Multi-scale modeling and optimization of CO₂ hydrogenation to light olefins via deep learning and optimization algorithms

This study presents a novel multi-scale modeling and optimization framework that integrates deep learning and advanced optimization techniques to enhance CO₂ hydrogenation for producing light olefins. A first-principles model, incorporating reaction kinetics (RWGS, FT, FTS) and heat and mass transfer dynamics, was developed for a one-dimensional fixed-bed reactor, generating a dataset of over 400,000 simulation rows. Among the evaluated deep learning architectures, the recurrent neural network (RNN) demonstrated superior predictive accuracy and robustness against 2 % Gaussian noise, establishing its efficacy as a surrogate for mechanistic models. Two optimization strategies were employed: (1) The interior-point method, leveraging gradient-based optimization, achieved propylene yields of 37.66 % by tuning inlet temperature and pressure (608.1 K, 1.6406 MPa) and 38.05 % by optimizing catalyst packing density (423 kg/m³), yielding 5.33 % and 5.72 % improvements over the baseline (32.33 %), respectively; and (2) Reinforcement learning (RL) with algorithms including DDPG, PPO, and TD3, where TD3 achieved the highest reward (34.02 %) in an RNN-based environment, demonstrating adaptive control under dynamic conditions. Comparative analysis reveals that the interior-point method excels in static, high-precision optimization, while RL offers robustness in dynamic, uncertain environments. This dual-optimization approach, augmented by domain randomization and mechanistic model augmentation to address plant-simulation mismatches, provides a robust foundation for intelligent carbon capture and utilization (CCU) systems, advancing CO₂ conversion and selective olefin synthesis.