Accelerating Multiphase Control: An AI Surrogate Framework

• Bridging the Real-Time Gap: Developed an AI framework that replaces slow, traditional physics simulations with Fourier Neural Operators, enabling the first real-time control of complex multiphase flows.
• Precision AI Decision-Making: Integrated Bayesian Optimization with neural surrogates to precisely track liquid levels in bubble columns, overcoming the mathematical “noise” that typically breaks standard control algorithms.
• Accelerating Industrial Innovation: Reduced the computational cost of multiphase flow management by orders of magnitude, providing a scalable foundation for smarter, more efficient chemical reactors and energy systems.

🔬 Graphical Representation of Computational Acceleration The acceleration plot highlights why surrogate-based control is practical for real-time multiphase systems: by replacing repeated CFD evaluations with a trained Fourier Neural Operator, the control pipeline achieves much lower latency while preserving predictive quality. In our benchmarks, the optimized training workflow (GPU-resident preloading plus CuPy preprocessing) reached up to 3.09x speedup per epoch, and inference throughput reached about 19,711 to 24,508 samples per second (0.041 to 0.051 ms per sample). This speed gain is critical because MPC requires many horizon evaluations at each control step, and faster model rollouts directly translate into faster optimization and more responsive closed-loop control. You can access the paperhere.