materials-chemistry/ai produced the result/Advanced Science 2026 · v2
Machine learning predicts where dopant atoms sit in hexaferrite crystals before refinement
Researchers trained tree-based models on 494 dopant records drawn from 371 refined hexaferrite structures. From chemical composition alone, the models predict which atomic site a dopant occupies, and attach a confidence level to each guess.
spectrum · one line per step, placed by what the step does · bright lines used AI
Pre‐Refinement Prediction of Dopant Site Occupancy in M‐Type Hexaferrites via Hierarchical and Uncertainty‐Aware Machine‐Learning
Advanced Science, 2026
doi:10.1002/advs.78493 · record aix-00279 v2 · checked 2026-11-04
- AI was for
- Classification, Property prediction
- Model family
- Gradient-boosted trees, Random forest, Linear model, Transformer
- Checked by
- Experimental1 tested
- Code
- not reported
The finding the paper is about came from the AI.
What this research was about
M-type hexaferrites are magnetic ceramics built from barium or strontium, iron and oxygen. Their iron atoms sit in five distinct kinds of site within the repeating crystal pattern, labelled 12k, 4f1, 4f2, 2a and 2b. Chemists tune the material's behaviour by doping it, swapping in a little of some other element. Which site that newcomer takes, and how it spreads across the five iron sites, shapes the result. Working this out normally means Rietveld refinement: fitting a model of the crystal to a diffraction pattern, the scattering fingerprint a beam leaves after passing through a powder. That fitting needs a starting guess about site occupancy, and arriving at a good one has relied on trial and error guided by experience.
The authors set out to supply that starting guess from composition alone, before any refinement is attempted, and to say how much each guess should be trusted.
Where AI came in
The models work in two stages. First, a classifier decides whether the dopant prefers the barium or strontium sublattice or the iron sublattice. Then, for iron-sublattice dopants, regression models estimate what fraction of each of the five iron sites the dopant takes. Both stages were trained from scratch on the collected records, using gradient-boosted trees and robust regression; a ready-made tabular model, TabPFN, was run alongside as a comparison. On 50 held-out records the classifier placed 48 correctly, and the occupancy predictions reached a macro mean absolute error of 0.0584.
What the models stand in for is the crystallographer's prior reasoning about where a dopant should go. Agreement between models, the TabPFN comparison and how far a case sits from the training chemistry were combined into a confidence level and a refinement-priority label, with 44 of the 50 validation records marked high or medium confidence. One sample, a magnesium and scandium co-doped strontium hexaferrite, was measured by neutron powder diffraction at 300 K and compared with the predictions. The models were then run over 96 rare-earth host and dopant combinations that have not been measured.
Written by AIxSci from the checked record below, to give context for readers outside the field. It is not part of the record.
The work
Technical · from the record
The authors assembled 494 dopant-level records from 371 Rietveld-refined M-type hexaferrite structures and fitted tree-ensemble and robust-regression models that predict, from composition alone, whether a dopant enters the Ba/Sr sublattice or the Fe sublattice and how it distributes over the five inequivalent Fe sites. On a composition-level external set of 50 records, the CatBoost sublattice classifier assigned 48 correctly (96% accuracy), and for the 37 Fe-sublattice records the occupancy models reached a macro mean absolute error of 0.0584 with 94.59% top-1 dominant-site accuracy and 91.53% pairwise ranking accuracy. Each prediction also carries a confidence level and a refinement-priority label derived from model agreement, a TabPFN baseline and training-domain coverage, with 44 of the 50 validation records assigned high or medium confidence. The framework was then run over 96 rare-earth host-dopant combinations, and predictions for a W-doped and Sc-doped case were compared with refined and neutron-diffraction occupancies.
How AI was used
Dopant amounts and site occupancies were extracted from refined CIF files into dopant-level records, and descriptors were built only from pre-refinement composition information: host type, target-dopant elemental properties, target-dopant amount and co-doping environment. Feature sets of increasing chemical detail were compared by grouped ablation within the training set, with a composition-level holdout keeping identical normalised formulas in one subset. Occupancy inference was formulated hierarchically: a first stage classifies Ba/Sr-site versus Fe-sublattice preference using the Fe-sublattice occupancy probability with a 0.5 threshold, where CatBoost was chosen from candidate classifiers by grouped five-fold cross-validation; a second stage, applied only to Fe-sublattice dopants, estimates fractional occupancies at the 12k, 4f1, 4f2, 2a and 2b sites, first through a homogeneous regression reference and then through a heterogeneous cooperative model that pairs a site-wise occupancy classifier with an occupancy-amount regressor. TabPFN was run as an independent small-sample baseline, and a reference model plus training-domain coverage were combined with classifier confidence to assign reliability levels and refinement-priority labels. SHAP, feature-group ablation and feature-perturbation analyses were used to examine which descriptors drive each stage, and the trained models were finally applied to unmeasured rare-earth host-dopant-amount combinations.
