From 8,000 Experiments to 28: AI Rewrites the Green Ammonia Catalyst

Eight thousand candidate experiments. That was the brute-force screen researchers at UNSW Sydney faced to find a better catalyst for green ammonia. Machine learning cut the list to 28. Four rounds of testing later, they had a five-metal alloy that produces seven times more ammonia than the benchmark.

What is gaining ground in AI-built ammonia

The discovery side is where the numbers move first. UNSW is not alone. Atmonia has handed Fujitsu a year-long program pairing high-performance computing with real catalytic data, and a Nature paper this year described CRESt, a multimodal AI-robotic platform that runs its own electrocatalysis discovery loop. Each closes the make-test-learn gap that used to take months.

As we wrote in August, AI-designed catalysts are already rewriting carbon capture. The discovery loop here is the same one. The lab bottleneck there was screening, too. Ammonia is the next substrate where the loop gets compressed, and the spillover into methanol and e-fuels is already visible in the same papers.

The second shift is production. Envision runs its Chifeng plant on two foundation models: Dubhe optimizes the hydrogen and ammonia train, Tianji reads weather to keep the whole system supplied from variable renewables. Applied Computing and KBR launched INSITE 3.0 on the same logic, an AI layer for low-emission ammonia operations. The catalyst still matters. The plant now runs because software arbitrages the wind.

The field of builders is widening past the labs. Topsoe signed a front-end engineering deal for a green ammonia project in Aqaba, Siemens Energy partnered with Shomax on zero-carbon fertilizer, and Nel ASA launched a next-generation electrolyzer platform under the Calluna project with EU backing. The catalyst is one input; the electrolyzer and the synthesis loop are the rest of the stack AI is now touching.

The moat is the data, not the patent. Every discovery loop Envision, Atmonia, or CRESt runs generates screened candidates and failure modes that train the next model. The first mover that accumulates the largest clean dataset of multi-metal electrocatalysis gets a compounding edge the late entrant cannot buy. That is why the screening story and the deployment story are one story: the plants in operation feed the models that design the next plants.

Why ammonia, not hydrogen, is the molecule that ships

Ammonia is already one of the most produced chemicals on earth, roughly 200 million tons a year. The industry knows how to liquefy, store, and ship it. Hydrogen does not have that logistics base at scale. For a principal weighing where decarbonization capital goes, ammonia is the carrier with terminals, pipelines, and class rules already written. Crack it back at the port and you have hydrogen. Burn it in a turbine and you have backup power. The molecule is the optionality.

The catch is the 2 percent problem. Conventional Haber-Bosch consumes about 2 percent of the world's energy and emits a comparable share of CO2. Green ammonia only earns its name when the hydrogen feed is renewable and the synthesis is efficient. That is exactly where AI touches both ends: cheaper catalysts lower the energy per ton, and foundation models squeeze the intermittency tax out of renewable-powered plants. Without both, green ammonia stays a premium niche. With both, it competes with the gray molecule on cost.

Shipping is the clearest demand signal. Long-haul vessels and fertilizer supply chains already move ammonia in volume, so the offtake does not need a new market to be invented. Green molecules priced near gray ones land directly into existing contracts.

What is losing ground

Manual trial-and-error screening is the obvious loser. When 28 runs beat 8,000, the economist's case for brute force collapses. The incumbent Haber-Bosch process is not going away, but its cost moat is narrowing as green electricity plus AI scheduling close the levelized-cost gap.

Single-metal optimization is quietly losing, too. The UNSW win was a five-metal alloy. The frontier is combinatorial, and humans do not screen five-element spaces by hand. The advantage now sits with whoever owns the data loop, not the chemist with the best intuition.

What is new this cycle

The February shipment was the first end-to-end commercial green ammonia cargo, certified to RFNBO and ISCC standards, bound for LOTTE Fine Chemical. That matters less as a single delivery than as proof the value chain, from renewable hydrogen to maritime logistics, is operational rather than sketched.

On the science side, a Nature Protocol paper laid out plasma-coupled electrochemical ammonia synthesis from air and water at ambient conditions. If ambient-pressure synthesis scales, the capex case for green ammonia changes: no high-pressure loop, no on-site hydrogen storage, containerized modules near the renewable source. The same paper family points at direct synthesis routes that skip the hydrogen intermediate entirely.

And the market itself is being re-rated. Green-ammonia synthesis catalyst spend is modeled to grow more than tenfold by 2035, the fastest line item inside a decarbonizing fertilizer and fuel chain. Capital is pricing the screening bottleneck as solved. The open question is whether tonnage follows the model on the timeline the market is now assuming.

The deployment economics to watch

Chifeng is the proof asset, not the norm. The plant quotes 320,000 tons a year now and 1.5 million by 2028, built on 100 percent renewable power. The number to watch is the levelized cost of hydrogen, because ammonia's price is mostly the hydrogen inside it. When AI scheduling lifts capacity factor on wind and solar, the synthesis train stops paying the curtailment penalty, and the cost curve bends the way solar did a decade ago.

The gap between announcement and first molecule still bites. Most green ammonia projects sit at final investment decision or earlier, and Chifeng is rare in actually shipping. For private capital, that spread is the entry map: the plants in operation set the benchmark, and the ones at FID are the call on execution risk. A project that has moved from press release to cargo has retired the risk that the others are still carrying.

Certification is the quiet gate. RFNBO and ISCC compliance are what let Chifeng's molecules clear European and Asian standards. As more regions write green-molecule rules, the certified producer gets pricing power the uncertified one cannot touch.

The financing tells the same story. Green ammonia is a capital-intensive build, and the projects that clear final investment decision are the ones with an offtake already signed. Chifeng's output is bound for LOTTE under a multi-year frame, which is why the plant financed and shipped while peers stalled at press release. For a principal, the filter is blunt: back the asset with a locked buyer and certified molecules, not the deck with the best slide. The catalyst breakthrough lowers one line of the model. The offtake and the power decide whether the model closes.

AI-built catalyst versus the legacy screen