Cross-platform hardware benchmark of style-based quantum GANs for data augmentation on superconducting and trapped-ion processors

Abstract

In the noisy intermediate-scale quantum era, controlled benchmarks of quantum machine-learning workloads across hardware modalities are needed to quantify how given algorithms behave under native provider execution stacks. This work presents such a benchmark for the style-based quantum generative adversarial network (qGAN) on a high-energy physics data-augmentation task. We compare two commercially available gate-model quantum computers: the IBM ibm_torino hardware, based on superconducting transmon qubits from the Heron chip and the IonQ aria-1 hardware, based on trapped-ion qubits. The generator architecture and trained parameters are kept fixed, and built-in mitigation is disabled when possible. We report quality and runtime metrics under each provider's native stack. The workflow uses circuit replication across available qubits, up to 48 on IBM and 24 on IonQ, to reduce the number of submitted jobs required for the target sample set. To our knowledge, this is one of the first controlled style-based qGAN hardware-to-hardware comparisons for this data-augmentation task. We observe that both platforms complete the task successfully, with marginal Kullback-Leibler divergences somewhat lower on aria-1, while end-to-end runtime is significantly shorter on ibm_torino. These results are an application-specific tradeoff benchmark, not a claim of algorithmic novelty.