Scenario
A raw material in a shipped formulation moves onto the SVHC Candidate List or REACH Annex XIV. The formulator must find a substitute that (a) is not itself restricted, and (b) preserves the technical function of the incumbent.
Worked example. DEHP (di(2-ethylhexyl) phthalate) — an Annex XIV plasticizer — is used in a sealant recipe and must be replaced.
Agent workflow
- Read the recipe and identify the flagged component by CAS/EC number.
- Deterministic lookup — match every component against the current SVHC Candidate List, Annex XIV, and Annex XVII, flagging the affected substance.
- Identify the incumbent's function class in the recipe (here: plasticizer) from master data and structural similarity.
- Generate substitute candidates of the same function class (e.g. DINCH, ATBC, DOTP/DEHT).
- Deterministically re-screen each candidate against the same three lists. This is the value — it catches the regrettable-substitution trap (for example, DIBP is itself an SVHC and is auto-rejected).
- Validated property comparison — MW, LogP, and key descriptors for the incumbent versus each candidate, to argue functional equivalence.
Demo output
A side-by-side table: the incumbent (flagged) versus three candidates, each with its regulatory status (clean / flagged) and property deltas. The magic moment is one candidate auto-rejected because it is itself listed.
Deterministic vs LLM
- Deterministic — regulatory lists and property math.
- LLM — candidate ranking rationale and the written explanation.
Every factual or regulatory claim is a deterministic lookup or a validated calculation, never an LLM guess. The agent reads and routes; the engine decides.
Data sources
- Native (offline): RDKit — MW, LogP, TPSA, and fingerprints for function/structure similarity.
- Public reference layer: ECHA SVHC Candidate List; REACH Annex XIV (Authorisation); REACH Annex XVII (Restriction); EC Inventory (CAS↔EC↔name); PubChem structures and identifiers.
Limits
Substitution output is a set of candidates for human review, not an approval. Property models carry applicability domains; uncertainty is surfaced, not hidden.
Skills orchestrated by this use case
This scenario routes an AI agent through the following curated skills.
-
Formulation Property Optimization
FormulationOptimize a multi-component formulation to hit property targets by defining the component/ratio design space under mixture constraints, screening candidates, and optimizing the blend against multiple objectives — orchestrating Paramus mixture-DOE, property-prediction, and optimization tools.When to use: a user needs to find component ratios that meet several property targets at once.Multi-tool WORKFLOW with mixture-design judgment. -
Excipient Compatibility Screen
FormulationScreen an active ingredient against candidate excipients for chemical and physical incompatibilities, ranking excipients by compatibility risk and flagging known problematic pairings — orchestrating Paramus reactivity, property-prediction, and knowledge-graph tools.When to use: a user needs to choose excipients that won't react with or destabilize the active.Multi-tool WORKFLOW with compatibility judgment.
Tools it reaches for
- RDKit molecular descriptors (MW, LogP, TPSA)
- RDKit fingerprint similarity
Browse the full deterministic layer in the tool browser.