly, rather than just reacting to what's in front of it.

Technical Summary

1. Problem Statement

Multiphase flows (e.g., gas-liquid systems like bubble columns) are ubiquitous in engineering but notoriously difficult to control in real-time due to:

• Strong Non-linearity: Rapid phase transitions, hysteresis, and complex interfacial physics.
• Sensor Limitations: Available sensors often lack the spatial and temporal resolution required to capture the full state of the system, leading to inaccuracies.
• Computational Cost: High-fidelity Computational Fluid Dynamics (CFD) models, which are necessary to resolve these dynamics, are too computationally expensive to embed directly into Model Predictive Control (MPC) loops. MPC requires repeated "rollouts" (simulations) over a prediction horizon to optimize control actions, a task that standard CFD solvers cannot perform at the speed of sensing and actuation.

The core challenge is to achieve closed-loop regulation of multiphase processes (specifically tracking liquid levels in a bubble column) with the speed required for real-time control, without relying on simplified reduced-order models that may fail to capture complex physics.

2. Methodology

The authors propose a Surrogate-Assisted Model Predictive Control (NMPC) framework that replaces the expensive CFD solver with a learned neural operator.

A. Neural Operator Surrogate (FNO)

• Architecture: The authors utilize a Fourier Neural Operator (FNO). Unlike standard neural networks (CNNs/RNNs) that map finite-dimensional vectors, FNOs learn mappings between infinite-dimensional function spaces ($G: \mathcal{A} \to \mathcal{U}$).
• Input/Output:
  • Input: A channel-stacked tensor containing:
    1. A short history of recent volume fraction fields ($\alpha$).
    2. Spatial coordinate channels ($x, y$).
    3. A candidate control sequence (inlet gas velocity) over the prediction horizon.
  • Output: The predicted future evolution of the volume fraction field ($\alpha$) over the horizon.
• Training: The FNO is trained on a dataset generated by OpenFOAM (using the twoPhaseEulerFoam solver). The network learns to forecast the spatiotemporal evolution of the phase-indicator field based on recent states and actuation signals.

B. Control Strategy (NMPC with Bayesian Optimization)

• Objective: Minimize a receding-horizon cost function that tracks a piecewise-constant liquid level setpoint while penalizing aggressive control moves.
• The Optimization Challenge: The controlled output (liquid level) is extracted from the predicted volume fraction field via a thresholding operation (identifying the interface height). This operation renders the cost function non-convex and non-smooth, making gradient-based optimization (which FNOs theoretically support) ineffective or unstable.
• Solution: The authors employ Bayesian Optimization (BO) to solve the 1D optimization problem (selecting the inlet velocity).
  • BO builds a Gaussian Process surrogate of the cost function based on a few FNO rollouts.
  • It uses an acquisition function (Expected Improvement) to balance exploration and exploitation, finding the optimal inlet velocity with very few evaluations.
• Loop: At each time step, the controller:
  1. Feeds the current state history and a candidate control sequence into the FNO.
  2. Predicts the future field evolution.
  3. Extracts the liquid level and evaluates the cost.
  4. Uses BO to find the optimal control input.
  5. Applies the input to the physical system (simulated) and repeats.

3. Key Contributions

1. Surrogate-Assisted MPC for Multiphase Flows: Demonstrates a practical framework for controlling complex, nonlinear multiphase systems in real-time by replacing CFD with FNOs.
2. Field-Level Forecasting: Unlike previous approaches that might predict only scalar outputs, this method predicts the entire spatiotemporal field (volume fraction). This preserves spatial structures (e.g., bubble plumes, recirculation zones) necessary for accurate level extraction and future generalization to partial observability.
3. Hybrid Optimization Scheme: Successfully integrates FNOs with Bayesian Optimization to handle the non-smooth cost functions arising from interface extraction, avoiding the need for differentiable programming in this specific context.
4. Real-Time Performance: Achieves inference speeds orders of magnitude faster than CFD, enabling the MPC loop to run at time scales compatible with industrial process control.

4. Results

The framework was validated on a 2D bubble column reactor simulation where the goal was to track a time-varying liquid level setpoint by adjusting the gas inlet velocity.

• Prediction Accuracy: The trained FNO accurately reproduced the global interface dynamics and large-scale structures. While it slightly under-resolved fine-scale structures within the liquid phase, the macroscopic behavior was sufficient for control.
• Control Performance:
  • Tracking: The controller successfully tracked piecewise-constant setpoints with short transients during switching.
  • Stability: The system exhibited well-damped responses with no sustained oscillations.
  • Steady-State: Small steady-state offsets were observed at extreme setpoints (due to actuator bounds), but the controller remained robust.
• Computational Efficiency:
  • Inference Speed: On an NVIDIA Quadro GV100, the FNO achieved a throughput of 19,711–24,508 samples per second (approx. 0.04–0.05 ms per sample).
  • Comparison: This is significantly faster than the time required for a single CFD step, making real-time MPC feasible.
  • Optimization: The Bayesian Optimization loop required very few FNO rollouts to converge to the optimal control action.

5. Significance and Future Work

• Industrial Applicability: This work provides a viable pathway for implementing high-performance control in fast multiphase unit operations (e.g., chemical reactors, additive manufacturing) where traditional CFD-MPC is too slow and reduced-order models are too inaccurate.
• Scalability: The field-level approach allows the framework to generalize beyond simple scalar regulation. It can potentially control complex distributions (e.g., catalyst bed activity) by leveraging the full spatial information retained in the surrogate.
• Future Directions:
  • Partial Observability: Extending the method to work with sparse sensor data (using Physics-Informed Neural Operators).
  • Gradient-Based Optimization: Exploring the use of the FNO's inherent differentiability for gradient-based MPC if the cost function can be smoothed.
  • Uncertainty Quantification: Integrating uncertainty-aware decision-making to improve robustness in industrial settings.

In conclusion, the paper establishes that Fourier Neural Operators are a powerful enabler for real-time Model Predictive Control of multiphase flows, bridging the gap between high-fidelity physics and the computational constraints of control systems.