Scenario

A compound developer must design or reformulate a plastic to hit thermal (glass transition, heat-deflection) and mechanical (modulus, strength, toughness) targets while controlling cost. The design space — base resin(s), blend ratios, additive package, and filler loading — is large, and compounding each candidate consumes material and time.

Worked example. A recycled-content polyolefin compound must recover the stiffness of virgin material by choosing the right filler and loading without making the part brittle.

Agent workflow

1. Read the target property spec and the available resins, additives, and fillers.
2. Blend miscibility screen — use solubility-parameter analysis to flag incompatible resin pairs and predict phase behavior of candidate blends.
3. Filler reinforcement screen — model stiffness/strength gains versus filler type and loading, and flag the loading where toughness starts to fall.
4. Additive package selection — choose stabilizers, impact modifiers, and processing aids that fit the base chemistry.
5. Tg screen and mechanical property screen — predict glass transition, modulus, and strength for the ranked candidates.

Demo output

A ranked set of candidate compounds with predicted Tg, modulus, and strength, a miscibility verdict per blend, and the filler loading that best balances stiffness against toughness. The magic moment: the agent finds the filler / loading that restores virgin-like stiffness while staying below the embrittlement threshold.

Deterministic vs LLM

• Deterministic — property predictions, miscibility / solubility-parameter math, and reinforcement modeling.
• LLM — the compound-design narrative and the additive/filler rationale.

Every property number and miscibility verdict is a validated calculation, never an LLM guess. The agent reads and routes; the engine decides.

Data sources

• Native (offline): polymer property prediction; solubility-parameter and miscibility analysis; filler-reinforcement modeling; DOE screening.
• Public reference layer: polymer, additive, and filler property references.

Limits

Screening-level guidance, not a substitute for compounding trials, standardized mechanical testing, or long-term aging. It narrows the compound design space; the extruder and the test bars confirm the winner.