PaperlayerView exact revision on arXiv

Source-linked AI summary

Knowledge-Driven XRD Phase Identification via Multi-View Retrieval and Explanation

Doaa Mohamed, Markus Stricker

arXiv:2609.31888v1cond-mat.mtrl-scics.LG

TL;DR

XRD phase analysis is challenging when experimental patterns diverge from references and classifiers struggle with ambiguous or unseen compounds. HyPhID combines classification, latent retrieval, explanation-guided similarity, and compositional evidence; it demonstrates strong performance on experimentally measured XRD for crystal-system and space-group prediction.

  • Problem

    XRD analysis remains challenging because experimental patterns can differ from ideal references, while single-label classifiers have limited interpretability and struggle with ambiguous or unseen compounds.

    Source: §1 · pp. 1–2

  • Method

    HyPhID combines a learned XRD representation with classification, latent-space retrieval, Integrated Gradients explanations, and composition-based evidence to aggregate ranked predictions.

    Source: §1 · p. 2

  • Results

    HyPhID demonstrates strong performance for crystal-system and space-group prediction on experimentally measured XRD while providing interpretable decision mechanisms.

    Source: §8 · pp. 9–10

  • Takeaways & Limitations

    Predictions can be supported by explicit diffraction evidence aligned with expert reasoning, alongside chemically grounded compositional priors.

    Source: §8 · pp. 9–10

    Performance drops on external datasets, where instrumental and sample variability pose a generalization challenge.

    Source: §7.3 · p. 8

Abstract

from arXiv · show

X-ray diffraction (XRD) is a experimental technique for determining the phase composition and structure of crystalline materials. However, interpreting XRD patterns is challenging, particularly in high-throughput materials discovery, where many novel materials may need to be characterized and no reference patterns are available. Consequently, machine learning is increasingly used to accelerate and automate the analysis while reducing errors associated with human interpretation. We propose a multi-decision framework for XRD phase analysis that integrates representation learning, similarity-based retrieval, and explainable decision support within a unified reference database. A convolutional autoencoder learns compact latent representations of XRD patterns that preserve structural similarity while remaining robust to variations arising from experimental noise and measurement conditions. By integrating multiple decision pathways within a shared latent space, the framework moves beyond single-label prediction toward ranked and interpretable phase analysis that mirrors expert practice. During inference, complementary decision mechanisms are applied, including latent-space classification and retrieval, explanation-guided similarity using Integrated Gradients, and composition-based similarity search. These mechanisms generate ranked candidate phase lists that are aggregated into a final prediction with an associated confidence score. Experiments on synthetic datasets demonstrate strong predictive performance, achieving 98.85\,\% accuracy for crystal system classification and 95.82\,\% accuracy for space group prediction on the test set, while maintaining robustness under realistic perturbations. The framework supports reliable, analyst-friendly identification of crystal phases and structures in high-throughput and exploratory materials discovery settings.

1 Introduction

XRD analysis is difficult when experimental patterns depart from ideal references or compounds are structurally ambiguous, while conventional classifiers often provide only single-label predictions. HyPhID combines latent representations, retrieval, explanation, and composition evidence to produce ranked crystallographic predictions.

1 Introduction

  • Deep-learning XRD approaches often remain single-label classifiers with limited interpretability and difficulty generalizing across synthetic and experimental domains, especially for ambiguous or unseen compounds.Experimental patterns can also deviate from references because of defects, disorder, impurities, or absent references for novel materials.Source: §1 · pp. 1–2
  • Latent embeddings support structural similarity retrieval, while combining parametric classification with non-parametric reference comparison offers complementary decision strategies.The introduction also motivates explanation-guided reasoning to connect model decisions with expert peak interpretation.Source: §1 · p. 2
  • Composition can constrain plausible structures when diffraction alone cannot uniquely identify a phase, and HyPhID fuses it with other evidence at the decision level.The framework uses compositional evidence as a complement to diffraction-based inference.Source: §1 · p. 2
  • HyPhID combines a convolutional-autoencoder representation with classifier output, latent retrieval, Integrated Gradients similarity, and composition retrieval to rank crystallographic predictions.The shared framework supports phase identification and prediction of crystal system and space group.Source: §1 · p. 2, Figure 1 · p. 3

2 Synthetic XRD Dataset Generation

The study builds balanced synthetic XRD datasets from CIF structures drawn from the Materials Project and Crystallography Open Database, then augments simulated patterns to represent experimental variability.

