Optimized batch identification of metallic alloy materials for precision machining operations using machine learning and image processing techniques

ABSTRACT

In modern machining operations, real-time identification of metallic alloy materials is critical to maintaining high-quality standards, reducing errors, and enhancing operational efficiency. This study introduces a robust machine learning and image processing framework designed for the rapid and precise identification of material batches during machining processes. The framework evaluates multiple machine learning models, including Support Vector Machine (SVM), Random Forest (RF), Naive Bayes (NB), k-nearest Neighbors (k-NN), Artificial Neural Networks (ANN), Radial Basis Function Neural Networks (RBFNN), and Logistic Regression, to determine their suitability under varying cutting conditions with high and low data volumes. The models were benchmarked on accuracy, training time, and inference speed to identify the most effective solution for real-time applications. SVM emerged as the most accurate model, achieving a precision of 91.2% with an inference time of 0.8 milliseconds, making it ideal for real-time tasks. ANN exhibited a comparable accuracy of 89.3% but suffered from a significantly higher inference time of 210.2 milliseconds, limiting its real-time feasibility. RF demonstrated quick training times but incurred a relatively high inference delay. Naive Bayes offered the fastest training and inference times with an accuracy of 79.8%, suitable for scenarios prioritizing speed over precision. This comparative analysis provides valuable insights for selecting the optimal machine learning model tailored to specific manufacturing requirements, fostering a data-driven approach to optimizing machining operations.

Keywords:
Real-Time Material Identification; Machine Learning; Metallic Alloys; Machining Operations; Image Processing; Support Vector Machine

1. INTRODUCTION

In manufacturing industries, accurate identification of alloy material batches is fundamental to ensuring consistent quality and adherence to production standards [¹]. Real-time batch identification of alloy materials has gained critical importance, as it facilitates continuous monitoring and robust quality control during machining operations [²]. Variability in alloy composition and properties can significantly affect machining efficiency, product quality, and overall operational costs [³]. Traditional methods for identifying alloy materials, often reliant on manual inspection or outdated tracking systems, are not only error-prone but also incapable of meeting real-time operational demands [⁴]. Consequently, developing automated approaches for reliable and efficient batch identification of alloy materials is essential for modern manufacturing environments [⁴]. Machine learning (ML) and image processing techniques have emerged as promising solutions for real-time alloy material identification [⁵]. ML algorithms, trained on comprehensive datasets, can uncover complex patterns in material features, enabling quick and accurate decisions with minimal human intervention [⁶]. Image processing further complements this by allowing computers to analyze visual data, such as surface textures, colors, and microstructural attributes of alloy materials [⁷]. By integrating ML with image processing, the physical characteristics of alloy materials can be systematically extracted, analyzed, and classified with high precision and efficiency [⁸].

This integration aligns with the industry’s shift toward automated, data-driven quality control mechanisms [⁹]. Real-time identification of alloy materials addresses critical issues inherent in conventional batch identification processes [¹⁰]. Misidentification of alloy batches can result in production delays, deviations in quality, and increased scrap rates [¹¹]. Moreover, failing to detect batch discrepancies in real-time can lead to costly rework and material wastage, especially in high-precision sectors such as automotive and aerospace manufacturing [¹²]. Implementing an ML and image processing-based approach to alloy material identification not only resolves these challenges but also enhances operational efficiency and reduces error-related costs [¹³]. The primary objective of this study is to investigate the feasibility and effectiveness of various ML algorithms for real-time alloy material batch identification in machining environments [¹⁴]. This study evaluates a range of algorithms, including Support Vector Machines (SVM), Artificial Neural Networks (ANN), and Logistic Regression, to determine their suitability for diverse machining conditions and data volumes [¹⁵]. Through a detailed comparative analysis, the study aims to identify algorithms that strike a balance between accuracy, training time, and inference speed, ensuring their optimal deployment in real-time applications [¹⁶].

Additionally, the study examines how specific ML algorithms perform under high and low data volume scenarios, shedding light on their adaptability and scalability in real-world settings [¹⁷]. This study aims to develop and evaluate a machine learning and image processing framework for real-time identification of metallic alloy materials in machining operations, thereby enhancing quality control and operational efficiency in industrial manufacturing [¹⁸]. By rigorously evaluating model performance under various operational conditions, the study offers practical insights into integrating ML models into existing manufacturing workflows while considering computational demands [¹⁹]. Ultimately, this study seeks to empower manufacturing professionals to select suitable ML models for their unique production environments and requirements [²⁰]. It addresses a pressing need in the manufacturing sector for enhanced processes in alloy material identification [²¹]. The research contributes to the growing body of knowledge on ML applications in industrial automation, providing actionable insights to improve quality control while minimizing waste [²²].

