🔬 materials science

A Hybrid Quantum Neural Network to Analyse Big Experimental Powder X-ray Diffraction Data

This paper introduces a hybrid quantum neural network framework that successfully extracts quantitative phase parameters from large-scale experimental powder X-ray diffraction datasets without iterative refinement, demonstrating its practical utility by accurately reconstructing spatial phase maps on an IBM quantum computer for complex solid oxide fuel cell and lithium-ion battery materials.

Original authors: H. Dong, S. D. M. Jacques, M. Q. Hlatshwayo, E. Papoutsellis, K. Georgopoulos, A. M. Beale, A. Vamvakeros

Published 2026-09-22

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: H. Dong, S. D. M. Jacques, M. Q. Hlatshwayo, E. Papoutsellis, K. Georgopoulos, A. M. Beale, A. Vamvakeros

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

In the world of materials science, understanding what something is made of and how its parts are arranged is often a matter of shining a powerful beam of X-rays through it. When these rays hit a crystalline solid, they scatter in specific patterns, creating a unique fingerprint that reveals the material's internal structure. For decades, scientists have used these fingerprints to analyze everything from new medicines to advanced batteries. However, modern research facilities can now generate these patterns at a staggering speed, producing hundreds of thousands of images in a single experiment. This flood of data has created a bottleneck; traditional computers struggle to process such massive volumes quickly enough to keep up with the pace of discovery, forcing researchers to wait or manually sift through information that should be analyzed in real time.

To address this growing challenge, a team of researchers has taken a bold step by testing a new kind of computer processor—one that operates on the principles of quantum mechanics rather than the standard binary logic of everyday machines. They did not simply simulate this process on a regular computer; they actually ran their analysis on a physical quantum computer located at a research facility in the UK. Their goal was to see if this emerging technology could decode complex X-ray patterns from real-world energy devices, specifically a solid oxide fuel cell and a lithium-ion battery, and determine the precise amount of each material present within them. The results showed that while current quantum hardware is still in its early stages, it is already capable of handling real experimental data, successfully mapping out the internal composition of these devices with an accuracy that matches established, traditional methods.

The researchers focused on two specific energy technologies: a micro-monolithic solid oxide fuel cell, which is a type of clean energy generator, and a commercial pin-type lithium-ion battery, the kind found in many portable electronics. Both devices are complex mixtures of different solid materials layered together. To understand how they work or how they degrade, scientists need to know exactly where each material is located and in what proportion. Traditionally, this requires a painstaking process called iterative refinement, where computers repeatedly adjust a model until it fits the X-ray data. This method is accurate but slow, especially when applied to the tens of thousands of data points collected in a single scan of a fuel cell or battery.

The team developed a new approach that combines a small quantum processor with a standard classical computer. They designed a system where the quantum part acts as a specialized feature extractor. Imagine the X-ray pattern as a long, complicated list of numbers. The quantum processor takes this list and compresses it into a much smaller set of numbers that still holds the essential information about the material's structure. This compressed information is then passed to a standard computer program, which quickly translates it into a map showing the weight and distribution of each material. To make this work on real, imperfect quantum hardware, the researchers first trained the system on a massive library of simulated data that included artificial noise, mimicking the errors that occur in physical quantum machines. Once the system learned to handle this noise in simulation, they transferred it to the actual quantum computer to analyze real experimental data.

The experiment involved analyzing roughly 10,000 X-ray patterns from the fuel cell and about 18,000 patterns from the battery. The researchers ran these patterns through an IBM quantum computer, which processed the data in a matter of minutes. The output was a detailed spatial map showing the distribution of the different phases within the devices. For the fuel cell, the system successfully distinguished between the anode, the electrolyte, and the cathode layers, clearly identifying the porous support structure and the thin layers of active material. For the battery, it mapped out the spiral structure of the jelly-roll design, separating the lithium cobalt oxide cathode, the graphite anode, and the thin metal foils that collect the electrical current.

When the team compared the maps generated by the quantum system to those produced by the traditional, gold-standard method of analysis, the results were strikingly similar. The quantum approach correctly identified the boundaries between different materials and accurately measured the relative amounts of each component. In the fuel cell, the error rates were minimal across the bulk of the material, with only slight discrepancies at the very sharp edges where signals are naturally weaker. In the battery, the system managed to separate the four distinct phases, including the metals and the carbon-based materials, with a high degree of precision. The system also successfully recovered scale factors, which indicate the intensity of the diffraction signals, allowing it to account for the vast differences in how strongly different materials scatter X-rays.

This work does not claim that quantum computers are currently faster or more accurate than the best classical supercomputers for this specific task. In fact, the researchers note that the quantum hardware is still subject to physical limitations and noise. Instead, the study serves as a proof of concept, demonstrating that physical quantum processors can be integrated into real-world scientific workflows to process large, complex datasets. By showing that a hybrid system can successfully extract quantitative parameters from real experimental data without needing to discard the information or rely solely on simulation, the team has laid a foundation for future applications. As quantum hardware continues to improve with more stable connections and better error correction, this method could eventually allow scientists to analyze massive streams of data in real time, unlocking new possibilities for understanding the materials that power our world.

Technical Summary: A Hybrid Quantum Neural Network to Analyse Big Experimental Powder X-ray Diffraction Data

Problem Statement
The quantitative analysis of experimental powder X-ray diffraction (XRD) data for complex multiphase materials is increasingly challenged by the sheer volume of data generated by modern facilities (e.g., synchrotrons), which can produce hundreds of thousands to millions of diffraction patterns per experiment. Conventional CPU-based approaches, including non-linear least-squares peak fitting and Rietveld refinement, face severe computational bottlenecks when processing these high-throughput datasets. While GPU-based machine learning offers automation, it remains constrained by classical bit architectures and requires significant resources. Furthermore, existing methods often struggle with the noise and complexity inherent in chemical imaging data like X-ray diffraction computed tomography (XRD-CT). The paper posits that novel data processing paradigms transcending standard von Neumann scaling limits are necessary, suggesting Quantum Machine Learning (QML) as a potential framework to map classical information into high-dimensional Hilbert spaces for efficient feature extraction.