2 Synthetic XRD Dataset Generation

  • 68,958 crystal structures—58,418 from Materials Project and 10,540 from COD—provide the CIF basis for the synthetic dataset.The source collection combines experimentally determined COD structures with mainly DFT-relaxed Materials Project structures.Source: §2 · p. 3
  • The space-group dataset samples about 2,000 structures for each of 22 representative groups, while the crystal-system dataset samples about 7,000 for each of seven systems.The 22 groups are selected from the 230 crystallographic space groups to maintain balance and sufficient training data.Source: §2 · p. 3
  • Each CIF generates 15 additional diffraction patterns using peak broadening, preferred orientation, background contributions, intensity scaling, and small peak shifts.These perturbations model instrumental and sample-related variability while preserving the underlying phase label.Source: §2 · p. 3

3 Model Architecture and Training

The model jointly learns reconstruction and classification representations, then trains them with objectives for peak-aware reconstruction, classification, smoothness, and supervised latent-space structure.

3 Model Architecture and Training

  • A physics-informed convolutional autoencoder with an auxiliary classifier jointly supports XRD classification, retrieval, and structured representation learning under experimental variability.Its training incorporates peak-aware reconstruction weighting and smoothness regularization.Source: §3 · p. 3

3.1 Architecture

  • The encoder maps diffraction patterns into separate reconstruction and classification latent vectors, with a decoder reconstructing the signal and a classifier predicting phase labels.The reconstruction latent retains signal, background, and noise information, while the classification latent captures discriminative structural features.Source: §3 · p. 3, §3.1 · p. 4

3.2 Training Objectives

  • Training balances reconstruction, classification, smoothness, and supervised contrastive objectives, with weights selected empirically using validation experiments.The peak-aware reconstruction loss emphasizes diffraction peaks, whose regions are identified by an intensity threshold relative to the maximum peak height; gradient sparsity encourages smooth reconstructions.Source: §3.2 · p. 4
  • Supervised contrastive learning clusters samples with the same phase label in latent space, supporting similarity retrieval and robustness to structural ambiguity.The objective is applied to projected embeddings.Source: §3.2 · p. 4

4 Integrated Gradients Database for Explainable Case-Based Prediction

The section uses Integrated Gradients attributions as a second retrieval space, supporting case-based predictions through similarity between attribution patterns. Stored training explanations are compared with query explanations, with retrieval invoked when additional interpretability is needed.

4 Integrated Gradients Database for Explainable Case-Based Prediction

  • Training attributions use the ground-truth class as the target, are stored with labels and dataset indices, then mean-centered and L2-normalized for similarity search.The Integrated Gradients calculation uses a zero-intensity baseline and numerically approximates the integral.Source: §4 · p. 5
  • Query explanations are computed when additional interpretability is needed, such as under high uncertainty or model-retrieval disagreement.Attribution similarity is measured with cosine similarity.Source: §4 · p. 5
  • Top-k attribution neighbors contribute class evidence through similarity-weighted voting, and class probabilities are obtained with temperature-scaled softmax.The resulting prediction is explanation-driven and supports case-based reasoning in attribution space.Source: §4 · p. 5
  • Similarity between query and stored Integrated Gradients attribution patterns supports explanation-driven retrieval and case-based prediction.This provides an attribution-space alternative to comparing diffraction patterns alone.Source: §4 · pp. 4–5

5 Adaptive Cosine Retrieval for Latent-Space Classification

The method performs non-parametric classification using a frozen autoencoder encoder and a database of training-sample latents. It retrieves cosine-similar neighbors and selects the neighborhood size yielding the most decisive class prediction.

5 Adaptive Cosine Retrieval for Latent-Space Classification

  • A frozen encoder embeds training samples once into a latent reference database, and cosine-similarity retrieval classifies each query against it.Query and stored latents are L2-normalized.Source: §5 · p. 5
  • The method evaluates multiple neighborhood sizes and selects the one with the highest top-1 class probability.Class evidence sums similarity contributions from neighbors of the same class; negative similarities may optionally be ignored.Source: §5 · p. 5
  • Similarity-weighted class scores become probabilities through temperature-scaled softmax, and the selected label is returned with top-n candidate classes.Lower temperature values produce sharper class probabilities.Source: §5 · p. 5

6 Embedding-Based Chemical Retrieval

The section compares learned chemical-composition embeddings with Magpie descriptors for retrieving chemically similar materials. It evaluates retrieval by shared-element accuracy and agreement between the two learned encoders.

