Onnes Simulator Aids Quantum Computer Cryogenic Fault Diagnosis

Diagnosing issues in the cryogenic systems essential for superconducting quantum computers is currently limited to basic threshold alarms. These alarms indicate a problem but not its specific nature. To improve this, a new system called Onnes has been developed, which acts as a physics-grounded digital twin simulator of a dilution refrigerator. This simulator incorporates a real-world noise fingerprint and drives a multi-agent LLM operations layer. Onnes was used to compare the performance of a zero-shot LLM agent panel against a supervised machine learning classifier in diagnosing cryogenic faults. The simulator includes six physics-grounded fault classes, some designed to be subtly confusable. Initially, the zero-shot LLM panel matched the classifier in fault detection but lagged in classification, particularly with confusable faults. However, by providing curated contrastive few-shot demonstrations and employing self-consistency voting, the LLM panel's classification accuracy dramatically improved from 68.5% to 99.0%, effectively matching the supervised classifier's 98.5% accuracy with minimal training data. A preliminary real-world check showed a detector trained on actual telemetry achieved 100% recall on injected faults with a low false-alarm rate, highlighting the potential for this simulation-driven approach to enhance quantum computing infrastructure reliability.