Onnes Physics-Grounded Multi-Agent LLM Simulator for Cryogenic Fault Diagnosis in Quantum Computing

Jul 16, 2026

Dilution refrigerators underpin superconducting quantum computers, yet fault diagnosis often relies on threshold alarms that signal an issue without identifying which physical fault is developing. Onnes delivers parity without training by using curated contrastive few-shot demonstrations and self-consistency voting to raise a zero-shot LLM agent panel’s cryogenic fault classification accuracy from 0.685 to 0.990, matching a supervised classifier while updating no parameters and using only six labeled examples.

LanguageEnglish

CategoryTechnology

FormatPDF

Pages17

Reading Time43 min

Abstract

Introduction

Onnes: A Physics-Grounded Multi-Agent LLM Simulator for Cryogenic Fault Diagnosis in Quantum Computing Infrastructure Praneeth Narisetty, Uday Kumar Reddy Kattamanchi, Shiva Nagendra Babu Kore Onnes Research | San Francisco, CA {praneeth, uday, shiva, [research}@onnes.ai](research}@onnes.ai) arXiv :2607 .05805v 1 [ cs .AI ] 7 Jul 2026 Abstract 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, and our headline result is parity without training: curated contrastive few-shot demonstrations and self-consistency voting raise a zero-shot LLM agent panel’s cryogenic fault-classification accuracy from 0.685 to 0.990—matching a supervised classifier (0 .985) with no parameter updates and only six labeled demonstrations—while a detector trained purely on real BlueFors telemetry posts a genuine real-hardware false-alarm rate of 6.4% . Onnes is a physics-grounded digital-twin simulator of a dilution refrigerator (a forward physics model with a learned real-fridge noise fingerprint, not a hardware-coupled bidirectional twin) that drives a live multi-agent large-languagemodel (LLM) operations layer. With it we run one of the first controlled head-to-head comparisons between a zero-shot LLM agent panel and a supervised machine-learning (ML) classifier on honestdiagnosis. The twin couples a real dilutioncooling floor, a noise-and-correlation fingerprint learned from real BlueFors logs, and six physics-grounded fault classes, three of them engineered to overlap on temperature but separate on flow and pressure. Across a 1000-turn evaluation, the zero-shot panel shows no statistically significant difference from the classifier on fault detection but trails on classification, its errors concentrating on the engineered confusable faults; the in-context lift closes exactly those cells, and an ablation attributes the gain almost entirely to the demonstrations. Run as a continuous monitor across a nine-run fault×seed sweep, the agent catches every developing fault within one poll interval, and a confidence gate suppresses pre-onset false alarms whose rate we find to be backend-dependent. As a first sim-to-real check we take the initial stage of our validation plan off the roadmap and run it: the same detector reaches 100% recall on physics faults injected onto real held-out windows, so its low false-alarm rate does not come at the cost of missed faults. All numbers are drawn verbatim from released run logs. Keywords: dilution refrigerator, quantum computing infrastructure, digital twin, LLM agents, in-context learning, self- consistency, fault diagnosis, label efficiency, supervised machine learning, anomaly detection, sim-to-real transfer 1 Introduction Superconducting and spin-qubit quantum processors operate at the base temperature of a dilution refrigerator, typically 10–35 mK at the mixing chamber (MXC) . The operational reality that shapes fault diagnosis here is threefold: cool-downs are multi-day and costly, so downtime is expensive and fridge health sits on the critical path of every experiment; hardware faults (leaks, blockages, quenches) are rare and often oneoff, so labeled fault episodes are scarce; and each new fridge is commissioned with no fault history of its own. In practice, monitoring is still built on threshold and rate-of-change alarms surfaced to a dashboard: they answer is a channel out of band? but not which physical fault is developing and what should the operator do? Closing that gap, turning telemetry into a named diagnosis and a recommended action under real-world label scarcity, is the problem we study. We take the position that this is a natural task for an LLM agent: the input is a modest, heterogeneous, multi-channel time series; the output is a structured judgment (detected / class / severity / action); and the reasoning be

FAQs

What problem does Onnes address in cryogenic fault diagnosis for quantum computing?

Onnes targets the gap in which current monitoring mainly triggers threshold alarms and lacks named, actionable physical fault classification under severe label scarcity.

How does Onnes achieve “parity without training” for zero-shot LLM agent panels?

It uses curated contrastive few-shot demonstrations plus self-consistency voting to boost accuracy from 0.685 to 0.990 without parameter updates, using only six labeled demonstrations.

What is the role of the physics-grounded digital-twin simulator in Onnes?

The twin simulates a dilution refrigerator using a forward physics model, a learned real-fridge noise fingerprint from BlueFors telemetry, and six physics-grounded fault classes engineered to be confusable in temperature but separable in flow and pressure.

Jul 16, 2026

Upload by Theodore