Paramus — Copolymer Composition Ranking

Overview

Turn "which comonomer ratio should I use?" into a reproducible enumerate → featurize → predict → trade-off-rank pipeline. The value is the multi-objective judgment across compositions, not any single property prediction.

Guidance vs execution: property values come from tested BRAIN models with provenance. Do not interpolate properties in ad-hoc code.

When to Use

• Choosing comonomer ratios to balance competing properties (e.g. Tg vs. flexibility)
• Ranking a small library of copolymer compositions against targets Do not use for: single-property screening of homopolymers (use the property-specific screen).

Workflow

Comonomers + ratio grid (or explicit compositions) + property targets/weights
  ↓ 1. Enumerate compositions            → build BigSMILES per ratio
  ↓ 2. Validate structures               → Paramus polymer parsing tool
  ↓ 3. Featurize                         → Paramus fingerprinting tool
  ↓ 4. Predict each target property      → Paramus BRAIN models (per property)
  ↓ 5. Multi-objective score             → weighted / Pareto trade-off
  ↓ 6. Rank & report                     → table + Pareto front + provenance

Procedure

1. Discover tools with search/get_schema; do not guess names.
2. Enumerate compositions across the requested ratio grid; validate every generated BigSMILES.
3. Featurize all compositions with one convention; batch where possible.
4. Predict each target property; capture value + uncertainty + model version per property.
5. Score: if the user gives weights, use a weighted objective; otherwise compute the Pareto front and present non-dominated compositions. State the rule explicitly.
6. Report ranked compositions + the Pareto front, flagging any dominated-but-uncertain candidate.

Domain Judgment

• Prefer Pareto reasoning over a single blended score unless the user supplies weights.
• A composition wins only if its advantage exceeds prediction uncertainty on the deciding property.
• Note when a ratio falls outside the models' applicability domain.

Fallbacks

• A property model is missing → rank on available objectives; clearly mark the missing dimension.
• Endpoint unreachable → stop and report; no local surrogate models.