In the past few years, transformer-based models have gained prominence in various domains, including material identification. For instance, the Generative Chemical Transformer (GCT) employs an attention mechanism to generate molecular structures that meet specific criteria, demonstrating the potential of transformers in material design [²²]. Similarly, the Molecule Attention Transformer (MAT) integrates inter-atomic distances and molecular graph structures into the attention mechanism, enhancing performance across diverse molecular prediction tasks [²³]. In the realm of machine vision, Vision Transformers (ViTs) have been introduced to address challenges in surface defect detection. A study on leather surface defect classification highlights the effectiveness of ViTs in anomaly detection and localization, even with low-resolution images and small datasets [²³]. Additionally, a novel detection transformer with a multi-scale fusion attention mechanism has been proposed for aero-engine turbine blade defect detection, demonstrating improved precision in multi-scale defect detection.

Image processing techniques are pivotal in modern machining operations, enabling real-time analysis and classification of metallic alloys [²³]. By extracting critical features such as texture, color, and surface patterns, these techniques facilitate accurate material identification, thereby enhancing quality control and operational efficiency [²⁴]. Integration of image processing with machine learning algorithms allows for automated decision-making, reducing reliance on manual inspection and minimizing errors [²⁵]. This synergy not only accelerates the manufacturing process but also ensures consistency and precision in machining tasks [²⁶].

As the industry increasingly adopts advanced technological solutions, the findings from this study hold the potential to enable more effective, data-driven decision-making in alloy material handling and processing [²⁷, ²⁸]. Future research could expand on this foundation by exploring the application of real-time material batch identification to a broader range of material types and machining conditions, while also enhancing the robustness of ML models through transfer learning and deep learning techniques [²⁹, ³⁰]. Moreover, integrating real-time feedback systems to adjust machining parameters dynamically based on identified alloy properties could further optimize operational efficiency [³¹]. Advancements in sensor technologies and high-resolution imaging systems are anticipated to improve feature extraction and boost identification accuracy [³², ³³]. This approach could also be extended to other industrial processes that require rapid material characterization, fostering broader adoption of intelligent, adaptive manufacturing systems across various sectors [³⁴].

2. MATERIALS AND METHODS

2.1. System design for real-time alloy material identification

The architecture of the real-time alloy material identification system integrates advanced machine learning models with a high-performance image processing unit [³⁵, ⁵³]. This framework includes an imaging sensor capable of capturing high-resolution details of alloy surfaces, a preprocessing module for image optimization, and a classification unit where machine learning models such as Support Vector Machine (SVM), Random Forest (RF), Naive Bayes (NB), k-Nearest Neighbours (k-NN), Artificial Neural Networks (ANN), Radial Basis Function Neural Networks (RBFNN), and Logistic Regression are employed [³⁶]. These components are interconnected to facilitate rapid data processing and minimize latency [³⁷]. The architecture is optimized for fast inference, ensuring efficient and accurate batch identification of alloy materials in real-time machining environments [³⁸].

2.2. Machining trials and data collection

Experiments were conducted to replicate various machining scenarios involving a range of alloy material batches with distinct compositions and surface properties [³⁹]. Each batch underwent machining under two conditions: high-volume cuts to simulate bulk processing and low-volume cuts to mimic precision machining [⁴⁰]. Standardized machining parameters, including feed rate, cutting speed, and depth of cut, were maintained across trials to ensure consistency [⁴¹]. These trials generated a comprehensive dataset of over 100,000 high-resolution images capturing variations in the texture, color, and microstructural features of alloy materials [⁴²]. A subset of these images was meticulously labeled to establish a ground truth dataset for training and validating the machine learning models [⁴³, ⁵³].

Data Scarcity vs. Model Complexity: ML models often require large datasets to achieve high accuracy. However, in materials science, data acquisition is labor-intensive and time-consuming, leading to small datasets that may not suffice for training complex models.

High Dimensionality vs. Limited Samples: Materials datasets frequently have high-dimensional feature spaces but limited sample sizes, which can result in overfitting and poor generalization of ML models.

Lack of Interpretability: Purely data-driven models may capture correlations without understanding underlying physical principles, making it challenging to interpret results and apply them to new materials systems.

Data collection is a foundational step in applying machine learning (ML) to materials science, directly influencing the performance and reliability of ML models. Effective data acquisition encompasses sourcing high-quality experimental and computational datasets, ensuring data consistency, and addressing issues like data sparsity and noise. Standardisation and normalisation are essential to harmonise diverse data types, such as crystallographic structures, electronic properties, and processing conditions. Additionally, integrating domain knowledge through feature engineering can enhance model interpretability and predictive accuracy.

