Calibrating Understanding: AI-Supported Discovery of Threshold Concepts in Analytical Chemistry

Threshold concepts (TCs) reorganize understanding in ways that are transformative, integrative, irreversible, bounded, and troublesome. Building on our prior work that framed TCs in Analytical Chemistry (AC) and leveraged AI to invert laboratory learning, we propose a two-stage, AI-forward workflow to surface TC candidates and connect them to laboratory and workplace practice.

We treat AC learning as the integration of intrinsic fundamentals—measurement aims and analytical figures of merit, sampling, validation, and sample preparation—with shared fundamentals contributed by adjacent disciplines (Chemistry, Physics, Mathematics, Biology, Engineering), such as light–matter interactions, statistics, signal processing, modeling, and instrumentation principles.

Stage 1 develops a self-calibrating, rubric-anchored prompt for an LLM. Using short excerpts from textbooks, lab manuals, and guidelines, an LLM assigns per-criterion TC scores and must cite sentence-level evidence. A second agent flags violations (missing or off-target evidence, misapplied criteria) and a third revises the prompt (clearer anchors, counter-examples, few-shot templates). Iterations continue until violations fall below a preset threshold and spot checks by an AC expert stabilize; all runs are versioned for auditability.

Stage 2 deploys the validated prompt at scale over a broader corpus (syllabi, practitioner artifacts, education journals) and fuses LLM outputs with light NLP signals (keyphrase salience, co-occurrence centrality) to prioritize candidates. An expert panel then adjudicates the shortlist, calibrates per-criterion judgments, and maps each TC to laboratory decisions, assessment tasks, and workplace-proximal activities, explicitly distinguishing intrinsic AC foundations from shared, cross-disciplinary ones.

Outcomes include a prioritized, evidence-cited list of TC candidates, concept maps linking intrinsic and shared foundations, and ready-to-adopt teach-and-assess designs.