Nighthawk-QCNN

96-qubit Quantum Convolutional Neural Network (QCNN)

Trained end-to-end on real IBM Quantum Heron r2/r3 hardware

(backend: ibm_fez/ibm_kingston) on February 4, 2026.

Task

Binary classification of parity of random Pauli-X excitations in 1D cluster state
(0 — even number → trivial state, 1 — odd number → non-trivial).

Technical Details

• Qubits: 96 (actively used in ansatz + preparation)
• Architecture: QCNN with 3 layers (conv → pool → conv → pool → conv → readout)
• Convolution operator: 4-parameter 2-qubit block (RY, RZ, CZ), shared parameters
• Pooling: static (measure + CZ, no conditional X due to compiler limitations)
• Readout: Z-probability on final qubit → MSE loss
• Trainable parameters: 72 (8 per layer × 3)
• Dataset: 24 samples (on-the-fly generation)
• Shots per evaluation: 384
• Optimizer: SPSA, 12 iterations
• Final loss (MSE): 0.2704 (after 36 evaluations)
• QPU time: ~7 minutes (IBM Heron r2/r3)
• Backend: ibm_fez (156 qubits, heavy-hex lattice, tunable couplers)

Training Convergence

MSE loss starts at ~0.268, dips to ~0.243 around evaluation 1.0, then rises again due to noise accumulation.

Repository Files

• Nighthawk.npy — trained parameters (72 values)
• qcnn.qasm — QASM3 description of the ansatz (parameter-free)
• results.csv — final training metrics
• training_log.txt — full log of loss evaluations and transpilation
• requirements.txt — dependencies for reproduction

Usage / Inference

from qiskit import qasm3
import numpy as np

# Load model
theta = np.load("Nighthawk.npy")
qcnn = qasm3.loads(open("qcnn.qasm").read())
qcnn.assign_parameters(theta)

print("Model loaded. Number of parameters:", len(theta))
# Next: compose with preparation circuit + run via Sampler 

Notes

• Proof-of-concept for scaling QCNN on NISQ hardware in 2026.
• Loss near random guess (0.25) due to high noise on Heron r2 — typical for NISQ.
• Why better results expected on ibm_miami (Nighthawk r1):
  • Square lattice topology (vs heavy-hex on Heron r2) → much better natural locality for convolutional layers
  • Higher CLOPS and lower gate errors → deeper circuits with less decoherence
  • Improved connectivity → fewer SWAPs during transpilation → lower overall error accumulation
  • Expected: noticeably lower final loss and higher effective classification accuracy
• Improvements: more shots, error mitigation (twirling/M3), run on Nighthawk (square lattice).

Pro Mundi Vita