2.2.1. Imaging hardware and environmental conditions

The high-resolution images utilized in this study were acquired using a Canon EOS 5D Mark IV DSLR camera equipped with a 50mm f/1.8 STM lens, offering a resolution of 30.4 megapixels. The camera was mounted on a Manfrotto 055XPRO3 tripod to ensure stability during image capture. Illumination was provided by a Softbox LED Light Kit with adjustable color temperature (3200K to 5600K) and brightness levels, ensuring uniform lighting across the sample surface. The lighting setup was positioned at a 45° angle to the sample surface to minimize glare and shadows. Additionally, a polarizing filter was used to reduce reflections and enhance surface detail visibility.

Environmental conditions were controlled within a temperature range of 22–24°C and a relative humidity of 45–50% to prevent condensation and ensure consistent imaging quality. The imaging sessions were conducted in a darkroom environment to eliminate ambient light interference. Image acquisition settings included an ISO of 100, aperture set to f/8, and a shutter speed of 1/125 seconds, providing optimal depth of field and sharpness. Each image was captured in RAW format to preserve maximum detail for subsequent processing.

Post-processing of the images was performed using Adobe Photoshop 2024, where adjustments were made to exposure, contrast, and sharpness to standardize image quality across the dataset. No alterations were made to the intrinsic features of the samples to maintain data integrity.

Initial Model Training: The process begins with training a machine learning model using an initial dataset of materials properties.

Prediction and Error Analysis: The trained model is then used to predict properties for a separate validation set. The discrepancies between predicted and actual values are analysed to identify systematic errors.

Model Refinement: Based on the error analysis, adjustments are made to the model. This may involve tuning hyperparameters, incorporating additional features, or employing different algorithms to better capture the underlying relationships.

Iterative Cycle: Steps 2 and 3 are repeated iteratively, with each cycle aimed at reducing prediction errors and improving model accuracy.

This iterative approach is exemplified in the study “Interpretable machine learning approach for exploring process-structure-property relationships in metal additive manufacturing”, where an error-targeted method was employed to optimize the laser powder bed fusion process for AlSi10Mg alloys. The methodology significantly reduced the amount of experimentation required, demonstrating the effectiveness of iterative refinement in predictive modelling.

2.3. Image preprocessing and feature extraction

The preprocessing of image data involved several critical stages, including noise reduction, normalization, and feature extraction [⁴⁴]. To eliminate background noise, each image underwent Gaussian filtering, enhancing the clarity of material features [⁴⁵]. Histogram equalization was applied to improve contrast and highlight subtle details in alloy textures and structures. Feature extraction techniques, such as the Gray Level Co-occurrence Matrix (GLCM) for texture analysis, colour histograms for surface colour profiling, and morphological operations for structural details, were employed to identify distinguishing characteristics of alloy materials [⁴⁶]. The extracted features were transformed into numerical vectors X = {x1, x2,…,xn}, where xi represents individual image features, serving as inputs to the machine learning models [⁴⁷, ⁴⁸]. This comprehensive preprocessing ensured the models could effectively distinguish between alloy batches based on their unique physical and structural attributes [⁴⁹].

2.4. Classification models and mathematical formulations

The study evaluated several machine learning models, each optimized with hyperparameter tuning to achieve the best results [⁵⁰, ⁵¹, ⁵³]. Key mathematical models included:

1. Support Vector Machine (SVM): The SVM aims to find a hyperplane H : wT x + b = 0 that maximizes the margin between two classes, where www represents the weight vector and b is the bias term. The optimal hyperplane minimizes:

   (1)$$min\frac{1}{2}\Vert {\text{w}}^{\text{2}}\Vert +C{\displaystyle {\sum }_{i=1}^N\epsilon i}$$

   subject to yi (wT xi + b ) ≥ 1 − ξi, with ξi ≥ 0.

2. Random Forest (RF): This ensemble model consists of multiple decision trees, where each tree Tj produces a class prediction hj(X). The final prediction H(X) is obtained by majority voting:

   (2)$$H(X)=mode\{h1(X),h2(X),\mathrm{...},hT(X)\}$$

3. Naive Bayes (NB): NB computes the posterior probability P(Ck∣X) for each class Ck based on feature likelihoods, given by:

   (3)$$P(Ck|X)=\frac{P(X|Ck)P(Ck)}{P(X)}$$

4. k-Nearest Neighbours (k-NN): For a given test instance, k-NN identifies the k closest neighbours based on Euclidean distance:

   (4)$$d(X,Y)=\sqrt{{\displaystyle {\sum }_{i=1}^n{(xi-yi)}^2}}$$

   The predicted class is the majority class among the neighbours.

