acceptodds
Withdrawn from ICLR 2027

PhaseMatcher: Autoregressive Prediction of Complete Phase Sets from Powder X-Ray Diffraction

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Recovering the complete set of crystalline phases in a mixture from its powder X-ray diffraction (PXRD) pattern is a central task in automated materials characterization. Peak overlap, abundance imbalance, and measurement-induced spectral variation obscure weak-phase evidence and make multiphase identification against large phase libraries difficult. We formulate multiphase PXRD analysis as exact-set prediction with unknown cardinality. PhaseMatcher predicts phases autoregressively from a large reference library and learns when the selected set is complete. After each selection, a physics-guided spectral decomposition module uses the original observation and all selected reference patterns to reconstruct the selected-phase contribution patterns and the residual spectrum. The normalized residual then guides the next-phase prediction, coupling phase identification with spectral decomposition. We construct PhaseMix-135K, a noisy synthetic multiphase PXRD dataset with physically motivated pattern variations, and a controlled mixture dataset from measured RRUFF single-phase patterns. PhaseMatcher achieves state-of-the-art exact-set recovery on the primary noisy benchmark and recovers complete phase sets on the controlled RRUFF dataset.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
powder X-ray diffraction, multiphase identification, autoregressive models, spectral decomposition, AI for science

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_At4iOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This work studies the problem of identifying the complete phase set from PXRD when weaker phases are covered by stronger phases and the number of phases is unknown a-priori. They introduce a method called PhaseMatcher, where they borrow the idea of autoregressive modeling to jointly model the contributions of all selected phases and the original reference such that those contributions will be calibrated at each step. There is another STOP head to determine when we should end the decomposition. They evaluate on two datasets, PhaseMix-135K and RRUFF, where they improve on both compared to the baselines.

Critical strengths
  • The paper's idea is simple but interesting. This paper uses an autoregressive model for phase identification. The model identifies one phase at a time and revisits previously estimated contributions, helping reduce residual distortion and identify weak phases that were previously missed.
  • This method is fast and can be scaled up.
  • There is a very detailed statistical analysis on the results based on different characteristicsof the given data.
Critical weaknesses
  • In the paper, the XRD are constructed without requiring the selected phases to coexist under realistic physical or chemical conditions. This randomly selected phases may have very different diffraction patterns, making them easy to demix. Although the appendix checks the overlap, more cases on chemically plausible mixtures with similar diffraction patterns would provide stronger evidence of the model’s ability to handle difficult cases.

  • Both benchmarks use synthetic mixtures instead of real samples. The Gaussian noise and background are also relatively small. It remains unclear to me whether the model maintains its performance under stronger noise, more complex backgrounds, or real experimental conditions.

  • Real samples can have strong preferred orientation, amorphous content, or phases missing from the library, which the generator does not model. So it is hard to tell whether the method, or its physics module, works on real data.

  • It is unclear to me whether the model performs better because it revisits earlier phase contributions or because it can adjust the reference patterns more flexibly. The joint scalar baseline only adjusts their intensity. A stronger baseline would also adjust peak positions and widths and compare how accurately it identifies the phases.

  • The external baselines perform surprisingly poorly. XDecomposer (Gao et al., 2026) and PhaseDifformer (Ito et al., 2026) were designed for source separation rather than identification, and PhaseDifformer was re-implemented under a modified protocol.

  • Add comparisons with the similar existing method, XRD-AutoAnalyzer (Szymanski et al., 2021), and with the refinement-based methods such as DARA (Fei et al., 2026) and RADAR-PD (Yadav et al., 2026), which are discussed in the related work but not compared.

Szymanski, Nathan J., Christopher J. Bartel, Yan Zeng, Mouhamad Diallo, Haegyeom Kim, and Gerbrand Ceder. "Adaptively driven X-ray diffraction guided by machine learning for autonomous phase identification." npj Computational Materials 9, no. 1 (2023): 31.

