PFAS substitutes get more than a candidate list: three agent stages close the loop from lab stock to manufacturability

Material substitution is split into requirement analysis, candidate generation and manufacturability assessment, and a rejection from the later stage is fed back to the earlier one, so a substitute is no longer just a candidate list. The earlier GraphAgents workflow centered on generating material hypotheses; it did not constrain proposals to lab-available materials, systematically assess manufacturability, or revise candidates using downstream processing feedback. In the THV substitution case, the system produced structured material requirements, an evidence-grounded substitute candidate and a lab-scale processing route; in the same run, the first candidate stack was returned as blocked on intellectual-property constraints, and only after that feedback was propagated did a second candidate receive a manufacturable verdict.

Candidate generation is treated as an iterative search with verification rather than a single generation, so each candidate passes a PFAS screen and then answers targeted validation questions. The single-call configuration with static retrieval scored lowest on candidate generation, showing that appending retrieved information to one prompt does not keep candidates anchored to verifiable material records. On the candidate-generation subsystem, the full workflow scored 5.22 with gpt-oss-20b and 6.26 with gpt-5.6-sol, while every reduced baseline scored 2.17 or lower; candidate anchoring draws on an internal material database built from 20 supplier specification sheets, averaging 18 extracted properties per material entry.

Manufacturability assessment acts as a downstream gate: when a candidate is judged blocked, its blocking constraints return to candidate generation until a candidate passes or the iteration limit is reached. In a comparison that held the trajectory fixed, propagating the blocking constraints back to candidate generation left candidate-generation scores similar, but raised the mean manufacturability score from 3.17 to 4.85. All four manufacturability criteria rose in that comparison, with manufacturing route from 3.3 to 5.4, practicality from 2.9 to 5.0, material-process compatibility from 3.6 to 5.0 and verdict grounding from 2.9 to 4.0; this feedback event was triggered only once across the repeated-seed experiments.

The workflow targets laboratory settings that need to replace a PFAS-containing material, and it outputs structured requirements, an evidence-grounded candidate and a proposed processing route, where the manufacturable verdict denotes evidence-supported processing plausibility rather than demonstrated synthesis. Evaluation centers on the THV substitution problem, and the multi-seed and matched-backbone comparisons characterize run-to-run variation and model dependence for that case; the two additional queries in the supplementary material serve as workflow demonstrations rather than a controlled benchmark.

The system can retrieve only from knowledge sources that have been curated and made accessible, and candidates are restricted to formulations involving a base material from the lab-available inventory, so its search space is a partial representation of the materials landscape. Numerical thresholds in the requirements did not appear in the query or retrieved passages but were introduced during synthesis and extraction, so they should be treated cautiously. Component contributions are not fully isolated in a factorial design, and the scores come from an LLM judge rather than experimental validation of the substitute, its properties or its processing route.