5. Artificial Neural Networks (ANN): The ANN model consists of multiple layers of neurons, with each neuron activation calculated as:

   (5)$$zj=\sigma \left({\displaystyle {\sum }_{i=1}^{n}wijxi+bj}\right)$$

   Where wij represents the weights, bj the biases, and σ the activation function, typically ReLU or sigmoid.

6. Radial Basis Function Neural Networks (RBFNN): RBFNN employs radial basis functions as activation functions. For an input vector XXX, the output is:

   (6)$$h(X)={\displaystyle {\sum }_{i=1}^n\text{w}}\varphi (\Vert X-Cj\Vert ))$$

   where ϕ(⋅) is a Gaussian function cantered at ci.

7. Logistic Regression (LR): LR models the probability of the binary outcome as:

   (7)$$P(y=1|X)=\frac{1}{1+e^{-}(\beta 0+\beta 1x1+\mathrm{...}+\beta nXn)}$$

   The model parameters β are optimized to maximize the likelihood of observed outcomes.

2.5. Definition of data volume conditions

The classification models were trained and evaluated under two specific data volume scenarios tailored to the identification of alloy materials:

2.5.1. High-data-volume scenarios

In these conditions, the models were provided with the complete dataset, encompassing a wide diversity of alloy material features such as texture, color, and structural characteristics [⁵⁰, ⁵¹]. This approach ensured the inclusion of extensive variations in alloy composition and machining-induced surface modifications, maximizing the models’ ability to generalize across different batches.

2.5.2. Low-data-volume scenarios

To simulate resource-constrained environments, the dataset was deliberately reduced in size while maintaining a representative subset of alloy material features [²⁶]. This condition tested the models’ adaptability and robustness when trained on limited data, which is critical in scenarios where data acquisition may be expensive or time-restrictive. The models’ performance was evaluated in both scenarios based on three key metrics: Accuracy: The percentage of correctly identified alloy batches. Training Time: The computational time required to train the models on the respective datasets. Inference Time: The time taken to classify an alloy material batch in real-time. Comparative analysis of these metrics across high- and low-data-volume conditions provided insights into the models’ scalability, effectiveness, and suitability for practical deployment in real-time alloy material batch identification systems [¹⁹]. This evaluation highlights the trade-offs between model accuracy and computational efficiency, enabling informed decisions for implementation in machining operations.

Table 1 summarizes the experimental setup for alloy material batch identification, detailing the various cutting parameter combinations employed to evaluate the performance of the machine learning models [¹⁵]. Each condition type (X, Y, Z, A, B, C) represents distinct combinations of cutting speed and feed rate tailored to machining operations on metallic alloy materials. These conditions were designed to assess the influence of tool wear, surface integrity, and machining precision on the accuracy of material batch identification [¹⁰]. Cutting speeds were set between 150 and 320 meters per minute, and feed rates ranged from 0.6 to 0.8 mm per revolution to replicate real-world machining conditions. Tool wear measurements were taken in micrometers, with recorded values spanning from 150 µm to 240 µm, incorporating a specified variation (± µm) to account for deviations caused by machining-induced thermal and mechanical stresses [²³]. The experimental conditions enabled the evaluation of key performance indicators, such as: The impact of alloy material properties on surface feature detectability.

Table 1
Summary of experimental setup with different cutting parameter combinations.

CONDITION CATEGORY	CUTTING SPEED (m/min)	FEED RATE (mm/REV)	WEAR MEASUREMENT (µM)	ALLOY GROUPS	EXPERIMENT COUNT	SAMPLE SIZE
X	150	0.6	220 µM ± 80 µM	ALLOY X, Y, Z	55	2020
Y	210	0.8	180 µM ± 90 µM	ALLOY X, Y, A	22	2760
Z	200	0.6	190 µM ± 70 µM	ALLOY B, C, D	78	3090
A	250	0.8	150 µM ± 110 µM	ALLOY C, D, E	65	4380
B	280	0.8	240 µM ± 120 µM	ALLOY D, E, F	90	5500
C	320	0.6	170 µM ± 100 µM	ALLOY A, B, C	72	4000

The relationship between tool wear and material classification accuracy. The adaptability of machine learning models under varying cutting parameter conditions. This setup provides a robust framework for understanding how machining parameters and material properties influence the real-time identification of alloy materials in industrial applications. Different alloy material batch groups (e.g., X, Y, Z, A, B, C) were processed under the defined machining conditions, with the number of experiments per group ranging from 22 to 90 [²⁰]. The total sample size per condition varied between 2020 and 5500 samples, ensuring a diverse and comprehensive dataset for evaluating the performance of machine learning models in identifying alloy materials under different machining conditions and tool wear rates [¹⁸]. Figure 1 illustrates the overall system architecture for real-time alloy material batch identification [¹¹]. It showcases the key components, including the imaging sensor, data pre-processing unit, feature extraction module, and the suite of classification algorithms [¹⁶].