Yadav, Lalit, Yongqiang Cheng, and Mathieu Doucet. "Automated multiphase identification and refinement in powder diffraction using mismatch-tolerant machine learning." arXiv preprint arXiv:2605.12478 (2026).

Fei, Yuxing, Matthew J. McDermott, Christopher L. Rom, Shilong Wang, and Gerbrand Ceder. "Dara: Automated Multiple-Hypothesis Phase Identification and Refinement from Powder X‑ray Diffraction." Chemistry of Materials 38, no. 3 (2026): 1364.

Other related comparisons could be informative.

  • A non-autoregressive multi-label classifier using the same encoder together with a count head. Without this comparison, it is unclear whether the gains come from autoregressive decomposition or simply from a strong encoder trained on 12 million synthetically mixing data.

  • Using more deformation synthesis techniques, such as per-phase shift, strain, and peak-width parameters, similar to Le Bail fitting (Le Bail et al., 1988) or CrystalShift (Chang et al., 2025).

Le Bail, A., Duroy, H., & Fourquet, J. L. (1988). Ab-initio structure determination of LiSbWO6 by X-ray powder diffraction. Materials Research Bulletin, 23(3), 447–452.

Chang, Ming-Chiang, Sebastian Ament, Maximilian Amsler, Duncan R. Sutherland, Lan Zhou, John M. Gregoire, Carla P. Gomes, R. Bruce van Dover, and Michael O. Thompson. "Probabilistic phase labeling and lattice refinement for autonomous materials research." npj Computational Materials 11, no. 1 (2025): 148.

Questions for authors
  • Can you test the method on realistic experimental multi-phase dataset? In this paper, the mixtures are synthetic, how does the model perform with much stronger noise? this can be tested on synthetic mixtures even without real experimental data.
  • How does the model handle a phase that is not in the library, or intensity changes larger than the ±10% used in training? Does it stop, or pick a wrong phase?
  • How are phases selected for mixing? Are they chosen randomly, or are physical and chemical constraints considered? How does the model perform when the selected phases have similar diffraction patterns and strongly overlapping peaks?
  • Can you compare with XRD-AutoAnalyzer, or with a subtraction baseline that also fits peak shift and width?
2 comments
AuthorsOpenReview ↗

Response to reviewer At4i

Thank you for your detailed comments. We added the following experiments and explanations.

  1. Phase selection and perturbations. We agree that the original benchmark randomly selects phases without considering chemical coexistence. PhaseMix-135K does not explicitly include preferred orientation, and its Gaussian noise and background are relatively weak. We therefore added PhaseMix-2 with richer perturbations, including preferred orientation, lattice deformation, instrument and sample broadening, stronger backgrounds, and Poisson noise. See Appendix D.10, Tables 32–34.

  2. Peak-position and peak-width fitting baseline. Following XRD-AutoAnalyzer’s iterative fitting-and-subtraction approach, we added and retrained a Scalar baseline that fits peak positions, widths, and intensities. Its Exact-set@1 is 66.97%, compared with 77.15% for PhaseMatcher Greedy. See Tables 1 and 3 and Appendix B.5.

  3. Transfer to measured spectra. Our supplementary file reports results on 91 measured spectra: 70 from DARA and 21 from OPXRD’s HKUST-A subset. These results provide evidence of transfer to experimental data. We use element constraints to narrow the candidates for these evaluations. The HKUST-A comparison is also in Appendix D.10, Table 35.

  4. XDecomposer results. Our measured phase recall is close to the component Top-1 reported in the original paper, but complete-set accuracy is much lower. The numbers are below. See Table 2 in the original paper and Appendix C.3, Table 15 in ours.

Phase countPaper component Top-1Our component Top-1Our phase recallOur Exact-set@1
265.31%78.41%66.79%34.32%
353.49%61.17%53.66%1.26%
441.90%51.96%46.54%0.03%

The two studies use different evaluation protocols. Low Exact-set accuracy does not mean low component-level accuracy, because complete-set identification requires every phase and the phase count to be correct.

