ChemNavigator: Agentic AI Discovery of Design Rules for Organic Photocatalysts

TL;DR

The paper addresses the difficulty of designing visible-light organic photocatalysts for hydrogen evolution due to enormous chemical space and limited interpretability of traditional ML approaches. It introduces ChemNavigator, an agentic AI that integrates large language model reasoning with DFTB calculations in a hierarchical multi-agent framework to drive hypothesis-driven discovery and rule extraction, supported by 130 open-vocabulary descriptors. In a 200-molecule campaign, the system autonomously derives six design rules governing frontier orbital energies—ether linkages elevating $HOMO$, carbonyl and halogen groups influencing $LUMO$ and band gaps, extended conjugation narrowing $E_g$, and contributions from cyano and amine groups—that align with established concepts of resonance, inductive withdrawal, and $π$-delocalization, and these rules are validated across solvated and higher-level theory benchmarks. Compared with prior ML analyses that reported only carbonyl effects, ChemNavigator provides quantitative effect sizes, feature interactions, and interpretable design guidance, illustrating a scalable path for AI-assisted, chemically grounded materials discovery that complements human intuition.

Abstract

The discovery of high-performance organic photocatalysts for hydrogen evolution remains limited by the vastness of chemical space and the reliance on human intuition for molecular design. Here we present ChemNavigator, an agentic AI system that autonomously derives structure-property relationships through hypothesis-driven exploration of organic photocatalyst candidates. The system integrates large language model reasoning with density functional tight binding calculations in a multi-agent architecture that mirrors the scientific method: formulating hypotheses, designing experiments, executing calculations, and validating findings through rigorous statistical analysis. Through iterative discovery cycles encompassing 200 molecules, ChemNavigator autonomously identified six statistically significant design rules governing frontier orbital energies, including the effects of ether linkages, carbonyl groups, extended conjugation, cyano groups, halogen substituents, and amine groups. Importantly, these rules correspond to established principles of organic electronic structure (resonance donation, inductive withdrawal, $π$-delocalization), demonstrating that the system can independently derive chemical knowledge without explicit programming. Notably, autonomous agentic reasoning extracted these six validated rules from a molecular library where previous ML approaches identified only carbonyl effects. Furthermore, the quantified effect sizes provide a prioritized ranking for synthetic chemists, while feature interaction analysis revealed diminishing returns when combining strategies, challenging additive assumptions in molecular design. This work demonstrates that agentic AI systems can autonomously derive interpretable, chemically grounded design principles, establishing a framework for AI-assisted materials discovery that complements rather than replaces chemical intuition.

Figures (4)

• Figure 1: Schematic showing individual effects of ether linkages (HOMO elevation) and carbonyl groups (band gap reduction). Observed band gaps for molecules with neither feature (baseline), ether only, carbonyl only, and both features. The combined effect (2.66 eV) exceeds the expected additive value (1.97 eV) by 0.69 eV, indicating diminishing returns when combining electron-donating and electron-withdrawing functionalization strategies.
• Figure 2: Schematic overview of the ChemNavigator AI discovery workflow, illustrating the iterative cycle of hypothesis generation, targeted molecular design, quantum chemical calculation, and statistical validation. Arrows indicate information flow between the multi-agent system components. All phases interact with a centralized Memory Bank (center) that stores molecules, calculations, hypotheses, and design rules. The workflow continues iteratively until the maximum cycles are reached or early stopping criteria are met.
• Figure 3: The system architecture is composed of five distinct layers working together to enable efficient task execution. At the top, the Discovery director (Orchestration Layer) manages workflows and coordinates the sequence of operations across components. Beneath it, the Agent Layers hosts specialized agents, each responsible for performing a specific task within the workflow. These agents rely on the Tool Layer, which provides utility functions and connects the system to external resources such as third‑party software, Python packages, and APIs. The system also incorporates two types of memory: short‑term memory, which maintains and updates shared state as agents interact, and a centralized long‑term database that stores persistent information for broader system use.
• Figure 4: Schematic overview of the seven-phase Agentic Discovery Protocol. The workflow illustrates the iterative cycle starting from Analysis and Strategy Selection, moving through Molecular Design and Structure Building, and concluding with Quantum Calculation, Evaluation, and Knowledge Extraction. Specific agents (Scientist, Director, Designer, Curator) manage distinct decision points, utilizing statistical thresholds to confirm design rules or flag hypotheses for further investigation