Figure 1
System design for real-time identification and basic functional concepts.

The figure outlines the data flow from image capture through pre-processing, feature extraction, model application, and final identification, highlighting the seamless integration of machine learning and image processing techniques for identifying metallic alloys in machining operations [²⁰]. Figure 2 presents the training and evaluation workflow, where each point of interest (POI)—such as material texture, surface finish, and machining-induced changes—is analyzed separately [²²]. Cross-validation is performed on an experiment-by-experiment basis, enabling the models to generalize across diverse machining conditions and alloy properties [¹⁸]. The figure underscores the iterative nature of model development, ensuring robustness and accuracy for each specific alloy material batch [¹⁰]. The methodology flow chart (Figure 3) offers a detailed, step-by-step visual overview of the research process, from data collection and pre-processing to model training, validation, and evaluation [²].

Figure 2
The working out and assessment process examined each point of interest (POI) individually, with cross-validation performed at the level of each experiment.

Figure 3
The detailed methodology flow chart for this alloy material study.

Each phase is delineated to showcase key activities such as parameter tuning, feature selection, model application, and performance metric collection. The chart provides a comprehensive roadmap of the study’s workflow, ensuring clarity in the approach to real-time alloy material batch identification [¹⁶]. Figure 4 outlines the essential functions of different machine learning models used in the study, including Support Vector Machine, Random Forest, k-Nearest Neighbors, Naive Bayes, Artificial Neural Networks, and Logistic Regression [¹⁹]. Each classification model’s unique function is depicted, providing an understanding of how they contribute to material identification through statistical learning, pattern recognition, and predictive analytics [¹⁰].

Figure 4
Basic functionality of machine learning category.

Torque is a measure of the rotational force applied to an object, causing it to rotate around an axis. It is calculated as the product of the force applied and the distance from the axis of rotation, typically expressed in newton-meters (N·m). In machining operations, torque is a critical parameter that influences the cutting force, tool wear, and overall efficiency of the process. Accurate torque measurements are essential for monitoring tool performance and ensuring the quality of the machined components.

The sample size for each experimental condition was determined by multiplying the number of alloy groups involved by the number of measurements taken per alloy group. This approach ensures that each alloy group’s performance is adequately represented under the specified cutting conditions.

For instance, in Condition X, where alloys X, Y, and Z were tested, the sample size was calculated as follows:

(8) $$\text{Sample Size = Number of Alloy Groups × Number of Measurements per Alloy Group}$$

(9) $$\text{Sample Size = 3 alloys × 20 measurements/alloy = 60 measurements}$$

This methodology was consistently applied across all experimental conditions to maintain uniformity and statistical robustness.

To generate the confusion matrix comparing predicted and true labels, we utilized the confusion_matrix function from the scikit-learn library in Python. This function computes the confusion matrix by comparing the true labels (y_true) with the predicted labels (y_pred) generated by our machine learning model.

For visualization, we employed ConfusionMatrixDisplay.from_predictions, also from scikit-learn, which provides a clear graphical representation of the confusion matrix, facilitating the interpretation of model performance. This method is widely used in machine learning to assess classification accuracy and identify areas where the model may need improvement.

In this study, Gray Level Co-occurrence Matrix (GLCM) and colour histograms were selected for feature extraction due to their proven effectiveness in capturing essential textural and colour information from images, which are critical for material identification tasks. GLCM analyses spatial relationships between pixel intensities, providing statistical measures such as contrast, homogeneity, energy, and entropy, which are sensitive to variations in surface texture and are widely used in industrial applications. Colour histograms offer a straightforward representation of color distributions within an image, aiding in distinguishing materials with different colour characteristics.

While alternative methods like Local Binary Patterns (LBP) and wavelet transforms also offer valuable features, they were not selected in this study. LBP is effective for capturing micro-textural patterns but may be less robust to noise and variations in lighting conditions. Wavelet transforms provide multi-scale analysis but can be computationally intensive and may require more complex parameter tuning. The choice of GLCM and colour histograms balances computational efficiency with the ability to capture essential features for accurate material classification in machining operations.

