AbstractWe respond to Schaeben et al.’s1 comment on our paper, “Machine Learning Enhanced Analysis of EBSD Data for Texture Representation.” While their observations are factually correct, they do not disprove our results. Our method, TACS, preserves the full distribution of crystallographic orientations and is validated with real-world data. We emphasize the importance of empirical validation over theoretical constructs in assessing machine learning methods’ practical performance.
ResponseIn their comment, Schaeben et al.1 present a critique questioning the validity of our TACS methodology, suggesting that our approach is fundamentally flawed based on their theoretical examples. They argue that our use of pole figures to assess the performance of TACS is inappropriate. However, visual inspection of pole figure maps remains a widely accepted and standard method for texture validation in materials science2,3,4,5,6,7,8. Moreover, preserving the crystallographic texture in pole figures is sufficient for many practical applications, such as crystal plasticity simulations. The TACS method was primarily developed with these applications in mind, which is why we used pole figure maps as part of our validation process. In fact, similar approaches have been used by others in the field. For example, Marki et al.8 developed a nonlinear optimization method to compact orientation distribution functions (ODFs) for crystal plasticity simulations. They reduced the number of crystal orientations without sacrificing accuracy by comparing compacted and full-size ODFs using the texture difference index (TDI) and visually validating their method through stereographic pole figures. Their results further affirm that pole figure maps can serve as an appropriate tool for validation.The authors suggest that we are mistaking examples for proof without providing a detailed mathematical interpretation. Proposing TACS methodology does not require mathematical proof as its use must require the user of the tool to assess the suitability of representation in the same way that Schaeben et al.1 have demonstrated the rotational difference in textures1. Moreover, as TACS is based on the K-means machine learning (ML) algorithm, it is widely accepted that ML algorithms are best validated by testing on a large number of datasets. Attempting to develop a mathematical proof would defeat the purpose of employing an ML-based approach, which is inherently designed to handle complex, real-world data. While TACS might not perform perfectly for every dataset, this is a common characteristic of ML-based approaches—they are not a universal solution for all scenarios. However, the TACS technique outperformed kernel density estimation (KDE) in 20 examples of our manuscript. The use of examples of representation using pole figure comparisons is exactly what users of TACS will perform for themselves to assess whether or not results are beneficial for their work.In addition, one key issue with the comment is the assumption that our method treats the three Euler angles (φ1, Φ, φ2) as independent variables. This is a fundamental misunderstanding of TACS. Our method treats these angles as a single entity, thereby preserving the full distribution of crystallographic orientations. The three angles provide information to evaluate an orthogonal rotation tensor between the crystal frame and the lab/sample frame so they must be taken together. The matching of pole figure maps is a consequence of this preservation.To demonstrate that TACS preserves the full distribution of crystallographic orientations, we provide sigma sections of the orientation distribution function (ODF) generated by TACS using reduced datasets of 350 and 100 crystal orientations, compared with the ODF from the raw dataset. These comparisons for three example datasets (cubic, hcp, and orthorhombic) are illustrated in Fig. 1. The close similarity of the ODFs demonstrates that TACS is capable of preserving the distributions of crystallographic orientations. Please note that these three examples were arbitrarily chosen from the 20 datasets that were used in the manuscript. All raw datasets and reduced-order datasets are available in the repository, along with a code to plot the sigma sections.Fig. 1Comparison of sigma sections from three example datasets showing the distribution of crystallographic orientations in the raw datasets and the reduced-order datasets generated by TACS.Full size imageThe comment also critiques our use of the Kolmogorov–Smirnov (K–S) test, arguing that it is not an ideal statistical measure. While we agree that the K–S test alone is insufficient, we employed it as a supplementary check to demonstrate that TACS quantitatively outperforms KDE in preserving the distribution of crystallographic orientations. We used a combined approach of pole figure comparison and the K–S test to evaluate the similarity of the full distribution of crystallographic orientations and agreed upon during the review. The significant difference in the K–S test scores demonstrated that TACS outperforms KDE in preserving the distribution of crystallographic orientations. We have shown that the KDE method cannot adequately capture crystallographic texture in pole figures without extensive parameter tuning, rendering it unsuitable for robust dataset size reduction. We addressed this limitation in detail in our manuscript. In addition, we demonstrate how the ML-based approach avoids human bias in reducing data sets with other techniques.ConclusionIn conclusion, the comprehensive approach employed by TACS—preserving the full distribution of crystallographic orientations and maintaining accuracy even with reduced datasets—demonstrates its effectiveness as a robust and practical tool for dataset size reduction for broad applications. TACS consistently outperforms conventional methods such as KDE in key areas, proving its value for a wide range of applications. It is misguided to suggest that conventional methods such as kernel density estimation (KDE) would be more suitable, as we have demonstrated in our manuscript that KDE fails to capture crystallographic texture accurately without extensive parameter tuning. This limitation is a clear reason why KDE is impractical for robust dataset reduction, which we addressed in detail. Finally, we expect any rigorous use of our methodology to be done so with full recognition of any inherent assumptions in Machine Learning techniques, as well as recognizing the inherent enhancement of the elimination of potential human bias in reducing data sets with other techniques.
