Catalysts underpin most of the chemical industry, yet discovering a better one has traditionally been slow, intuition-led trial and error across a near-infinite space of compositions and structures. AI-driven discovery changes the economics: machine-learning models predict which candidates are promising, so experiments focus only on the best leads.
Working principle
Models are trained on data from density-functional theory (DFT) simulations and experiments to predict descriptors that govern catalytic activity — for example the adsorption energy of a key intermediate (the basis of the Sabatier 'volcano' principle). A trained graph neural network can then estimate these properties for millions of candidate surfaces almost instantly, screening the space. Promising candidates feed an active-learning loop with high-throughput experiments that return data to improve the model.
| Aspect | Trial-and-error | AI-guided |
|---|---|---|
| Search | Intuition, sequential | Systematic, parallel in silico |
| Throughput | Few candidates | Millions screened |
| Cost | High (lab time) | Lower (compute first) |
| Driver | Experience | Data + descriptors |
CaveatAI does not replace chemistry — it prioritises it. Model quality depends on good training data, and a predicted catalyst must still be synthesised, tested and proven stable.
Applications
- Electrocatalysts for hydrogen evolution and CO₂ reduction
- Catalysts for ammonia synthesis and emissions control
- Discovery of new alloy and single-atom catalysts
References & further reading
- Nørskov et al., “Towards the computational design of solid catalysts,” Nature Chemistry, 2009.
- Chanussot et al., “The Open Catalyst 2020 (OC20) Dataset and Community Challenges,” ACS Catalysis, 2021.
- Tran & Ulissi, “Active learning across intermetallics to guide discovery of electrocatalysts,” Nature Catalysis, 2018.