Ten-Million-Scale Search for Ammonia Catalysts Using Generative AI ✛ ✦
Context
Ammonia synthesis from nitrogen and hydrogen is the cornerstone of modern chemical industry: without it, fertilizer production—and thus global food security—would be impossible. The traditional Haber–Bosch process, which uses iron-based catalysts, operates at high temperatures and pressures, consuming about 2% of the world's energy. To reduce energy consumption and the carbon footprint (today the carbon footprint of ammonia synthesis is enormous due to the use of methane and water), researchers are seeking new materials that can work under mild conditions. But the ammonia reaction isn't a single step; it's a complex network of dissociative and associative pathways, where key intermediates (N*, NH*, NNH*, HNNH*) place different demands on the catalyst surface. A catalyst ideal for one intermediate often fails for others. Thus, discovering a catalyst for ammonia synthesis is a compatibility problem across the entire reaction network, requiring a rethink of traditional screening methods that focus on single descriptors.
Methods
To tackle this complexity, the authors developed a platform based on generative transformers—machine learning architectures originally designed for natural language processing, but adapted to generate atomic structures. The transformer was pre-trained on the OC20 dataset, containing two million surface structures. It was then fine-tuned for the four key intermediates, enabling it to generate structures specific to each one. Large-scale generation on powerful GPU clusters produced roughly 15 million configurations per intermediate—an unprecedented volume for catalytic screening. The vast space was then compressed using hash-based deduplication and clustering to remove repeats and preserve diversity. The selected structures were evaluated using machine learning potentials from EquiformerV2—a neural network that predicts binding energies. This numerical simulation rests on the fundamental laws of quantum mechanics: solving the Schrödinger equation, first formulated by Erwin Schrödinger, and obeying the Pauli exclusion principle (Wolfgang Pauli), density functional theory (DFT) methods can calculate the electronic structure. Meanwhile, the distribution of electrons among states is governed by Enrico Fermi statistics, and the first law of thermodynamics dictates the energy balances on the surface.
Results
The results were stunning. Overlap analysis showed that at typical generation scales (10⁵–10⁶ structures), materials compatible with all four intermediates were virtually undetectable. Only at the ten-million level did 279 catalyst families emerge that satisfied the full spectrum of constraints. Among them, classical catalysts based on iron (including Fe-Pt) and ruthenium were recovered, confirming the screening's validity. But most importantly, previously unexplored compositions were found: Fe-V, Al-Pd-Zr, Al-Ti, Au-Ti, and Cu-Zr. Detailed DFT calculations revealed their distinct preferences for reaction pathways. For instance, Fe-V lowers the activation barrier for N₂ dissociation to 0.72 eV, operating via a dissociative mechanism, while Al-Pd-Zr stabilizes hydrogenated intermediates (NNH* and HNNH*) and follows an associative pathway with a 1.22 eV barrier for the first hydrogenation step. Interestingly, energy balance alone doesn't guarantee effectiveness: Cu-Zr easily splits N₂, but subsequent hydrogenation is energetically unfavorable, underscoring the importance of full network compatibility. In reaction thermodynamics, not only energy but also entropy is crucial, which together with the thermal effect determines the free energy.
Implications
This work shifts the paradigm of catalyst discovery: it shows that considering the entire reaction network is not a luxury, but a necessity. The method not only finds active sites but also maps out entire compatibility landscapes, where catalysts can be chosen for specific mechanisms. The recovery of known industrial catalysts confirms the approach's reliability, while the discovery of new families demonstrates its predictive power. For the chemical industry, this means the ability to purposefully design catalysts for complex, multi-step processes like ammonia synthesis through computational experiments that complement lab work.
Future development
In the future, this approach can be extended to other industrially relevant reactions—from CO₂ reduction to methanol oxidation—where similarly branched intermediate networks are at play. Improving the accuracy of ML potentials and incorporating kinetic factors (rate constants, solvent effects) will enable a shift from energy screening to full prediction of catalytic activity. Moreover, generative models will be able to account not just for composition, but also nanoparticle morphology, defects, and reaction conditions, bringing virtual screening closer to reality. Integration with self-driving labs could close the loop from hypothesis generation to experimental validation.
Impact
The results will impact several fields: in chemical engineering, by accelerating the development of catalysts for 'green' ammonia; in materials science, by advancing multiscale modeling methods; in artificial intelligence, by fostering generative architectures that can handle 3D structures and physical constraints.
Next steps
Next steps include experimental verification of the predicted materials (primarily Fe-V and Al-Pd-Zr) in lab reactors, and expanding the generative space to binary and ternary alloys, accounting for stability under reaction conditions. In parallel, the ML potential itself will be improved to enhance the accuracy of energy predictions.
Key open problems
This article is directly connected to an unsolved problem in chemical physics: how to rationally design catalysts for complex reaction networks, bypassing empirical approaches. It also touches on a fundamental question about scaling relations in catalysis: can we break the linear dependencies between binding energies of different intermediates and thus overcome the activity 'ceiling' predicted by volcano-shaped curves? The ten-million-scale computational study suggests the existence of rare but achievable materials that circumvent these limitations.
