The idea

We can search composition space with AI the way we search design space with code. Hypothesize, simulate, synthesize, test, learn. Loop again.

Why it matters

• Battery, alloy, and polymer cycles move from years to months.
• Process + recipe become a defendable advantage, not just the part.

Pattern

• Data layer: structured experiment records; units, uncertainty, lineage.
• Model layer: surrogate models (physics + ML) with cross-validation.
• Design loop: propose → screen → select → synthesize → test.
• Process twin: link recipe steps to outcomes; control for drift.
• Governance: version every recipe; lock release candidates.

First 30–60 days

• Pick one property to improve (e.g., cycle life, toughness, temp window).
• Consolidate past experiments into a tidy dataset; fix units and IDs.
• Train a baseline model; run a small Bayesian design-of-experiments.
• Execute 3–5 builds; capture full metadata; update the model.

Signals / KPIs

• Cycle time per design loop; % hits vs. baselines; variance explained (R²).
• Cost per validated improvement; scrap/rework rate; reproducibility.

Risks & mitigations

• Bad data: enforce templates; units; instrument calibration logs.
• Overfit models: holdout sets; external validation; ablation checks.
• Scale-up gaps: include process parameters early; pilot at two scales.