X-ray diffraction pattern analysis assisted by convolutional neural networks

ABSTRACT

This work presents an automated methodology for identifying inorganic compounds from powder X-ray diffraction (XRD) patterns using convolutional neural networks (CNNs). Crystallographic information files (CIFs) were automatically retrieved from the Crystallography Open Database (COD) and processed to generate theoretical diffractograms, including peak positions and intensities derived from structure factor calculations. A total of 33,180 inorganic compounds were processed. To enable CNN-based classification, diffraction patterns were transformed into RGB images using the DeepInsight approach combined with t-distributed stochastic neighbor embedding (t-SNE), allowing the homogenization of feature dimensionality across the dataset. The resulting images were used to train a convolutional neural network for feature extraction, followed by fully connected artificial neural network layers for multiclass classification. The proposed model achieved an average classification accuracy of 69% across 33,180 classes. Compared to traditional search–match approaches, this method reduces computational complexity and eliminates the need for manual interpretation of diffractograms. This work represents a fully automated pipeline integrating data acquisition, diffractogram generation, image transformation, and large-scale compound identification. The results demonstrate a modest yet promising performance, highlighting the dependence of the approach on the selected CNN architecture.

Keywords:
Crystallography; Software; Diffraction Patterns; Python; Intensities

1. INTRODUCTION

In the current era, interest in the use of artificial intelligence (AI) models has been steadily increasing. Their applications span a wide range of fields, including economics and finance [¹], agriculture [²], sports analytics [³], business strategy planning [⁴], and medicine. This growing adoption reflects the ability of AI-based methods to extract meaningful information from large and complex datasets.

Machine-learning (ML) techniques have become powerful tools in many scientific and engineering disciplines, such as computer science, telecommunications, chemistry, physics, materials science, and environmental science. Evidence of this trend can be observed over the past decade in the increasing number of ML-related publications reported in databases such as Web of Science [⁵].

Powder X-ray diffraction (XRD) is a non-destructive technique widely used for the characterization of crystalline materials. In this method, electromagnetic radiation with a typical wavelength of 1.541 Å is directed onto a sample of powder [⁶]. The resulting diffraction pattern consists of a superposition of Bragg reflections at characteristic angular positions. Analysis of these patterns allows for the qualitative and quantitative determination of the crystalline phases of the samples and other structural information [⁷].

Several software tools are available for XRD data analysis. Among the most widely used open-source solutions is the general structure analysis system (GSAS) [⁸]. Commercial alternatives such as HighScore Plus [⁹] and PROFEX [¹⁰] are also commonly employed in crystallographic analysis. However, a major limitation of these applications is their reliance on search-match algorithms, which require significant computational resources and often do not facilitate direct visual comparisons between compounds. Moreover, the accuracy of the results strongly depends on experimental conditions and the quality of the data. Proper interpretation frequently demands expert knowledge due to complex user interfaces and the large volume of information to be analyzed. Experimental artifacts, such as detector misalignment or poor sample preparation, may introduce spurious peaks and compromise the reliability of the analysis.

In recent years, several studies have explored the application of machine-learning techniques to XRD data analysis, addressing tasks such as phase identification, space-group classification, Rietveld refinement, and lattice parameter extraction. Most of these studies rely on diffractograms obtained from the Inorganic Crystal Structure Database (ICSD), which, as of 2023, contained information on approximately 291,293 crystal structures.

Table 1 summarizes representative studies from the state of the art in diffraction-pattern analysis, including their datasets, methodologies, and main results. Representative approaches have addressed search-match algorithms based on optimized Rietveld refinement [¹¹], crystal-system and space-group classification using convolutional neural networks [¹²], crystal-structure classification from diffraction patterns [¹³], crystallographic prediction from diffraction and chemistry data [¹⁴], prediction of space groups from ICSD-generated diffractograms [¹⁵], and artifact identification in X-ray diffraction data using gradient boosting methods [¹⁶].

Table 1
State of the art in diffraction-pattern analysis.

