XDecomposer: Learning Prior-Free Set Decomposition for Multiphase X-ray Diffraction
Hanyu Gao, Bin Cao, Yunyue Su, Tong-Yi Zhang, Qiang Liu · 2026
Read XDecomposer: Learning Prior-Free Set Decomposition for Multiphase X-ray Diffraction by Hanyu Gao, Bin Cao, Yunyue Su, Tong-Yi Zhang, Qiang Liu (2026) 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-2605-05866.txt.
Introduction
A powder diffraction pattern from a mixed material stacks the marks of every phase inside, and untangling them usually needs a candidate list. Hanyu Gao and four co-authors present XDecomposer, which needs no such list. It jointly splits a mixed pattern and names the phases without templates or prior knowledge of how many phases exist.
The method treats the task as a set prediction problem. Earlier retrieval and generation methods had assumed single-phase inputs and broke on mixtures. Prior-free splitting opens routine identification for the messy mixtures that real synthesis actually produces.
Catalog record
| CREATOR | Hanyu Gao, Bin Cao, Yunyue Su, Tong-Yi Zhang, Qiang Liu |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 75 KB |
| LICENSE | http://creativecommons.org/licenses/by-sa/4.0/ |
| RIGHTS | CC BY SA |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2605.05866
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
XDecomposer: Learning Prior-Free Set Decomposition for Multiphase X-ray Diffraction Hanyu Gao, Bin Cao, Yunyue Su, Tong-Yi Zhang, Qiang Liu Abstract Multiphase powder X-ray diffraction (PXRD) analysis remains a fundamental bottleneck in structure identification, as real-world synthesis often produces complex mixtures whose constituent phases (components) cannot be reliably disentangled. While recent advances in representation-based crystal retrieval and generation suggest the possibility of inferring structures directly from PXRD, existing approaches largely assume single-phase inputs and break down in multiphase settings. Here, we present XDecomposer, a prior-free framework for joint decomposition and identification of multiphase XRD patterns without requiring candidate phase lists, structural templates, or prior knowledge of phase number. We formulate multiphase diffraction analysis as a set prediction problem, where the model infers an unordered set of phase-resolved components, their mixture proportions, and corresponding structural representations within a unified architecture. A phase-query-driven decomposition mechanism, together with diffraction-consistent physical reconstruction, enables accurate source separation while preserving crystallographic fidelity. Extensive experiments on both simulated and experimental datasets show that XDecomposer substantially improves reconstruction accuracy and phase identification across diverse chemical systems, while maintaining strong generalization to unseen mixtures. These results provide a practical route toward data-driven, source-resolved multiphase XRD analysis and reduce long-standing dependence on prior-guided iteratively phase matching. The code is openly available at https://github.com/Licht0812/XDecomposer 1 1 footnotetext: Equal contribution. 2 2 footnotetext: Corresponding authors: Bin Cao ( [email protected] ) and Qiang Liu ( [email protected] ). 1 Introduction Powder X-ray diffraction (PXRD) is a foundational technique for determining crystal structure [ 11 , 4 , 54 ] . In a typical PXRD workflow, phase identification provides the basis for downstream structural refinement and quantitative analysis. Widely used refinement methods, such as Rietveld [ 40 ] , Pawley [ 37 ] , and Le Bail [ 24 ] , generally assume that the constituent phases are known a priori; when this assumption is violated, subsequent analyses can become unreliable. Phase identification is relatively well established for single-phase samples [ 31 , 8 ] . However, in practice, samples often contain multiple crystalline phases whose diffraction signals are superimposed in a single one-dimensional pattern. In such multiphase settings, the goal is no longer to identify a single phase, but to simultaneously determine all constituent phases and estimate their relative abundances. Unknown compositions, severe peak overlap, and experimental distortions further complicate this task [ 6 , 32 , 47 , 42 ] . As a result, multiphase phase identification has become a central bottleneck in PXRD analysis. From a machine learning perspective, single-phase identification can be cast as mapping a one-dimensional PXRD pattern to a single phase or structural representation. This setting has been widely studied using classification [ 8 , 6 , 32 , 41 , 26 , 25 , 48 ] , retrieval [ 10 , 23 , 53 , 58 ] , and generative modeling [ 28 , 56 , 13 , 15 ] . By contrast, multiphase analysis is not simply a multi-label extension of single-phase identification. A mixed pattern is generated by multiple latent diffraction components whose identities, number, and proportions are generally unknown. This makes the task more naturally