| name | onnes-llm-cryogenic-diagnosis |
| description | Physics-grounded digital twin + multi-agent LLM methodology for fault diagnosis in critical infrastructure. Uses a forward physics model with learned noise fingerprint to drive LLM agents for diagnostic tasks. Trigger words: fault diagnosis, digital twin, multi-agent LLM, cryogenic, physics-grounded, diagnostic simulator.
|
Onnes: Physics-Grounded Multi-Agent LLM Simulator for Fault Diagnosis
Source
- Paper: Onnes: A Physics-Grounded Multi-Agent LLM Simulator for Cryogenic Fault Diagnosis in Quantum Computing Infrastructure
- arXiv: 2607.05805v1 (2026-07-07)
- Authors: Praneeth Narisetty, Uday Kumar Reddy Kattamanchi, Shiva Nagendra Babu Kore
- Categories: cs.AI, cs.LG, quant-ph
Methodology
A unified framework combining physics-based simulation with multi-agent LLM reasoning for infrastructure fault diagnosis.
Core Architecture
┌─────────────────────────────────────────┐
│ Physics-Grounded Twin │
│ ┌───────────────────────────────────┐ │
│ │ Forward Physics Model │ │
│ │ + Learned Noise Fingerprint │ │
│ │ (from real hardware logs) │ │
│ └───────────────────────────────────┘ │
│ ↓ │
│ ┌───────────────────────────────────┐ │
│ │ Fault Classes (physics-grounded) │ │
│ │ - Some overlap on temperature │ │
│ │ - Separate on flow/pressure │ │
│ └───────────────────────────────────┘ │
└─────────────────────────────────────────┘
↓
┌─────────────────────────────────────────┐
│ Multi-Agent LLM Layer │
│ ┌─────────────┐ ┌─────────────┐ │
│ │ LLM Agent 1 │ │ LLM Agent 2 │ ... │
│ └─────────────┘ └─────────────┘ │
│ ↓ │
│ Confidence Gate (suppresses │
│ pre-onset false alarms) │
└─────────────────────────────────────────┘
Key Components
-
Digital Twin Simulator:
- Forward physics model of the target system
- Learned noise-and-correlation fingerprint from real operational logs
- Physics-grounded fault classes (engineered to be distinguishable)
-
Multi-Agent LLM Operations Layer:
- Zero-shot LLM agent panel for diagnosis
- Few-shot contrastive demonstrations for confusable faults
- Self-consistency voting across agents
-
Confidence Gate:
- Suppresses pre-onset false alarms
- Backend-dependent false alarm rate calibration
Performance Results
- Zero-shot panel: no significant difference from supervised classifier on detection
- With curated contrastive few-shot demonstrations + self-consistency voting:
- Classification accuracy: 0.685 → 0.990 (matching supervised classifier at 0.985)
- Requires only 6 labeled demonstrations, no parameter updates
- Continuous monitoring: catches every developing fault within one poll interval
- Sim-to-real: detector trained on real telemetry achieves 6.4% false-alarm rate, 100% recall
Implementation Workflow
class PhysicsGroundedTwin:
def __init__(self, physics_model, noise_fingerprint, fault_classes):
self.physics = physics_model
self.noise_fp = noise_fingerprint
self.faults = fault_classes
def simulate(self, scenario, fault_type=None):
"""Run physics simulation with optional fault injection."""
base = self.physics.forward(scenario)
noise = self.noise_fp.sample()
if fault_type:
fault_signal = self.faults.apply(fault_type, base)
return base + noise + fault_signal
return base + noise
class LLMDiagnosticAgent:
def __init__(self, llm, demonstrations=None):
self.llm = llm
self.demos = demonstrations or []
def diagnose(self, observation, use_demos=True):
"""Diagnose fault from observation."""
prompt = self._build_prompt(observation, use_demos)
return self.llm.generate(prompt)
def self_consistency_vote():
votes = [.diagnose(observation) _ (n_votes)]
._majority_vote(votes)
:
():
.threshold = threshold
():
confidence >= .threshold
Few-Shot Demonstration Strategy
- Contrastive Pairs: Select examples from confusable fault classes
- Minimal Set: 6 demonstrations sufficient to match supervised baseline
- Ablation: Performance gain attributed almost entirely to demonstrations
Application Domains
- Cryogenic systems: Dilution refrigerator fault diagnosis
- Industrial IoT: Equipment failure prediction
- Medical diagnostics: Multi-agent symptom analysis
- Infrastructure monitoring: Anomaly detection with physics constraints
Best Practices
- Learn noise fingerprint from real data — do not use synthetic noise alone
- Design fault classes to overlap on some dimensions but separate on others
- Use few-shot demonstrations for confusable cases, not all cases
- Implement confidence gating to suppress false alarms
- Validate sim-to-real transfer with held-out real data
Activation Keywords
fault diagnosis, digital twin, multi-agent LLM, cryogenic, physics-grounded, diagnostic simulator, noise fingerprint, contrastive demonstrations, confidence gate, self-consistency voting, sim-to-real