deCIFer: Crystal Structure Prediction from Powder Diffraction Data using Autoregressive Language Models
Frederik Lizak Johansen [email protected], Ulrik Friis-Jensen [email protected], Erik Bjørnager Dam [email protected], Kirsten Marie Ørnsbjerg Jensen [email protected], Rocío Mercado [email protected], Raghavendra Selvan [email protected] · 2025
Read deCIFer: Crystal Structure Prediction from Powder Diffraction Data using Autoregressive Language Models by Frederik Lizak Johansen [email protected], Ulrik Friis-Jensen [email protected], Erik Bjørnager Dam [email protected], Kirsten Marie Ørnsbjerg Jensen [email protected], Rocío Mercado [email protected], Raghavendra Selvan [email protected] (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-2502-02189.txt.
Introduction
Frederik Lizak Johansen and co-authors build deCIFer, a language model that predicts crystal structures from powder X-ray patterns, in a 2025 paper. Naming a new material's crystal layout is a key step in discovery, but powder data alone leave the answer underdetermined. Older methods lean on composition or symmetry hints rather than the measured pattern itself.
Trained on nearly two point three million known structures, the model reads a powder pattern and writes out a full structure file. Noise and blur were added to training patterns to mimic real instruments. On hard test sets it matched the true structure in most cases, setting a baseline the authors hope others will extend to messier real data.
Catalog record
| CREATOR | Frederik Lizak Johansen [email protected], Ulrik Friis-Jensen [email protected], Erik Bjørnager Dam [email protected], Kirsten Marie Ørnsbjerg Jensen [email protected], Rocío Mercado [email protected], Raghavendra Selvan [email protected] |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 73 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2502-02189
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
deCIFer: Crystal Structure Prediction from Powder Diffraction Data using Autoregressive Language Models Frederik Lizak Johansen [email protected], Ulrik Friis-Jensen [email protected], Erik Bjørnager Dam [email protected], Kirsten Marie Ørnsbjerg Jensen [email protected], Rocío Mercado [email protected], Raghavendra Selvan [email protected] Abstract Novel materials drive advancements in fields ranging from energy storage to electronics, with crystal structure characterization forming a crucial yet challenging step in materials discovery. In this work, we introduce deCIFer , an autoregressive language model designed for powder X-ray diffraction (PXRD)-conditioned crystal structure prediction (PXRD-CSP). Unlike traditional CSP methods that rely primarily on composition or symmetry constraints, deCIFer explicitly incorporates PXRD data, directly generating crystal structures in the widely adopted Crystallographic Information File (CIF) format. The model is trained on nearly 2.3 million crystal structures, with PXRD conditioning augmented by basic forms of synthetic experimental artifacts, specifically Gaussian noise and instrumental peak broadening, to reflect fundamental real-world conditions. Validated across diverse synthetic datasets representative of challenging inorganic materials, deCIFer achieves a 94% structural match rate. The evaluation is based on metrics such as the residual weighted profile ( R w p R_{wp} ) and structural match rate (MR), chosen explicitly for their practical relevance in this inherently underdetermined problem. deCIFer establishes a robust baseline for future expansion toward more complex experimental scenarios, bridging the gap between computational predictions and experimental crystal structure determination. 1 Introduction Characterizing the atomic structure of functional materials is essential for enabling progress in energy storage, electronics, and other emerging technologies. Powder X-ray diffraction (PXRD) is a widely employed experimental technique for this purpose: it measures a one-dimensional intensity profile whose peaks reflect the periodic arrangement of atoms in a crystalline solid. This makes structure determination an inverse problem: a three-dimensional periodic structure is mapped to a one-dimensional signal by a (largely) known forward process, and the inverse is often underdetermined, so multiple structures can match similar PXRD profiles. In experimental crystallography, structure solution from PXRD is typically carried out through a multi-stage workflow: preprocessing and peak finding, proposing an initial structural hypothesis, and then Rietveld refinement , which is a nonlinear least-squares fit of a simulated PXRD profile to the measured profile. The initial hypothesis is often obtained by identifying representative candidate structures from crystallographic databases, and using these as starting points for refinement. A key bottleneck is obtaining a good initial candidate, especially when peaks overlap, the pattern is noisy, or the correct structure type is not well represented in available databases. Crystal structure prediction (CSP) refers to the computational task of inferring these same structures given a set of constraints or observations. In traditional settings, these constraints are limited to high-level descriptors such as chemical composition or symmetry. While recent work has explored integrating experimental data into generative models, direct PXRD-conditioned CSP remains underexplored. Public structure databases predominantly archive refined structures, but do not routinely provide