6 Embedding-Based Chemical Retrieval

  • Chemical composition retrieval compares element-level and composition-level Qwen2.5-1.5B-Instruct embeddings with a Magpie descriptor baseline.Element-level vectors weight element embeddings by stoichiometric fractions; composition-level vectors mean-pool the final transformer layer.Source: §6 · p. 6
  • Normalized representations use cosine similarity to retrieve the top-k training compositions for each query.The learned embeddings are 2048-dimensional, while Magpie features are 145-dimensional.Source: §6 · p. 6
  • Chemical accuracy@k measures the fraction of retrieved compositions sharing at least half their constituent elements with the query, while retrieval overlap@k measures agreement between learned encoders.Both metrics are applied to the learned embeddings and Magpie baseline.Source: §6 · p. 6

7 Results and Discussion

The results examine Integrated Gradients (IG) decision patterns, compositional retrieval, and classification performance. They report structurally meaningful attribution patterns, stronger element-level chemical retrieval, and high internal but lower external classification accuracy.

7.1 Integrated Gradients Analysis of Model Decisions

  • IG attribution fingerprints cluster crystallographic classes by structural similarity, with higher-symmetry systems grouping around dominant peaks and lower-symmetry systems forming broader clusters.The authors interpret these patterns as evidence that predictions rely on physically meaningful diffraction features; Figure 2 shows hierarchical clustering from cosine distances between normalized fingerprints.Source: §7.1 · p. 6, Figure 2 · p. 7

7.2 Results of Embedding-Based Chemical Retrieval

  • Element-level Qwen embeddings outperform composition-level embeddings and Magpie descriptors in chemical retrieval accuracy and stability.The low overlap and weak correlation between element- and composition-level embeddings suggest they capture different notions of chemical similarity; Table 1 compares Qwen embeddings and Table 2 compares Qwen with Magpie.Source: §7.2 · pp. 6–7, Table 1 · p. 7, Table 2 · p. 8
  • Element-level Qwen embeddings are adopted for composition retrieval, which returns top-k chemically similar references and uses their crystallographic metadata as structural evidence.Figure 3 visualizes element-level and composition-level Qwen embeddings alongside aligned Magpie representations; element-level embeddings form compact clusters, while composition-level embeddings are more diffuse.Source: §7.2 · p. 7, Figure 3 · p. 8

7.3 Classification Performance

  • 98.85% crystal system accuracy and 95.82% space group accuracy on the internal holdout fell to 81.07% and 56.13%, respectively, on experimental RRUFF data.The external dataset contains experimentally reported, non-relaxed structures, unlike the mixed experimental and DFT-relaxed training data; the authors attribute the gap to domain shift.Source: §7.3 · pp. 7–8
  • HyPhID combines classifier, latent retrieval, IG similarity, and composition retrieval scores to produce ranked plausible crystallographic classes.The IG module mainly supports interpretability, while composition retrieval contributes complementary structural evidence; the framework is trained without composition-specific supervision.Source: §7.3 · p. 9

8 Conclusion

The conclusion presents HyPhID as a retrieval-enhanced, explainable framework for XRD phase identification and reports strong performance on experimentally measured patterns for both crystal-system and space-group prediction.

  • The framework combines convolutional-autoencoder representations, adaptive cosine retrieval, peak-level explanations, and composition embeddings for phase identification.Explanation-guided retrieval supports predictions with explicit diffraction evidence, while compositional retrieval provides chemically grounded prior information.Source: §8 · pp. 9–10
  • Table 3 reports accuracy for individual decision components and final HyPhID inference across crystal-system and space-group classification.Source: Table 3 · p. 10
  • Figure 5 illustrates source disagreement: the model predicts cubic correctly while retrieval and IG suggest alternatives, and retrieval selects SG 166 instead of ground-truth SG 225.Source: Figure 5 · p. 10
  • Experimental XRD evaluation demonstrates strong performance for both crystal-system and space-group prediction.Source: §8 · p. 10
  • The framework focuses on 22 space groups; extension to all 230 depends on sufficient data, while unseen compounds and mixed-phase signals remain future-work targets.Source: §8 · pp. 10–11
  • The implementation specifying architecture, training hyperparameters, data augmentation, and inference settings is available in the accompanying GitHub repository.Source: §8 · p. 11
Loading 2609.31888v1…