  1. Missing phases and larger intensity changes. If a true phase is missing from the library, the model looks for a similar pattern within the library. If relative-intensity changes exceed the training range of ±10%, the model still attempts a match. It may identify the correct phase or select another phase. Accuracy outside this range has not yet been quantified.

Thank you for your responses and clarifications. Overall, I do not consider the technical contribution to be sufficiently significant. I maintain my score.

Reviewer_2u2mOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary
  • Summary
    • Phase sets are... (never defined, but from Figure 1 they appear to be a collection of unique crystal structures with different compositions that might appear within a single crystalline sample) I had to look this up though.
    • This work introduces an autoregressive framework for spectral decomposition and phase-set identification using Transformers
Critical strengths
  • Strengths
    • The question the authors are exploring is interesting
Critical weaknesses
  • Weaknesses
    • No code, reproducibility questionable
    • Would have been good to say somewhere in the main text that you are working with Transformers for this (I found it buried in the appendix)
    • Not clear to me that there isn't data leakage (how do you det aside your hold-out test, and make sure the comparisons for the various different models you compare e.g. in Table 1 have all seen the same data and none of the hold-out data?)
    • The problem definition was not fully clear to me, so apologies if I've misunderstood the work and its importance, but I would have liked some more context on "when" someone would use a model like PhaseMatcher and what other tools they might be using as well, would it save them time, etc. What would it take for such a tool to be actually useful to an experimental scientist (e.g., what kind of performance metrics would you expect, what thresholds) given that the work is an application of ML to a scientific domain. This context is missing for me so it is hard to judge the impact or how someone would actually use the framework.
Questions for authors
  • Questions
    • Not clear to me why so different compositions+structures would be present in a single sample (e.g. Figure 1b example). Wouldn't you know something about the elemental composition of your sample before doing the decoding, and in this case be able to constrain your model during the spectral decomposition?
    • Why is an autoregressive approach the right approach here?
    • Are you using simulated or experimental data for training, or a mixture of both? if only the former, how does the model generalize to experimental spectral decomposition?
    • How does the cost of PhaseMatcher with Beam search compare to the other methods explored, e.g., in Table 3? You are reporting higher performance but usually this comes at greater compute cost, so what does it look like for you?
    • What are the failure modes of PhaseMatcher? Any particular PXRD patterns it could not resolve? for instance as they get noiser and noisers? where does your model break down, in other words?
2 comments
AuthorsOpenReview ↗

Response to reviewer 2u2m

Thank you for your comments and questions. We respond below.

  1. Phase sets and element information. We added the definition of “phase set” in the second paragraph of the revised introduction. A sample can contain starting materials, products, by-products, and impurities, so phases with different compositions and structures can coexist. Our main benchmarks search the full library without element information. For the measured spectra in the supplementary file, we use element information to narrow the candidates.

  2. Why autoregression. Our supplementary file describes the traditional Search–Match workflow: match a phase, estimate its contribution, and continue matching the remaining signal. This is similar to an autoregressive process. PhaseMatcher jointly re-estimates all selected contributions at each step and uses STOP to decide when to finish.

  3. Training data and transfer to measured spectra. PhaseMix-135K uses simulated data, while RRUFF mixtures are constructed from measured single-phase patterns. Training includes variations in peak positions, widths, intensities, background, and noise. Our supplementary file reports results on 91 measured spectra: 70 from DARA and 21 from OPXRD’s HKUST-A subset. The HKUST-A comparison is also in Appendix D.10, Table 35.

  4. Data and reproducibility. We checked that test observations were not used for training and that complete two- to four-phase combinations do not overlap between training and test sets. Implementation details are in Appendices A–C. We plan to release the code after acceptance.

  5. Computational cost. We added a runtime comparison in Appendix D.8. PhaseMatcher’s Greedy inference averages 0.0544 seconds per spectrum on one RTX 4090. The table also explains the differences in hardware and timing scope across methods.