Methodology
The authors propose a Hybrid Quantum Neural Network (HQNN) framework designed to extract quantitative parameters (phase weight fractions and scale factors) directly from one-dimensional powder diffraction patterns without iterative refinement. The architecture and workflow are defined as follows:

Network Architecture: The model utilizes a 6-qubit parametrised quantum circuit (PQC) acting as a non-linear feature extractor. The circuit consists of two repetitions of single-qubit rotational gates ( $R_x, R_y, R_z$ ) across all six qubits, followed by a ring-entanglement topology using Controlled-NOT (CX) gates. This yields 36 trainable quantum parameters.
Feature Extraction: The quantum circuit processes amplitude-encoded, normalised spectral vectors. Expectation values for Pauli observables ( $\langle X_i \rangle, \langle Y_i \rangle, \langle Z_i \rangle$ ) are measured for each qubit, generating an 18-dimensional feature vector.
Classical Decoding: These quantum features are fed into a classical dual-head Multi-Layer Perceptron (MLP).
- The Weight-Fraction Tower takes the 18 quantum features and outputs phase-specific weights constrained by a softmax function ( $\sum w_i = 1$ ).
- The Scale-Factor Tower concatenates the 18 quantum features with the pattern's maximum intensity (retained as a classical input) to predict unconstrained scale factors.
Training and Calibration Strategy: To address hardware noise and decoherence on Noisy Intermediate-Scale Quantum (NISQ) devices, the authors employ a four-stage workflow:
1. Noise-Aware Pre-training: The model is pre-trained on 200,000 synthetic diffraction patterns using a classical simulator with stochastic Pauli error injection (simulating gate noise) and finite-shot sampling.
2. Hardware Feature Acquisition: The quantum circuit parameters are frozen. The circuit is executed on physical IBM quantum processors (ibm_boston and ibm_kingston) to acquire feature vectors from a calibration set of 20,000 synthetic patterns.
3. Classical Head Fine-tuning: Only the classical MLP weights are retrained using the hardware-derived features to adapt to specific backend distortions.
4. Experimental Inference: The calibrated model is applied to experimental XRD-CT data.
Datasets: The framework was tested on two experimental systems: a three-phase solid oxide fuel cell (SOFC) containing LSM, YSZ, and NiO (~~10,000 patterns) and a four-phase lithium-ion battery containing LCO, graphite, Al, and Cu (~~20,000 patterns). Synthetic libraries were generated with known ground-truth parameters for training and calibration.

Key Results

Quantitative Accuracy: On held-out synthetic test sets, the hardware-calibrated HQNN demonstrated high correlation with ground truth. For the SOFC, $R^2$ values were 0.963 (LSM), 0.968 (YSZ), and 0.895 (NiO). For the battery, the model achieved a uniform $R^2$ of 0.943 across all four phases for scale factor prediction.
Spatial Reconstruction: When applied to experimental XRD-CT data, the HQNN successfully reconstructed spatial phase maps that showed strong structural alignment with conventional Rietveld refinement. The network accurately resolved microstructural boundaries, such as the porous anode matrix and the thin cathode layer in the SOFC, and the spiral jelly-roll architecture in the battery.
Error Analysis: Relative errors between HQNN predictions and Rietveld refinement were generally minimal (<5–10%) across bulk active materials. Discrepancies were primarily localized to sharp interfaces and boundaries with low signal-to-noise ratios.
Computational Efficiency: While classical pre-training took hours, the physical quantum feature acquisition was completed in approximately 47 minutes for the SOFC dataset and 64 minutes for the battery dataset on IBM hardware (covering 20,000 simulated calibration patterns plus all experimental patterns for each system). Classical head retraining and downstream inference were nearly instantaneous (milliseconds per pattern).

Significance and Claims
The paper positions this work as a foundational proof-of-concept rather than a demonstration of superior speed or raw accuracy compared to established classical deep learning (e.g., PQ-Net) or Rietveld refinement. The authors explicitly state that contemporary NISQ hardware is inherently slower and noisier than classical GPUs.

The primary significance claimed is:

First Direct Application: This represents the first direct application of physical quantum computing hardware to large-scale, real-world experimental materials characterisation data (tens of thousands of patterns), moving beyond isolated "toy" datasets.
Feasibility of Hybrid Pipelines: It demonstrates that hybrid quantum-classical architectures can successfully manage high-throughput data streams and execute quantitative parameter extraction on complex, multiphase physical systems within reasonable operational timeframes.
Robustness to Noise: The study validates that combining noise-aware simulator pre-training with fast downstream fine-tuning on QPU-derived observables can stabilize QNN inference against gate infidelities and shot noise.
Baseline for Future Development: The work establishes an empirical baseline for how variational quantum circuits handle real-world experimental data, providing a necessary step toward future applications where quantum processors might alleviate data congestion in materials science facilities.

The authors remain cautious, noting that theoretical challenges such as barren plateaus and data-loading bottlenecks persist, and that hardware maturation alone does not guarantee a provable quantum advantage. However, they argue that bridging the gap between abstract algorithms and tangible physical data processing is an essential step forward.

Technical Summary: A Hybrid Quantum Neural Network to Analyse Big Experimental Powder X-ray Diffraction Data

More like this