ReferencesSchaeben, H. & Gerald van den Boogaart, K. Comment on “Machine learning enhanced analysis of EBSD data for texture representation”. npj Comput. Mater. https://doi.org/10.1038/s41524-024-01324-4 (2024).Mayeur, J. R. & McDowell, D. L. A three-dimensional crystal plasticity model for duplex Ti–6Al–4V. Int. J. Plast. 23, 1457–1485 (2007).Article
CAS
Google Scholar
Riyad, I. A. et al. Modeling the role of local crystallographic correlations in microstructures of Ti-6Al-4V using a correlated structure visco-plastic self-consistent polycrystal plasticity formulation. Acta Mater. 203, 116502 (2021).Article
CAS
Google Scholar
Fan, X. G. et al. Modeling the anisotropy of hot plastic deformation of two-phase titanium alloys with a colony microstructure. Int J. Plast. 104, 173–195 (2018).Article
CAS
Google Scholar
Zhang, J., Li, X., Xu, D. & Yang, R. Recent progress in the simulation of microstructure evolution in titanium alloys. Prog. Nat. Sci. Mater. Int. 29, 295–304 (2019).Article
CAS
Google Scholar
Pinz, M., Benzing, J. T., Pilchak, A. & Ghosh, S. A microstructure-based porous crystal plasticity FE model for additively manufactured Ti-6Al-4V alloys. Int J. Plast. 153, 103254 (2022).Article
CAS
Google Scholar
Raabe, D. & Roters, F. Using texture components in crystal plasticity finite element simulations. Int J. Plast. 20, 339–361 (2004).Article
Google Scholar
Marki, R. E. & Knezevic, M. Nonlinear optimization for compact representation of orientation distributions based on generalized spherical harmonics. J. Mech. Phys. Solids 187, 105609 (2024).Article
Google Scholar
Download referencesAcknowledgementsThe authors acknowledge funding for this work was provided by the Center for Extreme Events in Structurally Evolving Material under U.S. Army CC-APG-RTP Division contract number W011NF-23-2-0073. The funder played no role in the study design, data collection, analysis, and interpretation of data, or the writing of this manuscript.Author informationAuthors and AffiliationsDepartment of Materials Science and Engineering, University of Wisconsin-Madison, Madison, WI, USAJ. Wanni, C. A. Bronkhorst & D. J. ThomaDepartment of Mechanical Engineering, University of Wisconsin-Madison, Madison, WI, USAC. A. Bronkhorst & D. J. ThomaAuthorsJ. WanniView author publicationsYou can also search for this author in
PubMed Google ScholarC. A. BronkhorstView author publicationsYou can also search for this author in
PubMed Google ScholarD. J. ThomaView author publicationsYou can also search for this author in
PubMed Google ScholarContributionsJ. Wanni: conceptualization, methodology, data curation, and writing—original draft. C.A. Bronkhorst: conceptualization, supervision, resources, writing—review and editing, and funding acquisition. D. J. Thoma: conceptualization, supervision, resources, funding acquisition, and writing—original draft.Corresponding authorCorrespondence to
D. J. Thoma.Ethics declarations
Competing interests
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.
Additional informationPublisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.Rights and permissions
Open Access This article is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License, which permits any non-commercial use, sharing, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if you modified the licensed material. You do not have permission under this licence to share adapted material derived from this article or parts of it. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by-nc-nd/4.0/.
Reprints and permissionsAbout this articleCite this articleWanni, J., Bronkhorst, C.A. & Thoma, D.J. Reply to: Comment on “Machine learning enhanced analysis of EBSD data for texture representation”.
npj Comput Mater 11, 73 (2025). https://doi.org/10.1038/s41524-025-01562-0Download citationReceived: 23 September 2024Accepted: 19 December 2024Published: 17 March 2025DOI: https://doi.org/10.1038/s41524-025-01562-0Share this articleAnyone you share the following link with will be able to read this content:Get shareable linkSorry, a shareable link is not currently available for this article.Copy to clipboard
Provided by the Springer Nature SharedIt content-sharing initiative