AUTHOR	ADDRESSED PROBLEM	DATASET	MODEL	ACCURACY	REMARKS
Woon Bae Park et al. (2017) [12]	Classify space group and crystal systems	150,000 ICSD XRD patterns	CNN	94.99%	The XRD pattern was treated as a picture
Jin-Woong Lee (2020) [13]	Identify phases of multiphase inorganic compounds	170 inorganiccompounds	CNN	98.13%	Used 1,785,405 synthetic XRD patterns
Henrik Schopmans (2023) [14]	Predict the space group of unseen structures	272,000 ICSD XRD patterns	CNN + ResNET	79.9%	Used synthetically generated crystals
Jeffery A. Aguiara (2019) [15]	Predict crystallographic structure from diffraction and chemistry data	431,000 COD patterns	ANN	85%	Used a SQL database to store patterns
Howard Yanxon (2022) [16]	Identify artifacts in X-ray diffraction data	58 raw experimental images	Gradient boosting	99.94%	Identified and separated the confounding factors in patterns

In the current work, we address the problem of recognizing the chemical composition of powder XRD patterns at the compound level. To this end, we used data from the Crystallography Open Database (COD) [¹⁷], which provides crystallographic information in standardized CIF format. An automated data-mining algorithm was developed to download, process, and generate diffractograms directly from these files. The resulting patterns were transformed into bitmap images using t-distributed stochastic neighbor embedding (t-SNE) through the DeepInsight methodology [¹⁸].

The resulting image-based dataset was used to train a convolutional neural network to recognize diffraction patterns automatically (Figure 1). Unlike traditional search-match approaches, this method significantly reduces computational requirements, minimizes the need for expert intervention, and rapidly identifies candidate compounds. Using this approach, an accuracy of 69% was achieved with a dataset of 33,180 inorganic compounds.

Figure 1
Example of CNN architecture.

To the best of our knowledge, no current software performs direct compound-level diffractogram recognition using a fully automated pipeline of this scale. The results presented here demonstrate the potential of the proposed methodology for applications in crystallographic analysis within chemistry, biology, and artificial intelligence.

The general strategy of the proposed approach is described in Section 2. A detailed description of the algorithm and the results obtained are presented in Section 3. Finally, a discussion and conclusion are provided in Section 4.

2. MATERIALS AND METHODS

The implementation of the artificial intelligence model was carried out in the phases listed in Figure 2. Information about each compound was obtained from the Crystallography Open Database (COD) [¹⁷].

Figure 2
Flow chart of the process used.

The COD is an open database sponsored by the University of Cambridge. It currently has 473,797 entries on organic, inorganic, metal organic, and mineral compounds. It is continually fed input from publications by researchers from all over the world. All the information about the compounds is stored in crystallographic information files (CIFs), which are standardized text files coded with a unique numeric identifier of seven digits [¹⁹]. Once the database was homologated and catalogued, a CNN was implemented.

3. RESULTS

We sought to develop a computational algorithm that would make it possible to download and clean up crystallographic information from the COD and calculate and display a given diffractogram with a list of peaks. There were several key criteria. The algorithm had to be designed to adapt and adjust to the structure and atomic complexity of the selected compounds. The routine had to be shielded against inconsistencies and syntax errors in the files, and the process had to be automated from start to finish, without the need for human intervention.

The algorithm was developed with its corresponding subroutines as shown in Figure 3. In the following discussions, the subroutines are described and the methodology is explained in detail.

Figure 3
Algorithms used for downloading and creating the database.

The compounds were selected from the COD named Inorganic Chemistry Journal, which has 49,741 entries, and filtered. Then a text file listing the URL addresses of the CIF files for each of the entries was created, as shown in Figure 4.

Figure 4
Examples of CIF file URLs.

3.1. CIF file reading

The downloading and reading of the selected CIF files were performed using a “for” loop that iterated through the entire list of URL addresses for the compounds (Figure 4). The file information was temporarily stored in memory, from which it could be readily extracted.

3.2. Lattice extraction from the space group

Exception handling was used to ascertain that the file contained the required data; if it did not, the file for the compound was discarded. The information required was the size and volume of the cell, the space group, and the coordinates of each atom in the compound; this information was stored in an array named “picos.”

3.3. Initial miller index range

An initial range of reflection indices for hkl was defined from 5 to -5, to be stored later in an array. To accomplish this, a subroutine was created to consider all possible combinations. An array called “hkl,” with 1,131 rows, was generated. The pseudocode used is shown in Figure 5.

Figure 5
Initial Miller indices range.

The space-group number had to be extracted, by using it as the value of the variable_symmetry_Int_Tables_number. With this information, the type of crystal system and the corresponding Bravais lattice could be determined.