  6. Failure modes. The main difficulty is distinguishing phases with similar patterns when the available observations and constraints are insufficient. As noise increases, weak peaks that help distinguish candidates become harder to detect, making it more likely that several phase combinations can explain the observation.

Thank you for the response. I will maintain my score at 2.

Reviewer_2FE8OpenReview ↗
Rating
2
Confidence
1
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes PhaseMatcher, an autoregressive framework for identifying unknown-size sets of crystalline phases from powder X-ray diffraction (PXRD) patterns. Instead of subtracting fixed reference patterns sequentially—which compounds errors when peaks overlap—PhaseMatcher uses a reference-conditioned spectral decomposition network to jointly re-estimate all selected phase contributions and the residual after each step. A separate STOP head predicts when the phase set is complete. Evaluated on PhaseMix-135K and RRUFF-based synthetic mixtures, PhaseMatcher substantially outperforms iterative matching, source-decomposition baselines (XDecomposer, PhaseDifformer), and a scalar-subtraction ablation.

Critical strengths

(Note: My background is in machine learning rather than PXRD characterization, so my evaluation focuses on the ML formulation and empirical results.)

  1. Effective Joint Re-estimation for Autoregressive Set Prediction: Jointly re-estimating all selected phase contributions from the original observation via pointwise softmax allocation is a clean, well-motivated way to prevent error accumulation in sequential residual matching. Ablations against sequential scalar subtraction (+12.4% Exact-set@1 on PhaseMix-135K and +23.2% on RRUFF) clearly show the value of this design, especially for weak overlapping phases.

  2. Clean Architecture and Thorough Evaluation: Separating the inputs to the Phase Head (normalized residual) and STOP Head (unnormalized observation) makes strong intuitive sense under noise. The empirical evaluation and ablations across 100K test mixtures are comprehensive and well-presented.

Critical weaknesses
  1. Evaluation Restricted to Synthetic Mixtures: Even on the RRUFF benchmark, the multiphase mixtures and perturbations are generated synthetically using the same parametric model (up to 4 phases, isotropic strain, Bernstein polynomial background) that informs the model's training and physics module. It is unclear how well the model generalizes to real experimentally measured multiphase mixtures.

  2. Small End-to-End Gain from Explicit Physics Guidance: In Table 27, removing the physics-guidance module drops Exact-set@1 by only 0.36% (Greedy) and 0.62% (Beam), suggesting that the neural encoder-decoder and softmax allocation already do almost all of the heavy lifting for phase identification.

Questions for authors
  1. Robustness to Early False Positives: During training, the decomposition network only sees prefixes of ground-truth phases. At inference, how does the joint decomposition behave if an incorrect phase is selected early on?

  2. Real Experimental Mixtures: Are there public benchmarks of physically measured multiphase PXRD mixtures where PhaseMatcher can be evaluated zero-shot, or with >4 phases?

1 comment
AuthorsOpenReview ↗

Response to reviewer 2FE8

Thank you for these thoughtful comments. We respond to the questions below.

  1. Early false positives. We train the decomposition model using prefixes of the ground-truth phase sequence. At inference, the predicted history may contain an incorrect phase. Our supplementary file analyzes how joint decomposition behaves after an incorrect phase is selected (Appendix D.3).

  2. Measured mixtures and phase counts. There is currently no large-scale public benchmark of measured, phase-labeled PXRD mixtures with high phase counts. We added results on 70 measured spectra from the DARA supplement and 21 labeled patterns from the HKUST-A subset of OPXRD, with details provided in the supplementary material.. The HKUST-A comparison is also reported in Appendix D.10, Table 35. Our positional encoding and dataset are limited to 4 phases, but the method can in principle be extended to more than four phases.

  3. Physics guidance. We agree that its end-to-end identification gains are small: 0.36 percentage points for Greedy and 0.62 for Beam. Its main purpose is to improve spectral fitting and provide a better starting point for downstream refinement, rather than to drive identification accuracy on its own.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.