Onnes: A Physics-Grounded Multi-Agent LLM Simulator for Cryogenic Fault Diagnosis in Quantum Computing Infrastructure
06:00 · July 8, 2026 · arXiv cs.AI RSS
Dilution refrigerators are the enabling infrastructure of superconducting quantum computers, yet their fault diagnosis is still dominated by threshold alarms that report that something is wrong, not what. We present Onnes, a physics-grounded digital-twin simulator of a dilution refrigerator (a forward physics model with a learned real-fridge noise fingerprint) that drives a live multi-agent LLM operations layer, and use it for a controlled head-to-head between a zero-shot LLM agent panel and a supervised ML classifier on cryogenic fault diagnosis. The twin couples a real dilution-cooling floor, a noise-and-correlation fingerprint learned from real BlueFors logs, and six physics-grounded fault classes, three engineered to overlap on temperature but separate on flow and pressure. Across a 1000-turn evaluation the zero-shot panel shows no significant difference from the classifier on detection but trails on classification, its errors concentrating on the confusable faults. Curated contrastive few-shot demonstrations and self-consistency voting then raise classification accuracy from 0.685 to 0.990, matching the supervised classifier (0.985) with no parameter updates and six labeled demonstrations; an ablation attributes the gain almost entirely to the demonstrations. Run as a continuous monitor across a nine-run fault-by-seed sweep, the agent catches every developing fault within one poll interval, and a confidence gate suppresses pre-onset false alarms whose rate is backend-dependent. As a first sim-to-real check, a detector trained purely on real BlueFors telemetry posts a real-hardware false-alarm rate of 6.4% and 100% recall on physics faults injected onto real held-out windows. All numbers are drawn verbatim from released run logs.
Summary
Onnes is a physics-grounded digital-twin simulator built to study fault diagnosis in dilution refrigerators, the cryogenic systems that maintain the 10–35 mK base temperatures required by superconducting quantum processors. The twin combines a forward physics model of the dilution-cooling cycle with a noise-and-correlation fingerprint learned from real BlueFors telemetry logs, then feeds identical telemetry windows to both a five-role multi-agent LLM panel and a supervised machine-learning classifier. Six physics-grounded fault classes are defined, three of them deliberately constructed to overlap on temperature signatures while remaining separable on flow and pressure readings.
In a 1000-turn paired evaluation the zero-shot LLM panel shows no statistically significant difference from the classifier on fault detection, yet its classification accuracy reaches only 0.685 because errors concentrate on the engineered confusable faults. Adding curated contrastive few-shot demonstrations together with self-consistency voting lifts classification accuracy to 0.990, statistically indistinguishable from the supervised baseline of 0.985, while using only six labeled examples and no parameter updates. An ablation isolates the demonstrations as the dominant source of the gain.
When the same agent panel is run continuously across a nine-run fault-by-seed sweep, every developing fault is flagged within one poll interval and a confidence gate suppresses pre-onset false alarms at rates that vary with the underlying language-model backend. A preliminary sim-to-real check confirms that a detector trained solely on real BlueFors windows achieves 100 % recall on injected physics faults while maintaining a 6.4 % false-alarm rate on held-out hardware data. The work therefore demonstrates that in-context techniques can close the performance gap between an LLM agent and a fully supervised classifier on a realistic, label-scarce cryogenic diagnostic task without retraining.
Why it matters
This research is highly relevant for Dutch AI and quantum researchers, given the Netherlands' strong strategic focus on quantum computing ecosystems like QuTech. It demonstrates a novel, actionable application of multi-agent LLMs and digital twins for maintaining critical quantum infrastructure.
