Machine learning opens greener route to urea production

Urea is a very important chemical, especially for fertilizers. However, the production of urea consumes large amounts of energy and relies heavily on fossil fuels.

But new discoveries from Griffith University and the Queensland University of Technology have revealed a new way to produce urea electrochemically, using electricity and exhaust gases such as carbon monoxide (CO) and nitrogen oxides (NO) instead.

“The challenge is that when CO and NO react over a catalyst, they usually do not form urea,” said co-first author Professor Qin Li from Griffith University.

“Instead, they tend to produce unwanted by-products such as ammonia and hydrocarbon compounds.

“This makes selective urea production very difficult.”

What did the researchers do?

Using a combination of quantum chemical simulations and machine learning, the research team uncovered a better catalyst design that encourages CO and NO to combine to form carbon-nitrogen bonds, rather than producing unwanted side reactions.

They learned:

• A pair of metal atoms fixed at the edge of a carbon material (called a double atom catalyst)

• How do these metal pairs interact with CO and NO simultaneously?

• Why do some metals promote urea formation while others do not?

As a result, they were able to consider 90 catalyst designs using highly accurate computer simulations and then rapidly screen more than 1,400 additional candidates using machine learning.

What were the key findings?

The most important discovery was not how each gas adhered to itself, but how strongly CO and NO adhered to the catalyst.

The research team identified a single number, called the co-adsorption energy, that reliably predicts whether a catalyst will produce urea or ammonia or hydrocarbons instead.

“We discovered a very narrow ‘sweet spot’ for this energy,” said co-lead author Dr. Yun Han.

“If the bonds between CO and NO are too weak, they fall off the surface.

“If the bonds are too strong, the gas will be over-reduced and undesirable by-products will form.

“Only moderate bond strengths favored urea formation.”

Why was machine learning important here?

The research team said it would take years to test thousands of catalyst designs in physics-based simulations.

To shorten that period, the team developed a machine learning model using simple atomic properties (from the periodic table) and structural information of carbon edges.

The model accurately predicted critical co-adsorption energies, allowing the researchers to narrow down 1,458 possible catalysts to 259 promising catalysts and validate only the best few in simulations.

“This approach dramatically accelerates catalyst discovery,” said co-lead author Professor Aijun Du, a computational chemist at the Queensland University of Technology.

“This study provides clear design rules for producing urea catalysts and shows how machine learning and chemistry can solve complex reaction problems.

“This brings urea production closer to a low-carbon, sustainable process and provides a reusable blueprint for designing catalysts for other green chemical reactions.

“Rather than relying on time-consuming and costly trial and error, we can systematically design catalysts that efficiently convert exhaust gases into fertilizer.”

The paper “Machine learning-assisted design framework for carbon edge-driven binary atom catalysts for urea electrosynthesis” ASC nano.