materials-chemistry/ai produced the result/Scientific Reports 2023 · v2
Machine learning picks a new metal mixture, which casting and X-rays confirm
Researchers trained five classifiers on 1200 experimentally made high-entropy alloys to predict which crystal structure a mixture would take. A random forest model predicted one new composition would be face-centred cubic; the alloy was then cast and X-rayed.
spectrum · one line per step, placed by what the step does · bright lines used AI
Phase prediction and experimental realisation of a new high entropy alloy using machine learning
Scientific Reports, 2023
doi:10.1038/s41598-023-31461-7 · record aix-00025 v2 · checked 2026-10-07
- AI was for
- Classification
- Model family
- Random forest, Support vector machine, Gradient-boosted trees, Clustering
- Checked by
- Experimental1 tested, 1 worked
- Code
- not reported
The finding the paper is about came from the AI.
What this research was about
Most familiar metals are one main element with small additions: iron with a little carbon, for instance. High-entropy alloys take the opposite approach, mixing several metals in roughly comparable amounts. The question that decides whether such a mixture is useful is how its atoms arrange themselves once it cools. They may settle into one simple repeating pattern, such as a face-centred cubic or a body-centred cubic arrangement, or into two such patterns side by side, or into a jumble of hard, brittle compounds called intermetallics. That arrangement, the phase, governs how the metal behaves. With so many elements and proportions to choose from, the possibilities vastly outrun what anyone can melt and measure.
Metallurgists have long summarised a candidate mixture with a handful of calculated numbers, worked out from the elements and their proportions: how much heat the mixing absorbs or releases, how much disorder it introduces, how unevenly sized the atoms are, how differently they hold onto electrons, and how many outer electrons they contribute on average. The researchers gathered 1200 alloy compositions from published experiments that had all been made by melting and casting, computed those five numbers for each, and set out to learn the link between the numbers and the phase that was actually observed.
Where AI came in
The learning was done by classifiers, programs that are shown examples already labelled with the right answer and learn to sort new cases into the same categories. Five kinds were trained from scratch on the five calculated numbers, with four possible labels: face-centred cubic, body-centred cubic, both together, or a mixture of intermetallic phases. The random forest, which pools the votes of many simple decision trees, reached an average test accuracy of 84%. A version retrained on data padded out with 192 artificially generated samples reached 92%, though the authors argue that per-phase predictions did not improve. The models were also run on five alloys kept outside training, whose phases had already been measured.
The models stood in for the melting and measuring that would otherwise be needed to find out a mixture's phase, and for simulation of the same question. The random forest was then given the five numbers for one composition not in the dataset, Ni25Cu18.75Fe25Co25Al6.25, and predicted a face-centred cubic structure at room temperature. That alloy was made by melting the pure metals together under argon, and X-ray diffraction, which reads a crystal's repeating pattern from how it scatters X-rays, showed peaks matching a face-centred cubic structure.
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 a dataset of 1200 high-entropy alloy compositions taken only from experimental studies that used melting and casting synthesis routes, described each alloy by five empirical parameters (mixing enthalpy, mixing entropy, atomic size difference, electronegativity difference and valence electron concentration), and trained five classifiers to assign alloys to FCC, BCC, FCC + BCC or mixture-of-intermetallic-phases classes. The random forest classifier reached an average test accuracy of 84%, a ROC-AUC score of 0.9649 and a tenfold cross-validation mean score of 0.9315; a random forest retrained on data augmented with 192 SMOTE-Tomek links samples reached an average accuracy of 92%, which the authors argue did not improve per-phase predictions judged by confusion matrices and AUC scores. A new composition, Ni25Cu18.75Fe25Co25Al6.25, was predicted by the random forest to be FCC, was then synthesised by vacuum arc melting, and its X-ray diffraction pattern showed peaks corresponding to an FCC structure.
