materials-chemistry/ai produced the result/Materials 2026 · v2
Tree-based models predict sliding wear of carburised and plain engineering steels
Researchers measured how three engineering steels wore away in dry sliding tests, then trained decision tree, random forest and physics-informed neural network models to predict that wear and to rank which test conditions mattered most.
spectrum · one line per step, placed by what the step does · bright lines used AI
Explainable Machine Learning with Optimized Tree-Based Models and Statistical Validation for Data-Driven Wear Prediction of Carburized and Non-Carburized Engineering Steels
Materials, 2026
doi:10.3390/ma19194045 · record aix-00272 v2 · checked 2026-10-20
- AI was for
- Property prediction
- Model family
- Random forest, Multilayer perceptron
- Checked by
- Held-out
- Code
- not reported
The finding the paper is about came from the AI.
What this research was about
Metal parts that rub against each other, such as gears and shafts, slowly lose material. Engineers call this wear, and it decides how long a component lasts. Wear depends on many things at once: the alloy, how hard the surface is, how fast the surfaces slide, how hard they are pressed together and how long the rubbing goes on. One common way to harden a steel surface is carburising, which adds carbon to the outer layer in a furnace. Because all these factors interact, wear is awkward to predict from first principles, and the usual answer is to test each combination on a rig and weigh how much material is lost.
The researchers ran dry sliding tests on three steels, EN3, 20MnCr5 and EN353, each both as received and carburised, using a pin-on-disk machine that presses a pin against a rotating disc. They varied sliding speed, applied load and test duration across three levels each, covering every combination, and recorded the mass lost in each case. They then asked whether models fitted to those measurements could predict wear for conditions held back from training, and which inputs the models leaned on.
Where AI came in
The measurements were turned into a table of load, time, speed, material and surface condition, with an extra column multiplying load by time. Three models were fitted to part of that table and asked to predict mass loss for the rest: a decision tree, which splits the data by a series of yes-or-no questions; a random forest, which averages many such trees; and a physics-informed neural network, a network given an extra nudge during training to keep its predictions in line with the idea that wear grows with load multiplied by time. Settings were chosen by trying combinations with cross-validation.
Scored against the measured values, the random forest reached a coefficient of determination of 0.7244, with error figures of 0.00495 and 0.00273, and the decision tree 0.6406, with 0.00565 and 0.00275. The physics-informed network did far worse, at -282.62, which the authors put down to the small dataset and the simplified physics constraint. The models stood in here for further rig testing, predicting wear for conditions rather than measuring them. A method called SHAP, which apportions a prediction among the inputs, was then run on the tuned random forest; it placed surface condition and material type above sliding time and the load-by-time term, with load and speed lower.
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
Dry sliding wear of EN3, 20MnCr5 and EN353 steels, in as-received and carburized conditions, was measured in pin-on-disk tests under a full factorial design of three sliding speeds, three loads and three durations. Decision Tree and Random Forest regressors tuned by GridSearchCV with 5-fold cross-validation, and a physics-informed neural network with a load–time constraint, were fitted to these data to predict mass loss. On the test data the Random Forest gave RMSE 0.00495, MAE 0.00273 and a coefficient of determination of 0.7244, the Decision Tree gave RMSE 0.00565, MAE 0.00275 and 0.6406, and the PINN gave RMSE 0.15872, MAE 0.12599 and −282.62, which the authors attribute to the limited dataset size and simplified physics constraint. SHAP attribution ranked surface condition and material type above sliding time and the load × time interaction, with load and speed contributing less.
How AI was used
Wear values measured by the weight-loss method across the factorial test matrix were assembled into a tabular dataset of load, time, speed, material and surface condition, with an engineered load × time interaction term. Decision Tree and Random Forest regressors were fitted to this dataset, their hyperparameters selected by GridSearchCV with 5-fold cross-validation, yielding a maximum depth of 5 for the Decision Tree and 300 estimators with unrestricted depth for the Random Forest. A physics-informed neural network was also trained, with an added loss penalty soft-constraining normalised predicted mass loss towards the normalised product of applied load and time, following Archard's wear principles. The fitted models were run over a held-out split to predict wear, scored with RMSE, MAE and the coefficient of determination, and re-evaluated across train–test split ratios from 0.1 to 0.4; residual plots and 95% confidence intervals were computed. SHAP was applied to the optimized Random Forest to rank input parameters, alongside the model's built-in feature importances, partial dependence curves and a feature correlation matrix.
The shape of the work
Structural · the record, drawn
Run dry sliding pin-on-disk wear tests
Physical execution, by hand or by robot.
Dry sliding wear experiments were conducted using a pin-on-disk tribotester (Ducom TR-20 Neo Series) under ambient laboratory conditions.where the paper describes this · verbatim
Assemble dataset and engineer interaction feature
Encoding data into features, descriptors, embeddings or graphs.
Feature engineering included the creation of interaction terms such as load × time to capture combined effects.where the paper describes this · verbatim
Fit and tune DT, RF and PINN wear models
Fitting model parameters, including fine-tuning an existing model. The AI stood in for statistical model.
hyperparameter tuning was carried out using GridSearchCV in conjunction with 5-fold cross-validationwhere the paper describes this · verbatim
Predict wear on held-out data
Running a trained model over new data to predict, classify or score. The AI stood in for physical experiment.
the predictions made by the Decision Tree, Random Forest, and PINN models are contrasted in Figure 4where the paper describes this · verbatim
Score models and test split sensitivity
Testing outputs against ground truth. Its result feeds back into an earlier step.
A 70:30 train–test split was employed to evaluate the performance of the developed models.where the paper describes this · verbatim
Attribute wear to input parameters with SHAP
Extracting understanding from model behaviour. The AI stood in for expert judgement.
residual analysis, confidence interval estimation, and SHAP-based feature importance analysis were employed to assess model reliability and interpretabilitywhere 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 comparative predictive performance of the DT, RF and PINN wear models and the SHAP-derived parameter ranking; both are outputs of the models themselves. The underlying tribological experiments were conducted conventionally, but the claims the paper advances are about the models.
using Decision Tree (DT), Random Forest (RF), and a Physics-Informed Neural Network (PINN)where the paper describes this · verbatim
Figure 6 presents the residual plots of the Decision Tree, Random Forest, and PINN models using the independent test dataset.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.
- DataWhether the data are available is not stated.
- ComputeThe hardware or time used is not stated.
- How many were testedThe paper gives no count of what was tested.
- Version of Decision Tree regressorWhich version of the model was used is not stated.
- Version of Random Forest regressorWhich version of the model was used is not stated.
- Version of Physics-Informed Neural Network (load–time constraint)Which version of the model was used is not stated.
About this article
Record aix-00272, version 2, checked by a person on 2026-10-20. 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). Tree-based models predict sliding wear of carburised and plain engineering steels. Record aix-00272, version 2, checked 2026-10-20. https://aixsci.org/articles/aix-00272/