Artificial intelligence techniques have great potential in generating insights into catalytic reaction mechanisms, opening the door to more selective and efficient catalysts. However, using machine-learning (ML) in a general way to study chemical reactions is challenging because broad approaches aimed at outlining all possible reaction pathways are impractical and algorithms that are mechanism-agnostic require human interpretation to connect features with phenomena. Now, a team of researchers led by Iowa State University (Ames, Iowa; www.iastate.edu) materials science professor Qi An has developed an ML framework to autonomously explore catalytic reaction pathways and mechanisms.
The technology is based on a type of ML called deep-reinforcement learning, which involves an ML agent “tasked with identifying plausible reaction pathways through interactions with a defined environment over time,” the researchers write. “Instead of laboriously screening all potential reaction steps, reinforcement learning can navigate reaction networks in an automated manner,” the team says.
The reinforcement learning (RL) framework has been termed high-throughput deep reinforcement learning with first principles (HDRL-FP). Principal investigator An explains: “The reaction-agnostic nature of HDRL-FP arises from its independence from the need for human experts to design specific RL representation of [the reaction] environment (for example states, actions or rewards) for a particular reaction. Instead, the RL environment is solely built on atomic positions, which are then mapped to the potential energy landscape derived from first principles.”
This framework facilitates the fast running of thousands of concurrent RL simulations on a single graphics processing unit (GPU). With GPUs and high-throughput strategies, the method can quickly and automatically identify the optimal reaction pathway from thousands of potential pathways, An said, adding “That effectively identifies viable reaction mechanisms amidst the extremely noisy data in real chemical reactions.”
Artificial intelligence platforms can enable rapid discovery and validation of novel materials and reactions
“The excellent generalizability and cost-efficiency of our framework are primarily a result of the high-throughput capacity enabled by the pioneering architecture of HDRL-FP,” An notes.
As a proof of principle, the team, which included collaborators from Salesforce AI research, used HDRL-FP to identify possible strategies to improve ammonia synthesis, where atmospheric nitrogen reacts with hydrogen over an iron catalyst via the Haber-Bosch process. The researchers anticipate their RL framework will be useful in studying a range of complex industrial catalytic reactions. They published results in a recent edition of Nature Communications.