To achieve this, a CSV file was created, characterizing the relationships among the space group number (space_group_IT_number), the crystal system, and the Bravais lattice. This information was stored in a DataFrame-type variable from the Pandas library. The CSV file can be displayed in the MS Excel interface, where the space-group number, the type of crystal system, and the corresponding Bravais lattice.

3.4. Reflection restrictions

The reflection restrictions for each type of cell were defined based on the corresponding space group, shown in Table 2.

Table 2
Reflection restrictions.

CELL TYPE	ALLOWED REFLECTIONS	REFLECTIONS NOT ALLOWED
Primitive (P)	Any H, K, L	None
Body-Centered (I)	H+K+L = 2n	H+K+L = 2n+1
Face-Centered (F)	H, K, L all even and odd	H, K, L mixed even and odd
Base-Centered (C)	H+K = 2n	H+K = 2n+1
Face Cubic-Centered (FCC)	H, K, L oddH, K, L even and H+K+L = 4n	H, K, L Mixed even and odd H, K, L all none and H+K+L = 4n
Rhombohedral (R)	–H+K+L = 3n	–H+K+L ≠ 3n
Hexagonal (HCP)	L even H+2K ≠ 3n	L even & H+2K = 3n

Using the “loc” method, the “picos” array was accessed based on the index and the type of unit cell. First-row filtering was performed, resulting in another array, called “Picosres.” (A fragment of the code used is shown in Figure 6). Then the interplanar distance was calculated for each index, depending on the type of crystal system, using the equations in Table 3 [²⁰].

Figure 6
Code for verifying valid results.

Table 3
Equations for calculating interplanar spacing.

6 SYSTEM TYPE	INTERPLANAR SPACING
Cubic	$\frac{1}{d^2}=\frac{h^2+k^2+l^2}{a^2}$
Tetragonal	$\frac{1}{d^2}=\frac{h^2+k^2}{a^2}+\frac{l}{c^2}$
Hexagonal	$\frac{1}{d^2}=\frac{4}{\text{3}}\left(\frac{h^2+hk+k^2}{a^2}\right)+\frac{l^2}{c^2}$
Rhombohedral	$\frac{1}{d^2}=\frac{\left(h^2+k^2+l^2\right)si{n}^2\alpha +2\left(hk+kl+hl\right)co{s}^2\alpha -\cos \alpha }{a}$
Orthorhombic	$\frac{1}{d^2}=\frac{h^2}{a^2}+\frac{k^2}{b^2}+\frac{l}{c^2}$

To calculate theta and 2 theta, it was necessary to verify the maximum geometric limit of $sin\theta =\frac{\lambda }{2d}$, where $\frac{\lambda }{2d}<1$

3.5. Interplanar and 2theta calculation

To achieve this, a separate fragment of code was created to calculate the right-hand side of the interplanar distance equations. The remaining operations were executed within an exception-handling block. In the event of any type of mathematical error, the entire row combination was discarded, and the procedure continued with the next indices. The code is shown in Figure 6.

3.6. Intensity calculation using the structure factor

The results were stored by routine in an array containing information about the indices, the interplanar distance, and the corresponding angles theta and 2theta for each compound. The factors a_i, b_i, and c_i were obtained from the IUCr Tables [²¹]. A DataFrame was created with the factors for each atom. The structure factor (equation 1) for each index was calculated using the imaginary and real parts, Fcos and Fsin, for each atom in the compound [⁶].

(1) $$I=F_{hkl}^2=\left[{\displaystyle \sum _{n}fncos\left[2\pi (hx+ky+lz)\right]}\right]{}^2+{\left[{\displaystyle \sum _{n}fnsin\left[2\pi (hx+ky+lz)\right]}\right]}^2$$

Equation 1. The structure factor.

In the code, a “for” loop was created that iterated over each atom in the compound, reading the coefficient corresponding to each atom at each pass and performing the required operations. A concatenated sum was also performed to obtain the cosine and sine parts. The resulting DataFrame was saved in a CSV file with the name of the unique COD code shown in Figure 7.

Figure 7
Code used for saving output file.

3.7. Generating the database of representative patterns and features

Once the compound was processed, the variables were cleared, and the next file was processed until the batch was complete. Each file was saved in a folder on a hard drive. If the user requires it, the diffraction pattern can be obtained by plotting the intensity against twice the angle (Figure 8).

Figure 8
Diffraction patterns obtained.

