Scenario

Finding a better heterogeneous catalyst by trial-and-error is slow. Computational screening can rank candidates by the descriptors that correlate with activity, grounded in each material's crystal structure, so the lab makes only the few most promising ones.

Worked example. A hydrogenation needs a cheaper catalyst than the incumbent noble metal; a set of mixed-oxide candidates must be ranked.

Agent workflow

1. Read the candidate materials.
2. Crystal structure analysis — parse structures and derive relevant surface / coordination descriptors.
3. Catalyst candidate screen — rank candidates by the activity-relevant descriptors.
4. Propose a short-list to synthesize.

Demo output

A ranked candidate list with the structural descriptors driving each ranking and a recommended synthesis short-list. The magic moment: a low-cost mixed oxide ranks alongside the incumbent, redirecting the campaign toward an affordable option.

Deterministic vs LLM

• Deterministic — the crystal-structure analysis and candidate screening.
• LLM — the candidate-ranking narrative.

Every descriptor is a validated calculation. The agent reads and routes; the engine decides.

Limits

Screening-level prioritization, not a prediction of measured turnover. It focuses synthesis; the reactor confirms activity.