BS–MS thesis
Entanglement detection and gate optimization in the system of three and four NMR qubits using deep learning and machine learning
S. Siyanwal
Overview
This thesis applies machine learning and deep learning to two tasks in NMR quantum computing. The first is detecting and classifying entanglement in two-, three- and four-qubit states. The second is optimizing quantum gates by designing radio-frequency pulse sequences with a genetic algorithm.
Problem / motivation
Entanglement witnesses detect only some entangled states. Full state tomography detects them all, but the number of measurements it needs grows exponentially with system size. Separately, implementing high-fidelity gates on NMR hardware requires breaking each gate into an optimized sequence of pulses.
Methods
Neural networks (Keras/TensorFlow) trained on simulated pure-state density matrices: 15,000 states (2 qubits), 100,000 (3 qubits) and 70,000 (4 qubits). Labels were entangled vs separable for 2 qubits, and separability classes for 3 and 4 qubits.
K-nearest neighbours and PCA + K-means on 50,000, 100,000 and 70,000 simulated pure states; PCA + K-means was also applied to 54 experimental two-qubit NMR states provided by V. Gulati, and their theoretical counterparts. A GAN with a complex-valued pseudo-siamese network was used for two-qubit states.
Gate optimization: a genetic algorithm, following the approach of Devra et al., that searches over sequences of hard pulses and delays (pulse width, phase and delay), using gate fidelity as the fitness function.
Results / outcomes
Neural networks: validation accuracy of 0.9997 (2 qubits), 0.9971 (3) and 0.9996 (4) on simulated pure states. They could not separate states within the fully entangled class, such as GHZ and W states.
K-nearest neighbours: best accuracy of 0.9981 (2 qubits, k = 3), 0.9862 (3 qubits, k = 1 or 3) and 0.9619 (4 qubits, k = 15 or 17).
PCA + K-means formed cluster boundaries for 2-, 3- and 4-qubit states; how accurate those boundaries are was left for further study. The GAN + pseudo-siamese network, trained only on separable two-qubit states, classified entangled states in the test set.
Genetic-algorithm pulse design: for a two-qubit homonuclear system, the thesis reports a gate fidelity of 0.8513 (about 8 hours of computation). It was not fully optimized and was not tested on the NMR hardware.
Open problems identified: training on noisy states, and working from incomplete tomography. The three-qubit classification on experimental NMR data was developed further into the published paper.