Post by Helmholtz AI

What does agentic AI for science look like in practice? Meet GENIUS: an AI co-pilot that lets materials scientists describe a simulation in plain language, and takes care of the rest of the setup. Congratulations to Celso Ricardo Rêgo, Mohammad Soleymanibrojeni, Roland Aydin, Diego Guedes-Sobrinho, Alexandre Cavalheiro Dias, Maurício Jeomar Piotrowski, and Wolfgang Wenzel on publishing "GENIUS: an agentic AI framework for autonomous design and execution of simulation protocols" in Communications Materials (Nature Portfolio)! Funded through the Helmholtz AI Project Call, GENIUS is an agentic AI co-pilot for computational materials science. It takes a plain-language description of a simulation and turns it into a validated Quantum ESPRESSO input file, combining a structured knowledge graph, a tiered hierarchy of LLMs, and a finite-state error-recovery loop that autonomously diagnoses and fixes failures. Benchmarked across 295 prompts spanning different materials and methods, GENIUS produced passing input files directly in around 80% of cases and autonomously recovered 76.3% of initial failures. 🔗 Read about GENIUS and the team's insights into agentic AI for science: https://lnkd.in/dJkPQQ9D The next Helmholtz AI Project Call round opens on 3 August - stay tuned.