The knowledge machine's view
This section is addressed to the author of the paper and their colleagues. Neighbouring works were found by meaning-based search across our archive, not by word overlap. We do not review: we show the field around the work.
What the work does
This work presents a transformer-based generative AI framework for discovering ammonia synthesis catalysts. The authors generated about 15 million structures for each of the four key intermediates, selected 279 materials compatible with all stages of the reaction network, and uncovered both known catalysts and new families.
Where its strength is
The strength of the work lies in the large-scale application of generative models to a complex reaction network with multiple intermediates, allowing compatibility with all stages to be considered rather than optimizing for a single descriptor.
Where it could go next
- The approach could be extended to other multi-step reactions, drawing on experience from the analysis of chameleon materials, where AI also searched for materials with adaptive properties.based on: How AI hunts for chameleon materials arXiv:2607.03578
- It would be interesting to apply similar generative screening to catalysts for hydrogen production, as in the work on titanium dioxide, where AI helped optimize photocatalysis.based on: How Dirty Titanium Dioxide Teaches AI to Produce Hydrogen arXiv:2606.01089
- The approach could potentially be integrated with methods from the roadmap for ions to speed up the evaluation of stability and conductivity of new materials.based on: Roadmap for Ions: How AI Paves the Way to the Batteries of the Future arXiv:2606.24480
Worth trying
- Test structure generation with additional stability constraints, using metrics from the work on superconductors.
- Check whether fine-tuning on experimental data improves the accuracy of binding energy predictions.
- Use clustering to identify common motifs among new catalysts, as in the analysis of crystals under the AI microscope.
Why this may matter
Accelerating the discovery of ammonia catalysts could reduce energy consumption and the carbon footprint of fertilizer production, which is critical for sustainable development.
Nearby works in our archive
- Crystals Under the AI Microscope: How a Machine Found 68,000 New Superconductorssimilarity 0.61
- Amino Acid Relative Found in a Cosmic Cloudsimilarity 0.58
- How Artificial Intelligence Becomes a Scientist's Partnersimilarity 0.58
- How AI hunts for chameleon materialssimilarity 0.56
- How Dirty Titanium Dioxide Teaches AI to Produce Hydrogensimilarity 0.56
- Roadmap for Ions: How AI Paves the Way to the Batteries of the Futuresimilarity 0.56
The archive is sparse around this work: the closest review is only 0.61, against a usual 0.64. It looks like few have drilled nearby.
Original abstract
as written by the authors arXiv:2606.22926
Catalyst discovery for ammonia synthesis is inherently a reaction-network challenge because catalytic performance is governed not by a single adsorbed intermediate, but by a surface's orchestrated compatibility with multiple distinct intermediates across competing dissociative and associative pathways. However, navigating ultra-large chemical spaces under such multi-intermediate constraints remains a formidable bottleneck for conventional screening workflows. Here, we report a reaction-network-level catalyst discovery framework driven by ten-million-scale generative exploration. By coupling adsorbate-specific generative Transformers with high-throughput machine learning potentials, we systematically map the structure-property landscapes of four critical intermediates (N*, NH*, NNH*, and HNNH*). Scale-dependent overlap analysis shows that the full four-intermediate compatibility space remains strongly under-sampled at conventional 105-106 generative scales, emerging exclusively under ten-million-scale exploration. By generating approximately 15 million configurations per adsorbate, followed by structural compression and machine-learning-potential predictions, we identified 279 highly potential target materials. This sparse compatibility space successfully recovers traditional Fe- and Ru-based motifs while uncovering previously unexplored catalyst families. Representative DFT calculations validate pathway-dependent mechanisms: Fe-V emerges as a dissociative-pathway lead by significantly lowering the initial N2 dissociation barrier, whereas Al-Pd-Zr efficiently stabilizes associative intermediates as an associative-pathway lead. These findings establish multi-intermediate reaction-network compatibility as a robust criterion for discovering advanced catalysts from multi-million generative chemical spaces.
Interestingly, ammonia is one of the most produced chemicals in the world: its annual synthesis exceeds 150 million tons, comparable to the mass of all humans on Earth. Meanwhile, nature's nitrogen fixation process by bacteria uses nitrogenase enzymes, whose active sites contain molybdenum or vanadium — elements also found in the new catalysts predicted by AI.
In Andy Weir's sci-fi novel The Martian, the astronaut synthesizes water from hydrogen and oxygen, and fertilizes plants with his own waste. But for a full-fledged Martian farm, ammonia is needed, and thus ultra-efficient catalysts for its synthesis from local resources. It's not far-fetched that such generative AI methods could help design catalysts for colonizing the Red Planet.
Key numbers
- Structures generated per intermediate: ≈15,000,000
- Compatible materials found: 279
- Screening energy window: less than 0.5 eV
- N₂ dissociation barrier on Fe-V: 0.72 eV
- Generation scale needed for overlap: tens of millions