standardized, model-ready paired PXRD patterns, since obtained data scans depend on instrument settings, sample preparation, and preprocessing choices. As a result, large aligned (structure, PXRD) corpora are uncommon, which has limited end-to-end PXRD conditioning in generative models. In this work, we present deCIFer , an autoregressive transformer-based model designed explicitly for PXRD-conditioned crystal structure prediction (PXRD-CSP) (Figure 1 ). To study PXRD conditioning despite the limited availability of paired experimental data, we construct a controlled paired setting: for each reference crystal structure we generate its corresponding PXRD profile using a defined forward simulator, and we use these paired examples for large-scale training and reproducible evaluation. This provides a controlled setting for large-scale training and evaluation, enabling our study of PXRD-conditioned generation. deCIFer can then be extended as richer experimental structure-PXRD datasets become available. Given a target PXRD profile, deCIFer generates a complete crystallographic structure description . We represent this description using the Crystallographic Information File (CIF) format, a standard structured text-representation used to exchange crystal structures and to interface with common crystallography tools, including refinement pipelines and structure databases. At the same time, this motivates a decoder-only autoregressive design: CIFs are variable-length structured text with strict syntax, and autoregressive decoding both supports fixing known CIF descriptors (such as chemical composition) during generation and enables sampling multiple PXRD-consistent hypotheses that can be ranked by PXRD-consistency. This enables controlled comparisons (PXRD-only vs. PXRD+descriptors). deCIFer is intended to complement traditional workflows as a hypothesis generator, proposing PXRD-consistent candidate structures which can then be filtered, ranked by PXRD-consistency, and refined using standard tools. To establish a practical baseline, we incorporate basic forms of synthetic experimental artifacts into our training and validation datasets, specifically Gaussian noise and instrumental peak broadening. These basic synthetic artefacts represent fundamental yet simplified aspects of real-world PXRD variability, chosen intentionally to provide a controlled, reproducible starting point for evaluation and for future extensions of the method. They act as a minimal simulation of measurement noise and resolution limits, analogous to data augmentation strategies in vision or speech domains. They do not capture more complex effects such as peak asymmetry, background drift, or preferred orientation, which we leave for future studies. The performance of deCIFer is demonstrated on diverse PXRD patterns from large-scale datasets representative of challenging inorganic materials. Our evaluations confirm the robustness of the approach and its ability to produce syntactically correct and structurally meaningful CIFs that accurately reproduce target diffraction patterns. With this, deCIFer serves as a foundational step toward bridging the gap between computational CSP and experimental crystal structure determination workflows. Our key contributions in this work are: 1. Integration of PXRD-based experimental conditioning into an autoregressive transformer model, enabling direct CSP in CIF format; a capability not demonstrated previously in transformer-based generative models. 2. Implementation of an effective conditioning mechanism to handle variable-length CIF sequences in autoregressive modelling. 3. Simulation of fundamental PXRD experimental artefact, Gaussian noise and peak broadening, as a practical baseline for real-world scenarios. 4. Comprehensive evaluation on two large-scale datasets: NOMA 1 1 1 NOMA stands for N OMAD ( nomad2019 ) , O QMD ( kirklin2015open ) , & M P ( materialsproject2013 ) A ggregation. and CHILI-100K ( FriisJensenJohansen2024 ) , including comparison to state-of-the-art CSP models and analysis of sampling consistency under varying conditions. These contributions aim to establish a reproducible foundation for integrating computational CSP with experimental workflows. (a) Overview of the deCIFer model. (b) Example generations using deCIFer. Figure 1 : (a) Overview of the deCIFer model, which performs autoregressive crystal structure prediction (CSP) from PXRD data, optionally guided by standard crystallographic descriptors (e.g., composition and space group) encoded as CIF fields, which can be provided to constrain generation or omitted for exploratory sampling. PXRD embeddings are prepended to the CIF token sequence, enabling the generation of structurally consistent CIFs directly from diffraction data. (b) Three examples from the NOMA test set showing deCIFer generations, each illustrating a reference structure, the generated structure and their corresponding PXRD profiles. 2 Background and Related Work CSP has traditionally relied on high-level descriptors, such as chemical composition or symmetry constraints, to guide predictions. Recently, approaches termed data-informed CSP have begun integrating explicit experimental data (particularly diffraction data) into their generative processes ( kjaer2023deepstruc ; guo2024abinitiostructuresolutions ; riesel2024crystal ; lai2025endtoendcrystalstructureprediction ) , marking a significant shift away from purely descriptor-driven CSP. Among the many experimental modalities that could inform such models,