GENIUS: An Agentic AI Framework for Autonomous Design and Execution of Simulation Protocols

TL;DR

GENIUS, an AI-agentic workflow that fuses a smart Quantum ESPRESSO knowledge graph with a tiered hierarchy of large language models supervised by a finite-state error-recovery machine, democratizes electronic-structure DFT simulations by intelligently automating protocol generation, validation, and repair, opening large-scale screening and accelerating ICME design loops across academia and industry worldwide.

Abstract

Predictive atomistic simulations have propelled materials discovery, yet routine setup and debugging still demand computer specialists. This know-how gap limits Integrated Computational Materials Engineering (ICME), where state-of-the-art codes exist but remain cumbersome for non-experts. We address this bottleneck with GENIUS, an AI-agentic workflow that fuses a smart Quantum ESPRESSO knowledge graph with a tiered hierarchy of large language models supervised by a finite-state error-recovery machine. Here we show that GENIUS translates free-form human-generated prompts into validated input files that run to completion on $\approx$80% of 295 diverse benchmarks, where 76% are autonomously repaired, with success decaying exponentially to a 7% baseline. Compared with LLM-only baselines, GENIUS halves inference costs and virtually eliminates hallucinations. The framework democratizes electronic-structure DFT simulations by intelligently automating protocol generation, validation, and repair, opening large-scale screening and accelerating ICME design loops across academia and industry worldwide.

Figures (9)

• Figure 1: GENIUS framework for autonomous Quantum ESPRESSO simulations. This schematic illustration depicts the end-to-end workflow of the GENIUS framework, designed to overcome technical barriers in DFT simulations. Users' Natural language prompts are interpreted by a recommendation system powered by a smart knowledge graph that encodes Quantum ESPRESSO parameter details and constraints. Large language models then generate the corresponding simulation protocols. The framework includes automated validation and an Automated Error Handling (AEH (error 1 and error 2) loop that utilizes a knowledge graph and Large language models to diagnose and correct failed runs iteratively. This integrated process autonomously translates user intent into validated Quantum ESPRESSO input files, ready for submission to the available computational resources.
• Figure 2: State diagram of the AI-driven framework for generating QE simulation protocols. The diagram is organized into four composite states (dashed boxes): RecommendationSystem parses the user's natural-language request ('Interface' → 'InitializeWorkflow'), retrieves materials data ('MaterialsDb') and simulation parameters (via 'DocumentCollection' → 'ConditionExtraction' → 'RetrieveCandidateParameters'), and evaluates them ('EvaluateParameters') to produce a structured input template. 'ProtocolGeneration' uses that template ('PrepareInputTemplate') to generate the actual QE input file ('QeInputGeneration'). 'ExecutionValidation' runs the simulation ('QeRun'), transitioning to 'Finished' on success. 'AutomatedErrorHandling' detects failures ('FailureDetected' → 'CheckRetries') and either retries execution ('AttemptCorrection'), switches to an alternative model ('SwitchModel'), or terminates at 'Failure' if all options are exhausted. Solid arrows show transitions labeled with the triggering action or condition; color and class styling differentiate data sources, main processes, and error-handling loops; more details can be found in the GitHub https://github.com/KIT-Workflows/agentic-workflow-framework.
• Figure 3: The three panels illustrate the structure and interface of the QE knowledge graph and how individual QE input parameters are formalized as machine-readable nodes and exposed through an interactive graph that makes their dependencies explicit.
• Figure 4: Comparison of workflow service JSON payload and the web interface.
• Figure 5: Self-Organizing Map (SOM) analysis of user input prompt embeddings.