Data-Driven Surrogate Modeling and Optimization for Ammonia Production

Solving large-scale flowsheet optimization using first-principles unit operation models remains a
significant challenge for conventional mathematical programming methods. This complexity
arises due to the high computational cost and difficulty in capturing non-linearities and
interactions within interconnected process units. Therefore, a new framework is needed to
holistically represent an entire process system by employing data-driven surrogate models,
which offer sufficient modeling accuracy while significantly reducing numerical complexity. By
simplifying the representation of individual process units, surrogate models facilitate the
optimization of complex flowsheets, allowing for the identification of both optimal structural
configurations and corresponding operating conditions.
In this study, ammonia production is used as a case study to demonstrate the feasibility and
effectiveness of the proposed approach. Surrogate models are constructed using data
generated from process simulators, effectively replacing traditional first-principles unit models.
An integrated software platform is developed for this purpose, combining the capabilities of
Aspen Plus for process simulation, MATLAB for input-output data sampling, and Python for
surrogate modeling using neural networks. A comprehensive superstructure is formulated to
consider multiple process routes and technologies for hydrogen production (methane steam
reforming, water electrolysis), nitrogen production (air separation units, membrane separations),
ammonia synthesis (Haber-Bosch reactor, microwave-assisted reactor), and downstream
separations (distillation, membrane separations). These alternative flowsheet designs are
systematically evaluated against performance metrics such as cost, sustainability, and
productivity, enabling a holistic assessment of different configurations.