Real-time control of multiphase processes with learned operators

Authors: Paolo Guida, Didier Barradas-Bautista Date: 2026-03-26 Paper ID: arxiv:2603.25308

Summary

This work addresses the challenge of controlling highly non-linear multiphase flows by proposing a surrogate-assisted Model Predictive Control (MPC) framework built upon learned operators. A Fourier Neural Operator (FNO) is trained to act as a fast surrogate, predicting the spatiotemporal evolution of the phase-indicator field (volume fraction) over a finite horizon based on recent states and control inputs. This learned surrogate model is then integrated into the MPC optimization loop to iteratively determine the optimal actuation signal with a low computational cost. The approach is validated on controlling the liquid level in a two-phase Eulerian bubble column by adjusting the gas flow rate, demonstrating the viability of FNO-based field forecasting for fast, closed-loop optimization of dynamic processes.

Key Contributions

• Proposing a surrogate-assisted Model Predictive Control (MPC) framework for multiphase processes using learned operators.
• Training a Fourier Neural Operator (FNO) to forecast the spatiotemporal evolution of a phase-indicator field (volume fraction) given recent states and candidate actuation signals.
• Demonstrating the framework on an optimal control problem for a two-phase Eulerian bubble column, tracking liquid level setpoints by adjusting gas flow rate.
• Showing that field-level forecasting with FNOs is suitable for closed-loop optimization due to their low evaluation cost.

Limitations

The current implementation is demonstrated on a single, controlled two-phase system (Eulerian bubble column) and does not explicitly address partial observability, which would require extensions to the learned model.

Key Concepts

• Fourier Neural Operator: A neural operator that learns mappings between infinite-dimensional function spaces by leveraging the spectral representation in the Fourier domain.
• Surrogate-Assisted Model Predictive Control: A control framework where a fast, learned surrogate model replaces the full, computationally expensive system model within the optimization loop of Model Predictive Control (MPC).

Limitations

The current implementation is demonstrated on a single, controlled two-phase system (Eulerian bubble column) and does not explicitly address partial observability, which would require extensions to the learned model.

Links

• ArXiv Abstract
• PDF