TeLLAgent: A Dual-Agent Tool-Enhanced LLM Framework for Autonomous Discovery of Organic Materials

TeLLAgent is an open-source innovative project. Through its dual-agent architecture (design + evaluation) that integrates large language models with professional scientific tools, it enables autonomous design and discovery of organic materials, opening up a new path for the automation of materials science, accelerating R&D cycles, and assisting scientists to focus on creative thinking.

Organic materials have great potential in fields such as optoelectronics and energy, but they face three major challenges:

1. Explosive chemical space: The number of molecular structures far exceeds the number of atoms in the universe, making exhaustive search impossible;
2. Complex property-structure relationships: Macroscopic properties are non-linearly correlated with molecular structures through multiple factors;
3. Need for interdisciplinary integration: Requires integration of knowledge from multiple fields such as organic chemistry and quantum chemistry. Traditional methods rely on empirical trial and error, which are slow and expensive; AI technology provides tools to accelerate discovery, but integrating AI with domain knowledge remains an open problem.

The core of TeLLAgent is a dual-agent collaboration architecture:

• Design Agent: Uses LLM to generate candidate molecules (based on target properties, drawing on known design principles, and considering synthetic feasibility), with creativity and literature integration capabilities;
• Evaluation Agent: Calls scientific tools (DFT calculations, molecular dynamics simulations, ML potential functions) to predict properties, ensuring computational accuracy and adherence to physicochemical principles;
• Collaboration Cycle: Design → Evaluation → Feedback → Optimization, combining creativity and rigor.

The tool-enhanced architecture is a key innovation:

• Tools Used: Quantum chemistry (Gaussian, ORCA), molecular simulation (LAMMPS, GROMACS), ML models, chemoinformatics (RDKit);
• Interface Principles: Standardized input/output, error handling, efficiency optimization (fast screening vs. precise calculation);
• Autonomous Workflow: Goal setting → Literature research → Hypothesis generation → Property calculation → Result analysis → Iterative optimization → Report generation It is highly automated; scientists only need to set goals and review results.

TeLLAgent can be applied in multiple domains:

• Organic electronics: High-mobility semiconductors, high-efficiency OLED luminescent materials;
• Energy: Donor/acceptor materials for organic solar cells, battery electrode materials;
• Biomedical: Biocompatible and degradable polymers;
• Catalysis: Organic catalysts, metal-organic frameworks (MOFs); For each scenario, optimization can be done based on specific performance indicators (e.g., band gap, solubility).

TeLLAgent represents important progress in the field of AI for Science:

1. Autonomous Scientific Discovery: Demonstrates the potential of LLM in the entire scientific process (from hypothesis generation to computational verification);
2. New Mode of Human-Machine Collaboration: Assists scientists in handling tedious computational screening, allowing them to focus on creative thinking;
3. Contribution to Open-Source Ecosystem: Provides an extensible platform where researchers can add new tools and agent capabilities.

TeLLAgent has limitations:

• High computational cost: High-precision quantum chemistry calculations are expensive, limiting chemical space exploration;
• Manual experimental validation required: Computational predictions need experimental synthesis and characterization;
• Generalization ability to be verified: Needs testing on a wider range of material categories. Future directions: Integrate robotic experimental platforms to achieve closed-loop autonomous discovery, develop efficient surrogate models to reduce costs, and expand to inorganic/composite material domains.