End-to-End Demonstration of Quantum Generative Adversarial Networks for Steel Microstructure Image Augmentation on a Trapped-Ion Quantum Computer
Samwel Sekwao · 2025
Read End-to-End Demonstration of Quantum Generative Adversarial Networks for Steel Microstructure Image Augmentation on a Trapped-Ion Quantum Computer by Samwel Sekwao (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2504-08728.txt.
Introduction
Steel inspection images are scarce for some crystal phases, and fake images from a quantum-assisted generator may fill the gap. Samwel Sekwao and co-authors ran that generator on real quantum hardware in this 2025 paper. The images show electron-scattering maps with five channels, covering two steel phases. A quantum circuit feeds the input layer of a large classical generator.
The hybrid design fought the usual collapse where a generator repeats one pattern. Image quality beat a purely classical baseline. A trapped-ion quantum computer ran the circuit end to end. Materials data sets are often small and hard to expand experimentally. Synthetic augmentation could speed high-throughput screening of alloys.
Catalog record
| CREATOR | Samwel Sekwao |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 43 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2504-08728
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
End-to-End Demonstration of Quantum Generative Adversarial Networks for Steel Microstructure Image Augmentation on a Trapped-Ion Quantum Computer Samwel Sekwao Abstract Abstract Generative adversarial networks (GANs) are a machine learning technique capable of producing high-quality synthetic images. In the field of materials science, when a crystallographic dataset includes inadequate or difficult-to-obtain images, synthetic images can be used for image augmentation to mitigate data scarcity and streamline the preparation of datasets for high-throughput analysis. We integrate quantum computing with GANs into a hybrid quantum-classical GAN to generate complex 5-channel electron backscatter diffraction (EBSD) images of two distinct microstructure phases of steel. By training a quantum circuit at the input layer of a large classical Wasserstein GAN (WGAN) model, we achieve higher image quality compared to a baseline classical GAN. The choice of WGAN also helps to mitigate mode collapse. We generate images from both ferrite and bainite microstructure phases in an end-to-end workflow. With respect to maximum mean discrepancy score, we find that the hybrid quantum-classical WGAN improves over classical Bernoulli GANs in 70% of samples. As the quantum computer is part of the training procedure, our method has potential to scale to larger number of qubits. Our results indicate that the WGAN model based on the quantum circuit ansatz may be effectively leveraged to enhance the quality of synthetic EBSD images on both quantum simulators and actual quantum hardware. Introduction The field of quantum machine learning (QML) has seen significant interest in recent years. A vast body of literature has emerged, demonstrating the potential of training parameterized quantum circuits to learn from data [ 11 , 35 , 21 , 43 , 16 , 31 ] . While there have been some efforts to understand the practical aspects of applying these methods to real world problems [ 18 , 44 , 30 , 32 ] , much of the existing work has been limited to either highly simplified problems or to primitive datasets where classical artificial intelligence (AI) methods already excel. In the field of generative modeling, quantum circuits are able to sample from probability distribution which are provably hard to generate using classical methods [ 49 , 28 , 20 , 12 ] . When applied to the task of modeling classical probability distributions, these models have shown evidence of improved expressivity and generalization capabilities, leading to superior models when evaluated on a parameter-by-parameter basis [ 49 , 23 , 24 , 15 ] . On the other hand, QML models have been limited in their ability to challenge modern classical generative machine learning (ML) models due to the limited qubit count and circuit depth available on modern devices in the noisy intermediate-scale quantum (NISQ) era [ 38 ] . Most quantum devices available to date are limited to a \mathcal{O}(10^{2}-10^{3}) qubit count and can only execute a few hundred two-qubit gate operations before noise overshadows the computation. One potential path forward is through the hybrid integration of both classical and QML models. In particular, one may include a quantum circuit as a component within a much larger hybrid quantum-classical ML model. The hybrid integration approach, introduced in Ref. [ 40 ] , leverages a parameterized quantum circuit to generate a correlated input distribution of binary strings. This distribution serves as the random latent vector input for a larger classical generative adversarial network (GAN) model. In contrast, the standard GAN implementation introduces randomness to the model by sampling a latent vector from a standard multi-dimensional probability distribution. The quantum approach is based on the associative adversarial network GAN (AAN-GAN) model. By sampling the input latent vector from a trained correlated distribution, the model performance can be improved [ 5 , 6 ] . In particular, the input distribution is trained to mimic the distribution of the penultimate output layer of the discriminator portion of the GAN, sampled over the set of real and generated images. In [ 40 ] , the authors demonstrated that a quantum circuit can effectively generate such correlated distribution over multi-dimensional binary strings, producing images with improved quality over the standard GAN approach for generating images of handwritten digits. In this work, we improve and expand upon the hybrid quantum-classical approach to the GAN model in several key directions. First, we train the quantum layer of the hybrid model to directly optimize the GAN loss function, as opposed to matching the output of the discriminator layer. Second, we apply this hybrid quantum-classical GAN model to a much more complicated data set. To accommodate this significant leap in complexity, we adapt the classical model to incorporate the recent advancements in the base GAN approach. To this end, we apply the Wasserstein GAN (WGAN) methodology to generate synthetic set of complex 5-channel electron backscatter diffraction (EBSD) images from different phases of steel microstructures. We would like to stress that these are not simply superficial differences. We will demonstrate that there are inherent limitations to integrating the quantum AAN-GAN routine to more complex hybrid models, and we suggest that the improvement observed in the previous works may have partly resulted from the simplicity of the classical model. Our implementation solves these issues and integrates the quantum and classical models tightly, making them inseparable into different components. In this paper, we will describe our improved hybrid approach to the quantum-classical GAN model and through numerical experiments we demonstrate that the inclusion of a quantum layer consistently improves over the best classical results. We achieve this through classical simulations of the quantum circuit layer and actual execution of our algorithm on a trapped-ion quantum computer. We validate the robustness of our method by training the entire model end-to-end on the actual quantum hardware. Overall, we believe that our new implementation provides a path towards scaling this hybrid quantum-classical GAN methodology to state-of-the-art levels. Quantum Wasserstein GANs have been used in other application contexts, including option pricing [ 22 ] , high-energy physics [ 9 ] , materials science [ 33 ] , anomaly detection [ 29 ] , and theoretical computer science [ 17 ] . The rest of this paper is structured as follows. In section 2, we present a brief introduction to GANs, WGANs, and quantum-circuit Born machines (QCBMs). We also describe the methodology used to train quantum enhanced WGANs, and introduce the metric used to evaluate GANs/WGANs trained on image data. Section 3 introduces the back-scatter diffraction images used to train the classical and quantum models. In section 4, we compare the performance of the classical models to the quantum enhanced models trained on both an ideal simulator and IonQ’s Aria-2 quantum processing unit (QPU) and describe briefly the experimental setup. Methods Generative Adversarial Networks (GANs) GANs [ 26 ] are one of the most popular forms of generative ML models in use today. GANs have shown a significant ability to produce highly realistic images and have also been applied to generate other types of synthetic data [ 39 , 14 , 48 , 41 , 19 ] . The prototypical GAN comprises two components: a generator ( G ) and a discriminator ( D ), both implemented as artificial neural networks. The generator’s objective is to generate synthetic images from a random input sampled from a latent