AI-driven quantum circuit compiler using Graph Neural Networks + Claude LLM — tested on real research circuits
quantumopt takes your quantum circuit and returns an optimized version that runs more accurately on IBM quantum hardware — with a plain English explanation you can cite in your paper.
pip install quantumoptfrom qiskit import QuantumCircuit
from quantumopt import compile
# Your research circuit
qc = QuantumCircuit(5)
qc.h(0)
qc.cx(0, 1)
qc.cx(1, 2)
qc.ry(0.5, 0)
qc.rz(0.3, 1)
# Compile and optimize
result = compile(qc, hardware="ibm_brisbane")
# Results
print(result.depth_reduction) # "31%"
print(result.gate_reduction) # "32%"
print(result.explanation) # Claude-generated report
print(result.optimized_circuit) # Ready to run on IBMTested on 41 real circuits from QASMbench (published research circuits):
| Metric | quantumopt | Baseline |
|---|---|---|
| Avg depth reduction | 13.2% | 0% |
| Avg gate reduction | 15.2% | 0% |
| Circuits improved | 34/41 | N/A |
| Circuits made worse | 0/41 | N/A |
| Best result | 89% | N/A |
Tested on 10,240 synthetic circuits:
| Metric | Result |
|---|---|
| GNN prediction accuracy (±10%) | 82% |
| GNN prediction accuracy (±20%) | 100% |
| Avg predicted improvement | 64.5% |
When ANTHROPIC_API_KEY is set, quantumopt generates
research-quality explanations:
"Transpilation of the target circuit for IBM Brisbane hardware yielded a 31.6% reduction in circuit depth (128 → 88 layers) and a 31.6% reduction in total gate count (326 → 223 gates). Among the applied optimization passes, merge_rotations contributed most substantially, as consecutive single-qubit rotation gates collapse into single parametrized operations. The resulting reduction in two-qubit gate exposure is particularly consequential for hardware execution, as each eliminated layer directly reduces coherence-time consumption against Brisbane's median T₂ timescales (~100–200 µs)."
- Your circuit is encoded as a Directed Acyclic Graph
- A trained GNN (82% accuracy) predicts optimization potential and recommends optimization actions
- Qiskit transpiler optimizes for target hardware
- Claude generates a hardware-specific explanation you can cite in your paper
VQE, QAOA, QFT, Grover, GHZ, Bernstein-Vazirani, Deutsch-Jozsa, Amplitude Estimation, Phase Estimation
- IBM Brisbane (default)
- More backends coming
# Required for AI explanations
export ANTHROPIC_API_KEY=your-key-here
# Optional — for real IBM hardware execution
export IBM_QUANTUM_TOKEN=your-token-here- Python 3.10+
- Qiskit >= 1.0.0
- PyTorch >= 2.0.0
- torch-geometric >= 2.4.0
- anthropic >= 0.20.0 (optional, for explanations)
If you use quantumopt in your research please cite:
Syamala, N. (2025). quantumopt: An AI-driven quantum
circuit compiler using Graph Neural Networks and Large
Language Models. GitHub.
https://github.com/nsyamala1/quantumopt
MIT License — free for research and commercial use.
Built by Naveen Syamala GitHub: github.com/nsyamala1 Issues: github.com/nsyamala1/quantumopt/issues