Learning-based Model Predictive Control of an Ammonia Synthesis Reactor

We investigate the application of state-of-the-art recurrent machine learning methods to tackle the computational challenges associated with the real-time solvability of a packed-bed ammonia synthesis reactor model for optimal control design. A high-fidelity multi-physics model is developed for the ammonia reactor by integrating reaction kinetics and complex transport phenomena. The model involves a system of interconnected algebraic, ordinary, and partial differential equations, providing a deeper understanding and improved prediction accuracy of pressure, temperature, and species concentration dynamics. However, due to its computational complexity, the resulting model is unsuitable for optimal control applications such as model predictive control (MPC). To overcome this limitation, we create a reduced-order model based on long short-term memory (LSTM) using time series data generated from offline simulations of process outputs. This surrogate model effectively captures the system's nonlinear dynamics, enabling MPC's successful implementation and optimal closed-loop performance.