AI x Quantum — TU Delft / Quantum Inspire

How might generative AI accelerate quantum computing?

An open research initiative exploring AI-driven quantum computing research. We build autonomous agents that run quantum experiments across multiple hardware backends, replicate published papers, and benchmark LLM capabilities on quantum tasks.

Live site: https://quantuminspire.vercel.app

Current results

Experiments (22 results across 7 study types)

Study	Backends	Key result
Bell State Calibration	Emulator, IBM Marrakesh, IBM Torino, Tuna-9	100% / 99.05% / varies fidelity
GHZ State (3q)	Emulator, IBM Marrakesh, IBM Torino, Tuna-9	100% / 98.14% fidelity
H2 VQE (2q)	Emulator, IBM Marrakesh, IBM Torino, Tuna-9	-1.1385 Ha emulator (chemical accuracy)
QRNG Certification	Tuna-9 raw + debiased, Emulator	Raw fails NIST; debiased passes all
Randomized Benchmarking	Emulator	99.95% gate fidelity
QAOA MaxCut	Emulator	87% approximation ratio
Quantum Volume	Emulator + Tuna-9	QV 16 (4q pass, 8/10 circuits)

Additional hardware experiments: connectivity probe (Tuna-9 topology), repetition code (3q QEC), detection code (emulator).

Paper replications (3 papers, 13 claims)

Paper	Claims tested	Pass rate
Sagastizabal 2019 (H2 VQE)	7	43% (emulator pass, hardware fail)
Peruzzo 2014 (HeH+ VQE)	3	100% (emulator)
Cross 2019 (Quantum Volume)	3	100% (emulator)

Hardware access

Backend	Qubits	Access
QI Emulator (qxelarator)	Configurable	Local, no auth needed
QI Tuna-9	9 (6 usable)	QI member 2108
IBM Marrakesh	156	IBM Quantum (free tier, 10 min/month)
IBM Torino	133	IBM Quantum
IBM Fez	156	IBM Quantum

Quick start

# 1. Website
npm install
npm run dev
# Deploy
vercel --prod

# 2. Python environment (Python 3.9-3.13 supported — 3.14 breaks qxelarator)
python3 -m venv .venv
source .venv/bin/activate
pip install -r mcp-servers/requirements.txt

# 3. MCP servers for Claude Code (optional — starts automatically)
#    The .mcp.json in this repo configures three quantum MCP servers.
#    They use .venv/bin/python, so step 2 must be done first.
#    Auth setup (needed for hardware, not for emulator):
#      - Quantum Inspire: qi login
#      - IBM Quantum: python -c "from qiskit_ibm_runtime import QiskitRuntimeService; QiskitRuntimeService.save_account(channel='ibm_cloud', token='YOUR_TOKEN')"

# 4. Run experiments
source .venv/bin/activate
python scripts/benchmark_harness.py --limit 10
python scripts/replications/replicate_sagastizabal.py
python scripts/replications/replicate_peruzzo.py
python agents/experiment_daemon.py

# 5. Tests
npm test                              # JS/TS (Vitest)
python -m pytest tests/               # Python

Architecture

Website (Next.js 14 + Tailwind + Three.js)

Route	Description
/	Research home — hero, experiments overview, agent architecture
/experiments	Experiment dashboard — grouped by type, backend badges
/experiments/[id]	Study detail — abstract, research question, results, visualizations
/replications	Paper replication dashboard — claims vs measured, cross-backend
/blog	Research blog (7 posts)
/learn	Interactive quantum learning page
/bloch-sphere, /state-vector, etc.	Interactive quantum visualizations

Agents (agents/)

Agent	Purpose
orchestrator.py	Pipeline coordinator
experiment_daemon.py	Queue -> submit -> analyze -> store results
benchmark_agent.py	LLM benchmark runner
replication_agent.py	Paper registry + run/analyze replications
replication_analyzer.py	Compare results vs published claims
qec_decoder.py	Quantum error correction decoder

MCP servers (mcp-servers/)

Server	Purpose
qi-circuits	Submit/check circuits on Quantum Inspire hardware
qrng	Quantum random number generation
ibm-quantum	IBM Quantum hardware access

Experiment result JSON schema (v1.0)

All result files in experiments/results/ follow this schema:

{
  "schema_version": "1.0",
  "id": "bell-calibration-001-ibm",
  "type": "bell_calibration",
  "backend": "ibm_marrakesh",
  "backend_qubits": 156,
  "job_id": "d65kqpoqbmes739d1k2g",
  "submitted": "2026-02-10T15:24:38Z",
  "completed": "2026-02-10T15:24:38Z",
  "parameters": { "shots": 4096 },
  "raw_counts": { ... },
  "analysis": { ... },
  "circuit_cqasm": "version 3.0\n...",
  "errors": null
}

• schema_version: always "1.0"
• backend_qubits: qubit count of backend (null for emulators)
• job_id: hardware job ID (null for emulator/local runs)

Stack

• Quantum: Qiskit 2.1, PennyLane 0.44, QI SDK 3.5.1, OpenFermion, PySCF
• AI: Claude, Gemini, GPT (via respective APIs)
• Web: Next.js 14, Tailwind, Three.js
• Hardware: Quantum Inspire Tuna-9 (9q), IBM Marrakesh (156q), IBM Torino (133q), IBM Fez (156q)
• Python: 3.9-3.13 (3.14 breaks qxelarator)

Links

• Quantum Inspire
• QuTech
• TU Delft
• Qiskit HumanEval paper
• Sagastizabal et al. (2019)
• Peruzzo et al. (2014)
• Cross et al. (2019)