It was necessary to perform test runs of the routine with small batches of files (10, 20, 50, and 100) until the final version was obtained (Figure 9).

Figure 9
Total files processed.

3.8. CNN database preprocessing

3.8.1. Homologation

The homologation of traits is required to obtain satisfactory training. Once 33,180 patterns were obtained, an analysis was carried out to detect variations in the number of features. The highest number of features was 171 and the lowest 5.

For this purpose, the DeepInsight algorithm [¹⁸] was used, a diagram of which is shown in Figure 10. Its function as the input to the CNN model is described in detail below.

Figure 10
DeepInsight algorithm. (Source: Sharma, Vans, Shigemizu, Boroevich, & Tsunoda, 2019).

T-distributed stochastic neighbor embedding is a convolutional algorithm that creates a probability distribution which represents the similarities between neighbors in a lower-dimensional space with Euclidean distances. The probability that trait x_i is similar to trait x_j was calculated with Equation 1:

(2) $$p_{ij}=\frac{\frac{exp(-||x_i-x_j|{|}^2)}{2{\sigma }^2}}{{\displaystyle {\sum }_{k\ne 1}\frac{exp(-||x_k-x_l|{|}^2)}{2{\sigma }^2}}}$$

Equation 2. Probability calculation for x.

For the DeepInsight algorithm, Equation 3 was used:

(3) $$\sigma =\frac{1}{\sqrt{2}}$$

Equation 3. Sigma value.

Furthermore, to obtain the corresponding coordinates y_i, y, y_j, an equation with the same σ value (Equation 4) was used:

(4) $$q_{ij}=\frac{\frac{\exp (-||y_i-y_j|{|}^2}{2{\sigma }^2}}{{\displaystyle {\sum }_{k\ne 1}\frac{\exp (-||yi-y_k|{|}^2}{2{\sigma }^2}}}$$

Equation 4. Probability calculation for y.

If the conditional probabilities p and q are equal, it means that the sample points y_i, y, y_j correctly model the similarity between the higher-dimensional samples x_i, y, x_j.

In this case, a vector of n x 2 and another of n x 1 dimensions, corresponding to the intensity and 2theta previously calculated for each pattern obtained, were used, as shown in Table 4.

Table 4
Intensity and 2 theta vectors used in the COD_1000118.csv file.

2THETA		INTENSITY
1	0.756395	376.1534
2	0.534616	406.0342
3	0.773454	568.9839
4	0.540532	418.3888
5	0.779403	1433.401
6	0.542548	1649.322
7	1.063774	402.0422
8	0.740725	807.1872
9	1.113129	2.126902
10	0.756721	1.821347
11	1.131189	63.91614
12	0.762288	397.3654
13	2.655846	25.29054

Then, the convex hull algorithm was applied to determine the minimum box that covers all the points. Since this box was not in the horizontal or vertical direction, a gradient rotation of the two lower corners of the box was performed, based on Equation 4 and Equation 5.

(5) $$Gr=\frac{yc2-yc1}{xc2-xc1}$$

Equation 5. Rotation gradient.

(6) $$\theta ={\tan }^{-1}(Gr)$$

Equation 6. Rotation angle.

Figure 11 shows how this method was applied to the point cloud obtained from the previous steps. Once the gradient angle was obtained, a corresponding rotation was performed.

Figure 11
Convex hull algorithm applied to scattered points.

The rotation matrix R was obtained from the sin θ and cos θ of the rotation angle. This was multiplied by the entire data set, producing a table that aligned with the horizontal and vertical axes, according to Equation 6, Equation 7 and Equation 8.

(7) $$R=\left[\begin{array}{cc}\cos \theta & \sin \theta \\ -\sin \theta & \cos \theta \end{array}\right]$$

Equation 7. Rotation matrix.

(8) $$Ac=|x_{d2}-x_{d1}|$$

Equation 8. X Cartesian coordinate.

(9) $$Bc=|y_{d3}-y_{d2}|$$

Equation 9. Y Cartesian coordinate

Using the ceil function, the Cartesian coordinates were converted into a bitmap (based on Equation 4, 11, 12 and 13), considering the minimum distance of the two points closest to each corner.

(10) $$A_p=ceil\left(A_c\times \frac{Precision}{d_{min}}\right)$$

Equation 10. Conversion from x Cartesian to x bitmap coordinate.

