AI Workflow Stabilises Complex X-Ray Data Analysis

Scientists are tackling a persistent challenge in materials science: reliably determining the structure of materials from X-ray diffraction (XRD) data, as initial AI-driven proposals frequently encounter difficulties during refinement due to peak overlap and weak diffraction consistency. Bin Cao from the Guangzhou Municipal Key Laboratory of Materials Informatics, Advanced Materials Thrust, The Hong Kong University of Science and Technology (Guangzhou), Qian Zhang and Zhenjie Feng from the Materials Genome Institute, Shanghai University, alongside Taolue Zhang, Jiaqiang Huang, Lu-Tao Weng, and Tong-Yi Zhang, working collaboratively across the Guangzhou Municipal Key Laboratory of Materials Informatics, Advanced Materials Thrust, The Hong Kong University of Science and Technology (Guangzhou), Materials Genome Institute, Shanghai University, and Material Characterization and Preparation Facility, Hong Kong University of Science and Technology (Guangzhou), present a new workflow, WPEM, that integrates physics-based constraints into a robust refinement process. This research is significant because WPEM demonstrably improves the accuracy and stability of structure refinement, even with complex datasets containing overlapping peaks, multiple phases, or amorphous components, effectively bridging the gap between artificial intelligence-generated structural hypotheses and physically admissible solutions derived from diffraction data.

Within the cool darkness of the X-ray beamline, scattered signals hold the key to a material’s atomic arrangement. Extracting that information from complex diffraction patterns has long been a bottleneck in materials science. Now, a new method promises to unlock hidden structures with greater accuracy and speed, even in messy, real-world samples. Scientists are increasingly turning to artificial intelligence to accelerate materials discovery and understanding, yet a fundamental challenge remains in bridging the gap between AI-generated structural hypotheses and experimentally verifiable models.

X-ray diffraction (XRD) continues as the primary non-destructive method for probing atomic structure, providing quantitative data on crystal symmetry, lattice parameters, and phase composition. A new workflow, termed WPEM, addresses a critical limitation in current AI-driven crystallography: the enforcement of physical and crystallographic consistency during refinement.

Existing AI approaches often struggle when peak intensities become unstable due to severe overlap in diffraction patterns, leading to implausible structures propagating into materials databases. Instead of relying solely on numerical fitting, WPEM directly incorporates Bragg’s law, the fundamental principle governing diffraction, as an explicit constraint within a batch expectation, maximisation framework.

This approach models the full diffraction profile as a probabilistic mixture, iteratively refining component intensities while ensuring peak positions remain Bragg-consistent. As a result, WPEM generates a continuous, physically admissible intensity representation, proving stable even in heavily overlapped regions or when dealing with mixed radiation or multiple phases.

Benchmarking against standard reference materials, including lead sulfate (\ce{PbSO4}) and terbium barium cobalt oxide (\ce{Tb2BaCoO5}), reveals that WPEM achieves lower Rp/Rwp values than established packages like FullProf and TOPAS under comparable refinement conditions. Beyond these benchmarks, the workflow’s generality is demonstrated through diverse applications, ranging from decomposition of multiphase titanium-niobium thin films to quantitative analysis of sodium chloride, lithium carbonate mixtures.

Also, WPEM successfully separates crystalline signals from amorphous halos in semicrystalline polymers, enables high-throughput lattice tracking in layered cathode materials, automates refinement of compositionally disordered ruthenium-manganese oxide solid solutions (CCDC 2530452), and even deciphers the phase composition of an ancient Egyptian make-up sample using synchrotron powder XRD. Scientists developed WPEM, a physics-constrained workflow for whole-pattern decomposition and refinement, addressing limitations in conventional methods where peak overlap hinders accurate structural modelling.

This technique transforms Bragg’s law into an explicit constraint within a batch expectation, maximisation framework, allowing for more stable and reliable analysis. By modelling the full diffraction profile as a probabilistic mixture density, WPEM iteratively infers component-resolved intensities while maintaining Bragg-consistency of peak centres. Once background subtraction is completed, the observed intensity profile is approximated as a sum of peak-shape functions, each described by parameters defining its position, width, and integrated intensity.

Specifically, each peak-shape function utilises a thin-tailed pseudo-Voigt density, a combination of Lorentzian and Gaussian components, to accurately represent the peak’s form. Unlike traditional methods that often link Lorentzian and Gaussian widths, WPEM optimises these parameters independently, accommodating heterogeneous broadening and severe peak overlap.

This independent optimisation is achieved through closed-form update rules iterated within the expectation-maximisation process, refining peak positions, widths, and mixing coefficients. Beyond simply fitting the data, WPEM enforces the Bragg-law constraint on peak centres, ensuring that the derived structural information is physically admissible. By providing Bragg-consistent, uncertainty-aware intensity partitioning, WPEM enables the combination of diffraction data with other analytical probes like transmission electron microscopy. Researchers benchmarked WPEM against standard reference patterns of lead sulfate and terbium barium cobalt oxide, demonstrating improved performance under matched refinement conditions compared to widely used packages like FullProf and TOPAS. Specifically, refinement of the \ce{PbSO4} standard yielded Rwp values of 0.0083 and χ2 values of 1.21, while \ce{Tb2BaCoO5} resulted in Rwp of 0.011 and χ2 of 1.87. Decomposition of a multiphase Ti, 15Nb thin film successfully quantified the relative contributions of each crystalline phase present. Operando lattice tracking was facilitated, enabling high-throughput analysis of layered cathodes and monitoring lattice parameter changes during charge-discharge cycles.

Automated refinement converged to a reasonable structure with atomic displacement parameters within a compositionally disordered Ru, Mn oxide solid solution (CCDC 2530452). Beyond inorganic materials, the composition of an ancient Egyptian make-up sample was deciphered using synchrotron powder XRD, identifying key crystalline phases and their relative abundances. By providing Bragg-consistent intensity partitioning, WPEM offers a refinement-ready interface for challenging XRD data.

Workflow precisely validates material structures using probabilistic diffraction modelling

For decades, materials scientists have relied on X-ray diffraction to unlock the atomic structure of substances, yet interpreting the resulting patterns has remained a surprisingly manual and often imprecise process. A new workflow called WPEM promises to automate and refine this analysis, moving beyond simply identifying potential structures to rigorously validating them against the fundamental laws of physics.

Instead of treating diffraction data as a collection of peaks, WPEM views it as a continuous distribution, allowing it to resolve overlapping signals and account for imperfections that typically plague real-world samples. Conventional methods struggle when signals from different phases or amorphous components become intertwined. However, WPEM’s approach, explicitly incorporating Bragg’s law into its calculations, creates a more stable and reliable intensity representation.

By modelling the full profile probabilistically, the system iteratively refines component intensities, offering a solution where standard packages often falter. Initial tests against known materials demonstrate improved accuracy, but the true power lies in its application to complex scenarios. This development isn’t just about better diffraction analysis; it’s about accelerating materials discovery and providing deeper insights into everything from ancient artifacts to next-generation batteries.