3. RESULTS AND DISCUSSION

3.1. Analysis of alloy machinability

The machinability assessment concentrated on analyzing the behavior of various alloy materials under differing cutting conditions and their impact on model predictions [¹¹]. Factors such as cutting speed, feed rate, and wear characteristics were critical in determining machinability. Alloys subjected to higher cutting speeds and feed rates exhibited increased tool wear, which affected the classification accuracy of the models [¹⁸]. Alloy groups with lower wear values, such as Alloy A and C, demonstrated improved precision in identification, suggesting that controlled machining parameters contribute to better classification performance [¹⁹]. These results underscore the significance of understanding alloy machinability to refine model performance in real-time alloy classification systems [²⁰].

3.2. Refinement and calibration of models

The machine learning models were refined through hyperparameter tuning to enhance prediction accuracy for alloy material identification. Adjustments included kernel type for SVM, tree depth for RF, and hidden layer configurations for ANN [¹⁰]. The optimization process led to an accuracy improvement of 4–7% across the models, with SVM and ANN showing the highest sensitivity to parameter changes [¹⁴]. For instance, the polynomial kernel with a degree of 3 significantly enhanced SVM’s classification capability, resulting in an accuracy of 91.2% [¹¹]. These findings highlight the importance of tailored calibration to accommodate variations in alloy properties and machining conditions, ensuring robust real-time identification [²⁰, ⁵³].

3.3. Predictive accuracy assessment

The performance of the machine learning models was evaluated under varying data volumes using metrics such as accuracy and robustness. In high-data-volume scenarios, SVM outperformed other models, achieving an accuracy of 91.2%, followed by k-NN and RF with accuracies of 86.3% and 84.5%, respectively [²⁰]. In resource-constrained, low-data-volume conditions, Naive Bayes emerged as a resilient choice with reasonable performance [³⁷]. These results highlight the necessity of selecting appropriate models based on data availability and operational requirements for efficient alloy material classification [³⁶].

Table 2 presents the analysed and optimal hyperparameters for each machine learning model used in the study [¹⁵]. Hyperparameter tuning is a critical process for optimizing model performance, and this table summarizes the values explored and the best configurations for each algorithm. For the SVM, the regularization parameter (C) was varied across a wide range of values, with the optimal value being 5.00E + 03 [¹⁹]. The kernel type, which determines the decision boundary, was found to perform best with a polynomial kernel, with a degree of 3 [²⁰]. The gamma parameter, which influences the shape of the decision boundary, was optimized at 1e−1, contributing to the model’s high performance. In the case of Random Forest (RF), several hyperparameters were adjusted to improve model accuracy [⁹].

Table 2
Analysed and ideal hyperparameters for each model.

MODEL	PARAMETER	EXAMINED VALUES	OPTIMAL VALUE
SVM	Regularization (C)	1e2, 5e1, 1, 1e−1, 1e−2, 5e3, 5e4, 1e5	5.00E + 03
	Kernel Type	Linear, Polynomial, Gaussian	Polynomial
	Degree	2, 3, 4, 5, 6	3
	Gamma	1e−2, 1e−1, 1, 1e1, 1e2	1e−1
Random Forest (RF)	Number of Trees	20, 50, 100, 150, 300, 600	150
	Splitting Criterion	Gini, Entropy	Entropy
	Max Depth	3, 6, 9, 12	6
	Max Features	3, 5, 6, 7	6
Naive Bayes (NB)	Smoothing Factor	1e−5, 1e−4, 1e−3	1e−4
Artificial Neural Networks (ANN)	Neighbors	3, 7, 15, 29, 60, 120	60
	Weighting Scheme	Uniform, Inverse Distance	Inverse Distance
	Leaf Size	10, 20, 30, 40	20
	Power Parameter (P)	1, 2, 4	2
	Distance Metric	Manhattan, Cosine, Minkowski	Cosine
Artificial Neural Networks (ANN)	Neurons per Layer	8, 16, 32, 64, 128, 256	128
	Hidden Layers	1, 2, 3, 5	3
	Dropout Rate	0.0, 0.3, 0.5	0.3
	Optimizer	RMSprop, Adam, Adagrad	RMSprop
Radial Basis Function Neural Network (RBFNN)	Neurons	8, 16, 32, 64, 128, 256, 512	128
	Beta	0.02, 0.1, 0.3, 1.5, 2.5, 4.0	2.5
	Initialization	Random, K-means	Random

The optimal number of trees was set at 150, which strikes a balance between model complexity and computational efficiency [¹⁰]. The splitting criterion, which determines how data is partitioned, was best set to ‘Entropy,’ and the maximum depth of the trees was optimized at 6 to avoid overfitting [¹⁷]. The model performed best when using 6 features for splitting at each node [¹⁹]. For Naive Bayes (NB), the smoothing factor was optimized at 1e−4, which is essential to prevent zero probabilities for unseen data during classification. For k-Nearest Neighbours (kNN), the number of neighbours was optimized at 60, with an inverse distance weighting scheme for better handling of nearby neighbours [¹⁵].