The shape of the work
Structural · the record, drawn
Build dopant-level dataset from refined structures
Obtaining raw data, whether by measurement, download or retrieval.
After structural screening, 371 qualified crystal structures were retained and transformed into 494 dopant‐level recordswhere the paper describes this · verbatim
Construct and select composition-derived features
Encoding data into features, descriptors, embeddings or graphs.
Features were constructed exclusively from pre‐refinement composition information.where the paper describes this · verbatim
Train and select occupancy models
Fitting model parameters, including fine-tuning an existing model. The AI stood in for expert judgement.
several candidate models were compared, and CatBoost was selected as the final classifier based on grouped cross‐validation performancewhere the paper describes this · verbatim
Predict sublattice and Fe-site occupancies
Running a trained model over new data to predict, classify or score. The AI stood in for expert judgement.
The first stage predicts whether the target dopant prefers the Ba/Sr sublattice or the Fe sublattice.where the paper describes this · verbatim
Assign reliability levels and refinement priority
Reducing a candidate set by filtering or ranking, in a single pass.
each prediction was assigned a reliability level and a refinement‐priority labelwhere the paper describes this · verbatim
Measure occupancies by neutron powder diffraction
Physical execution, by hand or by robot.
In this work, neutron powder diffraction was used to quantitatively resolve the crystallographic site occupancy of SrMg0.0498Sc1.7917Fe10.1598O19 at 300 K.where the paper describes this · verbatim
Validate predictions against refined occupancies
Testing outputs against ground truth. Its result feeds back into an earlier step.
Trained on 439 training records, this model correctly assigned 48 out of 50 external‐validation entries, yielding 96% classification accuracy.where the paper describes this · verbatim
Prospective screening of rare-earth dopants
Reducing a candidate set by filtering or ranking, in a single pass. The AI stood in for expert judgement.
Figure 6 summarizes pre‐refinement occupancy predictions for 96 host‐dopant combinationswhere the paper describes this · verbatim
What the record says
Technical · every part carries its own basis
+ in the paper~ our reading− not reported
How to read the quotations. A quotation shows where the paper describes something. It does not quote every value beside it: one passage locates a part of the work, and values without their own quotation are our reading of that passage.
The paper's reported result is the machine-learning framework itself and the occupancy hypotheses it produces; the sublattice assignments, fractional occupancies and refinement-priority labels exist only as model outputs
The candidate occupancy‐amount regressors included Huber regression, gradient boosting, XGBoost, LightGBM, and CatBoost.where the paper describes this · verbatim
In this work, neutron powder diffraction was used to quantitatively resolve the crystallographic site occupancy of SrMg0.0498Sc1.7917Fe10.1598O19 at 300 K.where the paper describes this · verbatim
The data that support the findings of this study are available from the corresponding author upon reasonable request.where the paper describes this · verbatim
What this paper did not report
Technical · absence is published deliberately
- CodeWhether the code is available is not stated.
- Trained model weightsWhether the trained model is available is not stated.
- ComputeThe hardware or time used is not stated.
- Version of CatBoostWhich version of the model was used is not stated.
- Version of XGBoostWhich version of the model was used is not stated.
- Version of LightGBMWhich version of the model was used is not stated.
- Version of Random ForestWhich version of the model was used is not stated.
- Version of ExtraTreesWhich version of the model was used is not stated.
- Version of Gradient BoostingWhich version of the model was used is not stated.
- Version of Histogram-Based Gradient BoostingWhich version of the model was used is not stated.
- Version of Logistic RegressionWhich version of the model was used is not stated.
- Version of Huber regressionWhich version of the model was used is not stated.
- Version of TabPFNWhich version of the model was used is not stated.
About this article
Record aix-00279, version 2, checked by a person on 2026-11-04. The record describes the paper; it does not assess whether the paper's findings are right. The paper is published under CC-BY; quotations are at most 25 words. How we work · Report an error
Cite this record: AIxSci (2026). Machine learning predicts where dopant atoms sit in hexaferrite crystals before refinement. Record aix-00279, version 2, checked 2026-11-04. https://aixsci.org/articles/aix-00279/