Quantum computers depend on moving information through tiny, noisy systems where small errors can compound rapidly. One way to build such a device is to arrange qubits in a ring and steer information from one site to another by tuning a fixed magnetic field at each site; each design is called a controller. The most useful way to score a controller is not its noise-free behaviour but how well it holds up when the system is being disturbed by random environmental noise. The standard score, the first-order Robustness-Infidelity Measure (RIM₁), is the controller's mean error under dephasing. Small values indicate robust designs, and large values indicate fragile ones. Computing it directly takes expensive Monte Carlo simulations, which makes screening many designs slow.

This project trains a machine-learning surrogate that predicts RIM₁ in milliseconds, using a published dataset of 1,500 controllers for 5- and 6-qubit rings from the Cardiff–Swansea–USC quantum control group. Two models are compared: a Support Vector Regressor fed hand-crafted physics features, and a Convolutional Neural Network fed the raw controller and metadata. Both predict robustness accurately on held-out designs, with the simpler kernel model slightly outperforming the deep one on this small dataset.

The project then asks what the network has learned. A Sparse Autoencoder, a recent interpretability technique borrowed from large language model research, decomposes the network's internal state into a small set of features. Statistical tests and causal interventions together show that the network's representation is organised around physically meaningful quantities, combined non-linearly in ways that no single hand-crafted feature captures. The full pipeline runs end-to-end on a laptop in a couple of minutes.