formulated as a single-channel blind source separation (BSS) problem: given a mixed PXRD pattern, a model must jointly identify the constituent phases, reconstruct the diffraction signal of each phase, and estimate their relative proportions. This formulation highlights the underdetermined nature of the problem, in which multiple unobserved components must be recovered from a single observed signal, often without component-level supervision (see Appendix A.3 ). A natural approach is to exploit the additive structure of diffraction signals. Existing methods therefore adopt a sequential identification paradigm [ 47 , 18 ] : they iteratively identify a dominant phase, subtract its contribution, and repeat on the residual. However, this strategy suffers from a fundamental limitation: phase proportions are unknown, making each subtraction inherently inaccurate. Moreover, peak overlap causes subtraction to distort signals from other phases, leading to error accumulation across iterations. As a result, even strong single-phase models often fail in multiphase settings, revealing the inadequacy of sequential identification and motivating the need for a new paradigm. Here, we formulate prior-free multiphase XRD analysis as a blind source separation problem, in which a model directly decomposes a mixed diffraction pattern into an unordered set of single-phase components and their proportions. Existing approaches either rely on predefined phase sets or sequential subtraction, which limits their ability to jointly recover all components. To this end, we propose XDecomposer for direct whole-pattern decomposition. It is trained in two stages: self-supervised pretraining on large-scale single-phase data to learn general diffraction patterns, followed by supervised training on multiphase mixtures for component-level decomposition. XDecomposer integrates convolutional encoding for local peak features, Transformer-based global context modeling, a query-based mechanism to separate individual phases, and physically constrained reconstruction. Experiments on simulated and real-word experimental datasets show that XDecomposer accurately decomposes mixtures, reconstructs phase-specific signals, and estimates phase proportions, establishing a new perspective for multiphase PXRD analysis. The contributions are as follows: • Problem formulation. We cast multiphase PXRD analysis as a permutation-invariant set prediction problem under unknown phase cardinality, providing an alternative to sequential identification and enabling direct, prior-free decomposition of mixture patterns. • Modeling approach. We introduce XDecomposer, a query-based decomposition model that combines global context modeling with phase-conditioned latent modulation to separate phase-specific components from a single mixed diffraction pattern. • Physics-aware reconstruction. We incorporate physically motivated constraints, including non-negativity and mixture consistency, through a mask-based reconstruction mechanism, improving stability and plausibility of the recovered components. • Empirical evaluation. We conduct extensive experiments on both simulated and real-world datasets, showing that XDecomposer consistently outperforms strong baselines across multiple metrics and exhibits robust generalization to unseen mixtures. 2 Related Work Query strategy For single-phase materials, identification typically matches measured diffraction patterns to reference entries, such as the Powder Diffraction File, using peak positions and relative intensities [ 18 ] (see Appendix C.1 ). For multiphase systems, the observed pattern is treated as a superposition of phase signals and analyzed sequentially by identifying a dominant phase, subtracting its contribution, and repeating on the residual. Such methods therefore cannot jointly infer all mixture components and, despite their effectiveness in controlled settings, depend strongly on candidate-phase priors and are sensitive to peak overlap, noise, and incomplete or biased reference databases. Learning Strategies Recent advances leverage machine learning to automate phase identification from PXRD patterns. For single-phase materials, the problem is well-studied and typically formulated as classification, retrieval, or generation. Classification methods treat phase identification as a supervised labeling task, directly predicting the phase identity from diffraction patterns [ 8 , 6 , 32 , 41 , 26 , 25 , 48 ] . Retrieval-based approaches instead learn representations that enable matching a query pattern against a database of known phases [ 10 , 23 , 53 , 58 ] . Generation methods model the inverse diffraction process, reconstructing crystal structures from PXRD patterns to facilitate phase identification [ 28 , 56 , 13 , 15 ] . In contrast, multiphase analysis remains underexplored. Existing methods typically extend these formulations to multi-label classification [ 26 , 25 , 12 , 14 ] or regression [ 1 , 2 ] , often with fixed candidate phase