ChemAgents: A Hybrid Agentic Framework Coupling Large Language Models, Molecular Transformers, and Physics-Based Simulations for Accessible Computational Chemistry

Read full text Summarise with VIRAT
Authors: Niket Sharma
Cited by: 0

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

The rapid advancement of artificial intelligence in chemistry has produced a powerful yet fragmented ecosystem of predictive models, creating a significant accessibility gap for domain experts who lack specialized machine learning expertise. To bridge this divide, we introduce ChemAgents, a practical, domain-specific agentic AI framework designed to make molecular property prediction and simulation more intuitive and reproducible. Expanding upon previous conversational agent architectures such as ChemCrow [1] and ChemToolAgent [2], ChemAgents employs a layered system of 13 specialized agents, a custom toolset integrating PubChem, LAMMPS molecular dynamics, and a novel Simulation-AI Coupler, alongside an aggressive duallayer caching architecture (in-memory LRU + persistent SQLite) that reduces repeated-query latency by up to 200×. The framework offers two complementary interaction modes: (i) a direct conversational agent for rapid task routing and (ii) a hybrid LLM-tool agent that leverages multistep ReAct reasoning for complex inquiries. A novel confidence-triggered hybrid prediction scheme dynamically routes molecules between fast AI screening and computationally expensive physics-based validation. The core specialist models were developed by fine-tuning the ChemBERTa-2 foundation model [3] on the MoleculeNet benchmarks ESOL, Tox21, and ClinTox [4], demonstrating competitive performance against established GCN baselines. By combining generalist LLM reasoning with domain-specialized transformer models and traditional simulation engines, ChemAgents demonstrates a practical path toward democratizing access to advanced computational chemistry workflows.

Fields of Study & Subjects

Computational Drug Discovery Methods Machine Learning in Materials Science History and advancements in chemistry

Publication details

Journal / source
ChemRxiv
Publication date
25 March 2026
Language
EN
Document type
Preprint