Workshop: AgenticAI4HPC’26: The First International Workshop on Agentic AI for HPC
Authors: Evan Antoniuk (Lawrence Livermore National Laboratory), Shehtab Zaman (Binghamton University), Tal Ben-Nun and Bhavya Kailkhura (Lawrence Livermore National Laboratory), Kenneth Chiu (Binghamton University), and Brian Van Essen (Lawrence Livermore National Laboratory)
Abstract: While modern language models can perform high-level chemical reasoning, they often struggle to propose solutions to chemical design tasks that adhere to known principles of chemical validity or synthesizability. Here, we introduce CHARGE, a Chemistry-tool Augmented Reasoning-based Generative framework that combines large-scale reasoning models with automated tool calling to tackle molecular discovery and design challenges. We integrate cheminformatics utilities and neural-network property predictors into the reasoning loop, enabling autonomous hypothesis generation, verification, and refinement without user intervention. Across 10 molecule design tasks, CHARGE consistently designs molecules with improved properties compared to the unaugmented LLM baseline. For the task of predicting chemical synthesis outcomes, we show that the tool-assisted reasoning of CHARGE increases synthesis prediction accuracy from 12% to 40%. Together, these results point towards an agentic AI chemistry workflow, whereby LLMs provide flexible scientific reasoning and orchestration, while robust cheminformatics tools verify and guide the molecular design.
Back to AgenticAI4HPC’26: The First International Workshop on Agentic AI for HPC Archive Listing
Back to Full Workshop Archive Listing