Leveraging Physics-Informed Neural Networks for Faster and More Accurate Simulation of Gas-to-Liquid Operations and Design

Abstract We present a novel approach to multiscale modeling of Gas-to-Liquid (GTL) processes based on Fischer-Tropsch (FT) synthesis, leveraging Physics-Informed Neural Networks (PINNs). We address the computational challenges inherent to theory-based, ground-up modeling of FT reactors by developing and integrating PINN models for two critical scales: microkinetics and catalyst pellet. The PINN-based approach demonstrates significant computational efficiency gains compared to traditional numerical methods, achieving speedups of 5 and 7 orders of magnitude in case of microkinetics and the pellet-scale model respectively, when employing the ability of GPU to process multiple sets of reaction conditions within the same run. Notably, this increased computational efficiency is achieved without compromising accuracy, with relative errors generally below 1% for most hydrocarbon product reaction rates across a wide range of industrially relevant operating conditions. The developed framework enables near real-time simulation of complete FT reactor tubes, with a full solution obtained in approximately 4 seconds on a single consumer-grade workstation (equipped with 1x V100 GPU). Results from these simulations show qualitative agreement with characteristics observed in real-world reactors. The scalability and parallelization capabilities of neural networks further enhance the potential for rapid, large-scale simulations, creating new possibilities, in line with the industry trend towards digital twins for reactor optimization, process control, and plant-wide simulations, which were previously considered to be computationally intractable.