For Random Forest (RF), the optimal number of leaves per tree was determined to be 20, ensuring an effective balance between model complexity and generalization. The distance metric for splitting data during tree construction was set to Cosine similarity, which improved the model’s ability to capture relationships between alloy characteristics [¹³]. In ANN, the optimization involved configuring 128 neurons per layer and 3 hidden layers. A dropout rate of 0.3 was selected to mitigate overfitting, while the RMSprop optimizer was used due to its ability to efficiently update weights and enhance training performance [²¹]. For Radial Basis Function Neural Networks (RBFNN), the optimal configuration included 128 neurons and a beta parameter of 2.5, ensuring the model could effectively capture the underlying patterns in alloy properties. Random weight initialization further supported the model’s robustness in identifying intricate material characteristics [²⁹]. Table 2 provides a comprehensive summary of the hyperparameter optimization for each model, critical for achieving high performance in alloy material identification tasks [³⁰].

3.4. Computational efficiency and resource analysis

The computational efficiency of each model was assessed by measuring training and inference times, with a focus on their potential for real-time application in alloy classification [¹²]. Naive Bayes exhibited the shortest training and inference times, making it an excellent choice for low-latency environments. Random Forest required moderate computational resources, offering a balanced trade-off between accuracy and speed. While ANN and RBFNN achieved higher accuracy, their complexity led to increased computational time, which may limit their real-time applicability in resource-constrained settings [¹⁸]. This balance between model performance and computational efficiency is crucial, and SVM and Naive Bayes emerged as the most balanced models for real-time alloy material classification, delivering both robust results and efficient performance [¹¹].

Table 3 presents a performance comparison of various machine learning models used for alloy material batch identification under both high and low information capacity cutting conditions [⁸]. The table evaluates the performance of seven machine learning models: SVM, RF, NB, k-NN, ANN, Radial Basis Function Neural Networks (RBFNN), and Logistic Regression (LR) [¹³]. The models are compared based on their accuracy, training time, and inference time. Among the models, SVM achieved the highest accuracy at 91.2%, followed by ANN at 89.3% [¹⁸]. The table also includes training and inference times, reflecting the computational efficiency of each model. Naive Bayes was noted for its extremely fast training time of 0.03 seconds, making it ideal for low-data conditions, while more complex models like ANN and RBFNN exhibited longer training and inference times due to their complexity [²⁹]. This comparison illustrates the trade-off between accuracy and computational efficiency, helping to identify the most appropriate model for real-time machining operations, depending on data volume and operational requirements [¹].

Table 3
Performance comparison of various models for high and little information capacity wounding circumstances.

SI.NO	MODEL	ACCURACY (%)	TRAINING TIME (S)	INFERENCE TIME (MS)
1	Support Vector Machine	91.2	72.4	0.8
2	Random Forest	84.5	4.1	58.3
3	Naive Bayes	79.8	0.03	0.7
4	k-Nearest Neighbors	86.3	0.2	3.1
5	Artificial Neural Networks	89.3	19.5	210.2
6	Radial Basis Function Neural Networks	78.4	5.9	258.7
7	Logistic Regression	83.5	0.5	0.5

Figure 5 illustrates the machinability evaluation of different batches, emphasizing variations in tool insert lifespan even under similar cutting speeds. This highlights the importance of material properties in machining performance. Figure 6 demonstrates how increasing the tolerance for data inclusion affects prediction accuracy, comparing results under high and low data volume conditions. Figure 7 showcases a confusion matrix that visually represents the accuracy of the batch classification process. Finally, Figure 8 shows a significant improvement in accuracy to over 98.23% when assuming that Batches A, B, C, D, and E share similar machinability characteristics, underscoring the importance of material consistency in improving prediction accuracy.

Figure 5
The machinability of the verified substantial consignments was evaluated, revealing variations in insert lifespan despite similar cutting speeds.

Figure 6
The effect of expanding the open-mindedness for incorporating information into the model working out on the calculation accurateness for cutting condition combinations with high information capacity and little information capacity was examined.

