Transmon-QKD-ML

What I built

• Research supervised by Aswath Babu (IIIT Dharwad) and Prof. V. Ravishankar (IIT Delhi), written up with Aswath Babu as "Predicting BB84 Error Rates from Transmon Calibration Data: Learned Surrogates versus Physical Baselines on a 156-Qubit IBM Heron Processor".
• Over 25 days, logged 115 calibration snapshots (17,940 per-qubit records) of the IBM Heron r2 processor ibm_marrakesh and ran 105 BB84 experiments plus 90 Bell-state tests, collected by an automated GitHub Actions pipeline.
• Showed that the Monte-Carlo training simulator has an exact closed form at the label-noise ceiling (R² 0.989 vs 0.990), so ML surrogates trained on it cannot add information.
• On hardware, calibrated readout error alone predicts QBER with Spearman ρ = 0.61 and median error 0.0076, against ρ = 0.47 and 16 false abort flags for the simulation-trained random forest.
• Measured large calibration drift (median Jaccard 0.25 for the top-10 qubits across snapshots under a day apart) and recommended latest readout calibration plus a short pre-flight check for qubit selection. Data and code are open.

Stack