(11) $$B_p=ceil\left(B_c\times \frac{Precision}{d_{min}}\right)$$

Equation 11. Conversion from y Cartesian to y pixel coordinate.

(12) $$x_p=round\left(1+\frac{(x_c-x_{min})\times A_p}{x_{max}-x_{min}}\right)$$

Equation 12. Calculation for the x pixel coordinate.

(13) $$y_p=round\left(1+\frac{(y_c-y_{min})\times A_p}{y_{max}-y_{min}}\right)$$

Equation 13. Calculation for the y pixel coordinate.

This routine was applied to the 33,180 patterns in the database, obtaining distributions like the examples shown in Figure 12 and Figure 13. The image output parameters are given in Table 5. With these parameters, the homologation of the characteristics of the patterns was complete.

Figure 12
DeepInsight results in a COD_4314932 pattern.

Figure 13
DeepInsight results in a COD_4314923 pattern.

Table 5
Image output parameters.

PARAMETER	VALUE
Extraction Method	TSNE
Pixel Numbers	600 × 600
X Variable	Intensity 2 Theta
Y Variable	Peak Index
File Output	PNG
Color Channel	RGB

In addition to the image file, a histogram with the indices and the intensity with its corresponding label was obtained at a resolution of 640 × 480 pixels, which had to be down-sampled to 600 × 600 (Figure 14). The size of the database was 369 GB, with a total processing time of 1,272 hours.

Figure 14
Histogram obtained from COD_1100959 pattern.

Although powder X-ray diffraction data are inherently one-dimensional (intensity vs. 2θ), the dataset employed in the present study exhibits substantial variability in the number of diffraction peaks per sample, ranging from 5 to 171 peaks.

This variability complicates the direct use of conventional 1D CNNs, since additional preprocessing procedures, such as padding, truncation, and interpolation, are typically required to standardize the input dimensions. Such operations may alter the original diffraction information and affect the preservation of features.

To address this limitation, the DeepInsight (t-SNE) approach was employed to transform diffraction features with vectors of varying lengths into fixed-size 2D representations. This transformation preserved the relationships among the features and enabled CNNs optimized for image-based pattern recognition to be used.

In addition, the spatial organization generated by the t-SNE mapping placed similar features near each other in regions of the image allowing the CNN to exploit both local and global structural relationships among the diffraction peaks.

3.8.2. Catalogation

For cataloging, the K-fold cross-validation procedure [²²] was used, the main objective of which was to divide the total dataset into training sets, constituting 80% of the size of the original dataset, with the validation and test sets making up the remaining 20%, at 10% each. Once the sizes were established, the training proceeded, and the assertiveness index was obtained. Then another training run was performed, processing the test and validation sets through the entire database, as shown in Figure 15, where the number of patterns considered for each set is indicated.

Figure 15
K-fold cross-validation.

3.8.3. CNN architecture

When the cataloging was completed, four images were obtained: three in each of the RGB channels of the convolved image and one that corresponded to the diffraction pattern. The size of the latter histogram was readjusted in order to standardize it to fit with the convolution images. All images had a size of 600 × 600 pixels.

The architecture was designed with three convolution layers and three max pooling layers, with the objective of reducing the number of features to fit the parameters indicated in Figure 16. Once the processing was completed, a vector of 1 × 256 characteristics was obtained, which served as input for the artificial neural network (ANN), whose architecture is shown in Figure 17.

Figure 16
CNN feature vector dimensionality across layers.

Figure 17
ANN architecture.

The input layer consisted of 256 neurons for each feature obtained in one of the patterns, followed by five hidden layers, alternating between 17 and 38 neurons, until an output layer of 33,180 neurons was obtained, each of which corresponds to a pattern and activates when it recognizes the pattern. Finally, SoftMax was used as an activation function, so that the network reports the probability that the pattern shown corresponds to the one indicated.

In the first training session, a mean accuracy index of 69% was obtained (see Figure 18).

Figure 18
Accuracy index.

To further evaluate the accuracy of the proposed model, its classification performance was analyzed across representative crystal systems with different degrees of symmetry. Table 6 summarizes the accuracy of the results obtained for the different crystallographic systems. The results indicate that the model is more accurate when classifying highly symmetric crystal systems, such as cubic structures, and that its performance decreases progressively for low-symmetry systems, such as monoclinic structures. This behavior is consistent with the characteristics of powder X-ray diffraction patterns. Highly symmetric structures generally produce more regular and distinguishable diffraction signatures, with less peak variability and peak overlap; in contrast, low-symmetry systems generate more complex diffraction patterns, with a greater number of reflections and more peak overlap, which makes it more challenging for a CNN to discriminate among compounds. The results suggest that the proposed model is capable of learning the structural features of crystallographic symmetries. Furthermore, the trends observed attest to the consistency and validity of the proposed learning approach.