How AI was used
Phase labels for 1200 literature-sourced alloy compositions were grouped into four classes and encoded numerically; five physical descriptors were computed from composition using published empirical formulae, outliers removed, missing values imputed with several imputers (simple, KNN, MICE) and features scaled with a robust scaler. K-nearest neighbours, support vector machine, decision tree, random forest and XGBoost classifiers were then fitted to these descriptors in their default settings using scikit-learn and XGBoost in Python, with five repeated runs averaged and performance read off accuracy, precision, recall, F1, ROC-AUC (one-vs-rest) and tenfold cross-validation; the random forest was additionally hyper-parameter tuned, and separately refitted on a dataset resampled with SMOTE-Tomek links for comparison. The trained classifiers were run on five alloy compositions held outside the training and test sets to compare predicted with reported phases, and the random forest was then run on the descriptors of a new composition to predict its phase before that alloy was synthesised by vacuum arc melting and characterised by X-ray diffraction.
The shape of the work
Structural · the record, drawn
no AI
Curate experimental HEA dataset restricted to melting and casting routes
Obtaining raw data, whether by measurement, download or retrieval.
This study extracted a dataset of 1200 unique compositions of HEAs experimentally synthesised from the melting and casting routeswhere the paper describes this · verbatim
no AI
Compute five empirical physical parameters as features
Encoding data into features, descriptors, embeddings or graphs.
The ∆Hmix for available HEAs in the dataset were calculated using Miedema’s rulewhere the paper describes this · verbatim
no AI
Encode labels, remove outliers, impute and scale features
Cleaning, filtering, normalising or labelling data already obtained.
outlier detection was performed to remove the outliers from the datasetwhere the paper describes this · verbatim
AI
Train and compare five base classifiers and a tuned RFC
Fitting model parameters, including fine-tuning an existing model.
To use a variety of available machine learning algorithms in their vanilla form (base models)where the paper describes this · verbatim
AI
Resample dataset with SMOTE-Tomek links and retrain RFC
Fitting model parameters, including fine-tuning an existing model.
we tried to resample our data using SMOTE-Tomek links method for V-RFC modelwhere the paper describes this · verbatim
AI
Predict phases of unseen literature alloys with all five models
Testing outputs against ground truth.
The predictive capability of all five classifiers was further tested for alloys that were not considered for training or testing the datasetwhere the paper describes this · verbatim
AI
Predict the phase of a new alloy composition with RFC
Running a trained model over new data to predict, classify or score. The AI stood in for simulation.
RFC algorithm indicated that this new HEA would stabilise as FCC phase at room temperaturewhere the paper describes this · verbatim
no AI
Synthesise and characterise the new HEA
Physical execution, by hand or by robot.
All elemental metals were melted together by vacuum arc melting under inert gas (high purity Ar) environmentwhere 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 reported new alloy composition was selected from the random forest classifier's phase prediction, which was then checked by XRD; the paper's central result is the model's prediction and its experimental realisation
we tested five robust algorithms namely, K-nearest neighbours (KNN), support vector machine (SVM), decision tree classifier (DTC), random forest classifier (RFC) and XGBoost (XGB)where the paper describes this · verbatim
The peaks from X-ray diffraction revealed an FCC structure in corroboration with the ML predictionswhere the paper describes this · verbatim
A detailed description of the complete dataset is provided as supplementary informationwhere 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 K-nearest neighbours classifier (V-KNN)Which version of the model was used is not stated.
- Version of Support vector machine (V-SVM)Which version of the model was used is not stated.
- Version of Decision tree classifier (V-DTC)Which version of the model was used is not stated.
- Version of Random forest classifier (V-RFC)Which version of the model was used is not stated.
- Version of XGBoost (V-XGB)Which version of the model was used is not stated.
- Version of Hyper-parameter tuned random forest classifier (HT-RFC)Which version of the model was used is not stated.
- Version of SMOTE-Tomek links augmented random forest classifier (ST-RFC)Which version of the model was used is not stated.
- What step 4 replacedThe paper gives no basis for what the AI stood in for.
- What step 5 replacedThe paper gives no basis for what the AI stood in for.
- What step 6 replacedThe paper gives no basis for what the AI stood in for.
About this article
Record aix-00025, version 2, checked by a person on 2026-10-07. The record describes the paper; it does not assess whether the paper's findings are right. The paper is published under CC-BY-4.0; quotations are at most 25 words. How we work · Report an error