Optimal Dynamic Regimes for CO Oxidation Discovered by Reinforcement Learning.

Area of Science:

• Catalysis
• Chemical Engineering
• Artificial Intelligence

Background:

• Metal nanoparticles are crucial heterogeneous catalysts for activating molecules and lowering reaction energy barriers.
• Reaction yield is influenced by adsorption, activation, desorption, and reaction dynamics, which depend on gas composition, temperature, and pressure.
• Steady-state conditions can lead to catalyst deactivation; dynamic control offers potential for improved performance.

Purpose of the Study:

• To apply reinforcement learning (RL) for dynamic control of CO oxidation on a palladium catalyst.
• To investigate optimal control strategies, including stationary, periodic, and nonperiodic regimes.
• To demonstrate the benefits of dynamic control in heterogeneous catalysis.

Main Methods:

• Utilized a policy gradient reinforcement learning algorithm trained in a theoretical environment.
• Parametrized the RL model using experimental data for CO oxidation.
• Controlled the reaction based on CO and O2 partial pressures over successive time steps.

Main Results:

• The RL algorithm successfully learned to maximize the CO2 formation rate.
• Identified optimal stationary, periodic, and nonperiodic control regimes.
• Provided insights into the advantages of dynamic control for catalytic processes.

Conclusions:

• Reinforcement learning is a viable approach for optimizing heterogeneous catalytic reactions.
• Dynamic control strategies, particularly periodic and nonperiodic regimes, can outperform stationary conditions.
• This work promotes the adoption of RL in catalytic science for enhanced process control and efficiency.