Table 6
Classification accuracy across representative crystal systems with different levels of symmetry.

CRYSTAL SYSTEM	SYMMETRY LEVEL	ACCURACY (%)
Cubic	High	82.4
Tetragonal	Medium	75.8
Orthorhombic	Medium	69.3
Hexagonal	Medium	66.7
Monoclinic	Low	58.9

Given the large number of compounds of different classes involved in the present study (33,180 compounds), additional evaluation metrics were employed to provide a more complete assessment of classification performance. Figure 19 shows the Top-k accuracy values obtained for the proposed model. While the Top-1 accuracy reached 69%, the Top-3, Top-5, and Top-10 accuracies reached 84%, 91%, and 96%, respectively.

Figure 19
Top K classification accuracy.

In these results, the correct compound frequently appears in the highest-ranked predictions generated by the model. This is especially promising for XRD search-match applications, which typically analyze several candidate compounds, as opposed to applications that rely solely on a single prediction.

4. DISCUSSION

The results demonstrate the feasibility of using convolutional neural networks for the large-scale identification of inorganic compounds from the X-ray diffraction patterns of powders. Although the obtained accuracy of 69% is lower than that reported in studies focused on space-group or crystal-system classification, it is important to emphasize that the current work addresses a significantly more complex problem: direct compound-level identification across 33,180 distinct classes.

Previous studies have typically limited classification to a restricted number of categories, such as crystal systems or space groups, or have relied on synthetic datasets with controlled noise levels. In contrast, the dataset in this study was generated from information files for compounds of varying structural complexity, atomic composition, and symmetry, which resulted in diffraction patterns with a high degree of variability.

The transformation of diffractograms into RGB images using the DeepInsight framework proved effective for homogenizing feature dimensionality and enabled CNN training without manual feature engineering. However, this transformation may also have caused information loss, which partially explains the moderate accuracy of the classification.

Another limiting factor was the restricted availability of computational infrastructure, which constrained both the depth of the CNN architecture and the number of training epochs. It can be expected that attention mechanisms, deeper architectures, and optimized hyperparameter tuning could substantially improve performance. Additionally, the generation of synthetic diffraction patterns and data-augmentation strategies could further enhance generalization.

Despite these limitations, the proposed methodology offers a novel alternative to traditional crystallographic search-match techniques, eliminating the need for exhaustive peak-by-peak comparisons and expert intervention. The approach is particularly suitable for high-throughput crystallographic analysis and automated materials-discovery pipelines.

5. CONCLUSION

This work presents a fully automated methodology for identifying inorganic compounds from powder X-ray diffraction (XRD) patterns, using convolutional neural networks. The proposed pipeline integrates data acquisition from the Crystallography Open Database with diffractogram generation based on crystallographic information files, feature homogenization through DeepInsight, and multiclass classification, using deep-learning models.

A dataset comprising 33,180 inorganic compounds was constructed and processed without manual intervention. The trained model achieved an average classification accuracy of 69%, demonstrating the feasibility of compound-level identification in the context of a highly complex and large-scale classification problem. Unlike previous studies focused on crystal-system or space-group classification, this work addresses direct compound recognition, which significantly increases problem dimensionality.

The main limitations of the current implementation are related to computational constraints that restricted the depth of the CNN architecture and the extent of hyperparameter optimization. Additionally, the image-based representation of diffractograms may have led to partial information loss and affected classification performance.

Future work will focus on improving the accuracy of the model through optimized deep-learning architectures, synthetic data generation, and advanced data-augmentation techniques. The integration of experimental diffraction data and the extension of the methodology to multiphase samples are also planned. Overall, the proposed approach provides a scalable and automated alternative to traditional XRD analysis methods, with potential applications in materials science, chemistry, and crystallographic research.

6. ACKNOWLEDGEMENTS

We extend our appreciation to CNMN-IPN and LabCREA-CIITEC for their support in performing the characterizations carried out in this study. Alexandro I. Medina Velázquez acknowledges the CONACYT Ph D scholarship. The authors also would like to thank Henry Jankiewicz for the editing work that he did for this paper.

