An autonomous agentic query framework with MCP tools and modernized NIST SRD-46 for accelerating aqueous speciation analysis
Year: 2026
Type: article
Abstract: Large language models (LLMs) could automate reasoning over structured scientific databases in chemistry and materials science, but this requires purpose-built agentic architectures, not simply better models. We present an agentic workflow for systematically querying and extracting data from the NIST Standard Reference Database 46 (SRD-46) to perform autonomous speciation and thermodynamic analysis, with three contributions: (1) a hierarchical toolset, served by Model Context Protocol (MCP) framework, which converts natural-language scientific questions into validated SQL over a database enriched with canonical chemical identifiers, equilibrium-reaction maps, and fingerprint-similarity matrices; (2) an induction-deduction-validation reasoning loop with compactors that preserve key chemical context across multi-turn sessions under dynamic memory limits within the same session; and (3) a soft time/turn-budget hook that maintains coherent reasoning within resource constraints. The resulting framework is LLM-agnostic and spans queries from direct lookups to cross-species comparisons to open-ended tasks like estimating unknown Fe complexation constants and selecting ligands for acidic Fe electrodeposition. We benchmark four frontier models (Claude Opus 4.6, Claude Sonnet 4.6, GPT-5, and GPT-5.4) on 49 prompts across four complexity levels, repeated five times (~1,000 answers), using automated claim-level validation and regex-based tool-result matching. All models retrieve reliably on direct-lookup tasks, confirming the architecture bridges natural-language reasoning and structured database access. Beyond lookups, behaviors of models diverge: GPT-5.4 converges rapidly with high per-claim verifiability but limited exploratory depth, whereas Opus 4.6 pursues broader, cross-examined reasoning at the cost of slower convergence and more unsupported claims. This reveals a fundamental tradeoff between convergent efficiency and exploratory breadth on fuzzy scientific queries that scalar benchmarks cannot resolve. It also suggests that fuzzy scientific queries require expert-stewarded reasoning, not faster convergence criteria or higher benchmark metrics alone. Our results demonstrate that domain-expert-designed agentic architectures and evaluation frameworks are essential infrastructure for AI-driven scientific discovery, and that mixed-model deployment strategies may be needed to balance efficiency and explanatory depth for different classes of scientific questions.
Source: Digital Discovery
DOI: 10.1039/d6dd00466k
Institution: Argonne National Laboratory
Language: English
Subfield: Materials Chemistry
Field: Materials Science
Domain: Physical Sciences
Open Access status: diamond
OpenAlex ID: W7167829840