7. DATA AVAILABILITY

The entire dataset supporting the results of this study has been made available on SciELO Data and can be accessed at https://github.com/itzalex09/DRXreader.

8. BIBLIOGRAPHY

• ^[1] DE LOURDES GUTIÉRREZ CORDERO, M., SEGOVIA-VARGAS, M.J., ESCAMILLA, M.R., “Análisis del riesgo de caída de cartera en seguros: metodologías de “Inteligencia Artificial” vs “Modelos Lineales Generalizados”, Economía Informa, v. 407, pp. 56–86, Nov. 2017. doi: https://doi.org/10.1016/j.ecin.2017.11.004.
  » https://doi.org/10.1016/j.ecin.2017.11.004
• ^[2] SEGOVIA, J.S.B., ROJAS, F.A.D., QUISHPE, M.W.V., “Estudio del uso de técnicas de inteligencia artificial aplicadas para análisis de suelos para el sector agrícola”, Recimundo, v. 5, n. 1, pp. 4–19, 2021. doi: https://doi.org/10.26820/recimundo/5.(1).enero.2021.4-19.
  » https://doi.org/10.26820/recimundo/5.(1).enero.2021.4-19
• ^[3] VICENTE-MARTÍNEZ, J.A., MÁRQUEZ-OLIVERA, M., GARCÍA-ALIAGA, A., et al., “Adaptation of YOLOv7 and YOLOv7_tiny for soccer-ball multi-detection with DeepSORT for tracking by semi-supervised system”, Sensors, v. 23, n. 21, pp. 8693, Oct. 2023. doi: https://doi.org/10.3390/s23218693. PubMed PMID: 37960393.
  » https://doi.org/10.3390/s23218693.
• ^[4] ABDEL-BASSET, M., MOHAMED, M., SMARANDACHE, F., “An extension of neutrosophic AHP–SWOT analysis for strategic planning and decision-making”, Symmetry, v. 10, n. 4, pp. 116, Apr. 2018. doi: https://doi.org/10.3390/sym10040116.
  » https://doi.org/10.3390/sym10040116
• ^[5] SURDU, V.-A., GYŐRGY, R., “X-ray diffraction data analysis by machine learning methods—A review”, Applied Sciences, v. 13, n. 17, pp. 9992, Sep. 2023. doi: https://doi.org/10.3390/app13179992.
  » https://doi.org/10.3390/app13179992
• ^[6] BADIS, K., MERINE, H., RAMLI, Y., et al., “Effect of polymers nature and stirring speeds on physicochemical properties and the controlled release of allopurinol-loaded microspheres”, Journal of the Mexican Chemical Society, v. 66, n. 1, Dec. 2021. doi: https://doi.org/10.29356/jmcs.v66i1.1583.
  » https://doi.org/10.29356/jmcs.v66i1.1583
• ^[7] WILL, G., Powder diffraction, 1 ed., v. 58, Berlin, Springer, 2006. doi: https://doi.org/10.1007/3-540-27986-5.
  » https://doi.org/10.1007/3-540-27986-5
• ^[8] TOBY, B.H., VON DREELE, R.B., “GSAS-II: the genesis of a modern open-source all purpose crystallography software package”, Journal of Applied Crystallography, v. 46, n. 2, pp. 544–549, Mar. 2013. doi: https://doi.org/10.1107/S0021889813003531.
  » https://doi.org/10.1107/S0021889813003531
• ^[9] DEGEN, T., SADKI, M., BRON, E., et al., “The highscore suite”, Powder Diffraction, v. 29, n. S2, pp. S13-S18, Oct. 2014. doi: https://doi.org/10.1017/S0885715614000840.
  » https://doi.org/10.1017/S0885715614000840
• ^[10] DOEBELIN, N., KLEEBERG, R., “Profex : a graphical user interface for the Rietveld refinement program BGMN”, Journal of Applied Crystallography, v. 48, n. Pt 5, pp. 1573–1580, Oct. 2015. doi: https://doi.org/10.1107/S1600576715014685. PubMed PMID: 26500466.
  » https://doi.org/10.1107/S1600576715014685.
• ^[11] LUTTEROTTI, L., PILLIÈRE, H., FONTUGNE, C., BOULLAY, P., CHATEIGNER, D., “Full-profile search–match by the Rietveld method”, Journal of Applied Crystallography, v. 52, n. 3, pp. 587–598, 2019. doi: https://doi.org/10.1107/S1600576719005044
  » https://doi.org/10.1107/S1600576719005044
• ^[12] PARK, W.B., CHUNG, J., JUNG, J., et al., “Classification of crystal structure using a convolutional neural network”, IUCrJ, v. 4, pp. 486–494, 2017. doi: https://doi.org/10.1107/S205225251700714X.
  » https://doi.org/10.1107/S205225251700714X
• ^[13] RA, M., BOO, Y., JEONG, J.M., BATTS-ETSEG, J., JEONG, J., LEE, W., “Classification of crystal structures using electron diffraction patterns with a deep convolutional neural network”, RSC Advances, v. 11, n. 61, pp. 38307–38315, 2021. doi: https://doi.org/10.1039/D1RA07156D.
  » https://doi.org/10.1039/D1RA07156D
• ^[14] AGUIAR, J.A., GONG, M.L., TASDIZEN, T., “Crystallographic prediction from diffraction and chemistry data for higher throughput classification using machine learning”, Computational Materials Science, v. 173, 109409, 2020. doi: https://doi.org/10.1016/j.commatsci.2019.109409.
  » https://doi.org/10.1016/j.commatsci.2019.109409
• ^[15] SCHOPMANS, H., REISER, P., FRIEDERICH, P., “Neural networks trained on synthetically generated crystals can extract structural information from ICSD powder X-ray diffractograms”, Digital Discovery, v. 2, 2023. doi: https://doi.org/10.1039/D3DD00071K.
  » https://doi.org/10.1039/D3DD00071K
• ^[16] YANXON, H., WENG, J., PARRAGA, H., XU, W., RUETT, U., SCHWARZ, N., “Artifact Identification in X-ray Diffraction Data Using the Gradient Boosting Method”, Lecture Notes in Networks and Systems, pp. 508–515, 2024. doi: https://doi.org/10.1007/978-3-031-47718-8_34.
• ^[17] GRAŽULIS, S., DAŠKEVIČ, A., MERKYS, A., et al., “Crystallography Open Database (COD): an open-access collection of crystal structures and platform for world-wide collaboration”, Nucleic Acids Research, v. 40, pp. D420-D427, Jan. 2012. doi: https://doi.org/10.1093/nar/gkr900. PubMed PMID: 22070882.
  » https://doi.org/10.1093/nar/gkr900.
• ^[18] SHARMA, A., VANS, E., SHIGEMIZU, D., et al., “DeepInsight: A methodology to transform a non-image data to an image for convolution neural network architecture”, Scientific Reports, v. 9, n. 1, pp. 11399, Aug. 2019. doi: https://doi.org/10.1038/s41598-019-47765-6. PubMed PMID: 31388036.
  » https://doi.org/10.1038/s41598-019-47765-6.
• ^[19] BROWN, I.D., MCMAHON, B., “CIF: the computer language of crystallography”, Acta Crystallographica. Section B, Structural Science, v. 58, pp. 317–324, Jun. 2002. doi: https://doi.org/10.1107/S0108768102003464. PubMed PMID: 12037350.
  » https://doi.org/10.1107/S0108768102003464.
• ^[20] DAVID, B., Structure determination from powder diffraction data, 2001, https://www.iucr.org/__data/assets/pdf_file/0020/21629/cpd25.pdf, accessed in June, 2026.
  » https://www.iucr.org/__data/assets/pdf_file/0020/21629/cpd25.pdf
• ^[21] PRINCE, E., International Tables for Crystallography: Mathematical, physical and chemical tables, Chester, England, International Union of Crystallography, 2006. doi: https://doi.org/10.1107/97809553602060000103.
• ^[22] BERRAR, D., “Cross-validation”, In: Ranganathan, S., Gribskov, M., Schönbach, C. (eds), Encyclopedia of Bioinformatics and Computational Biology, Amsterdam, Elsevier, pp. 542–545, 2019. doi: https://doi.org/10.1016/B978-0-12-809633-8.20349-X.
  » https://doi.org/10.1016/B978-0-12-809633-8.20349-X

Publication Dates

• Publication in this collection
  31 July 2026
• Date of issue
  2026

History

• Received
  04 Feb 2026
• Accepted
  08 June 2026