Figure 7
A misunderstanding matrix illustrating the calculation accurateness of the alloy batch material documentation algorithm is provided.

Figure 8
Assuming that Batches A, B, C, D, and E exhibit similar machinability characteristics, the accurateness of the documentation algorithm improves to over 98.23%.

3.5. Comparative discussion on model performance

The comparative analysis reveals that no single machine learning model excels across all performance metrics, as each model demonstrates distinct strengths based on accuracy, speed, and data handling capacity [¹⁰]. The SVM model, with its high accuracy under high-data conditions, is particularly suitable for tasks requiring high precision, such as in quality-critical machining operations. On the other hand, Naive Bayes and Random Forest stand out for their faster computational times, making them better suited for real-time and moderate-accuracy applications [¹³]. This highlights the importance of selecting the right model based on the specific operational context, including data volume and required accuracy, to achieve the best performance in industrial machining applications [¹⁶].

The results of this study confirm that SVM, with its 91.2% accuracy, outperformed other models like Random Forest and k-Nearest Neighbors in alloy material batch identification. This makes SVM a reliable choice for real-time machining operations where high data volumes are involved. These findings are consistent with previous research, which supports the effectiveness of SVM in precision tasks, particularly in industrial applications [²]. Additionally, the Naive Bayes and Random Forest models demonstrated shorter training and inference times, making them suitable for low-latency applications, an important consideration for real-time implementations [¹⁶].

Furthermore, the assumption that Batches A, B, C, D, and E share similar machinability characteristics led to a substantial increase in prediction accuracy to 98.23%, highlighting that batch homogeneity plays a crucial role in improving model performance. This is in line with prior studies that emphasize the importance of material consistency in enhancing machine learning-based predictive systems in manufacturing [¹⁹]. These findings reinforce the notion that optimizing batch classification can lead to significant improvements in process efficiency and predictive accuracy. Future research could explore the impact of additional cutting conditions and material properties on model performance, as well as explore new computational methods for achieving faster real-time predictions [²³].

Validate the accuracy metrics reported for the machine learning models, statistical hypothesis testing was conducted to assess whether the observed differences in performance are statistically significant. The 5×2 cross-validation procedure, as proposed by DIETTERICH [⁵²], was employed to mitigate the optimistic bias associated with traditional k-fold cross-validation methods. This approach involves performing two-fold cross-validation five times and then applying a paired Student’s t-test to compare the performance of different models shows in Table 4. A p-value threshold of 0.05 was used to determine statistical significance. A p-value less than or equal to 0.05 indicates that the difference in model performance is statistically significant, suggesting that the observed difference is unlikely to be due to random chance. Conversely, a p-value greater than 0.05 suggests that the difference is not statistically significant [⁵³].

Table 4
Results of the hypothesis tests for each pair of models.

MODEL COMPARISON	P-VALUE	T-STATISTIC	SIGNIFICANCE
SVM vs. Random Forest	0.003	−5.318	Significant
SVM vs. k-NN	0.131	−1.802	Not Significant
SVM vs. Naive Bayes	0.368	0.988	Not Significant
SVM vs. ANN	0.003	−5.191	Significant
SVM vs. RBFNN	0.028	−3.062	Significant
SVM vs. Logistic Regression	0.003	−5.318	Significant

These statistical analyses provide a more robust validation of the model performance metrics and offer confidence in the comparative effectiveness of the machine learning models employed in this study.

4. CONCLUSION

This study demonstrates the efficacy of machine learning algorithms, particularly Support Vector Machine (SVM), in real-time alloy material batch identification during machining operations. By evaluating various models—including RF, k-NN, NB, ANN, RBFNN, and LR—we identified that SVM offers the highest accuracy at 91.2%, while maintaining reasonable computational efficiency. The study also highlights the trade-offs between accuracy and computation time, with models such as Naive Bayes and Random Forest providing faster processing times but with slightly lower prediction accuracy, making them more suitable for applications requiring low latency. The impact of material machinability on model performance was also significant. The assumption that Batches A, B, C, D, and E have similar machinability properties resulted in a remarkable increase in prediction accuracy to over 98.23%, emphasizing the role of material homogeneity in improving the reliability of batch identification algorithms. Furthermore, expanding the tolerance for data inclusion during model training proved beneficial, demonstrating that a broader range of data leads to enhanced prediction accuracy, both in high and low data volume conditions. The confusion matrix provided in the study validated the model’s predictive capabilities, showcasing its ability to accurately classify batches under varying cutting conditions.

The findings of this research have significant implications for instantaneous one-to-one care and decision-making in machining processes. This approach could be particularly beneficial for industries like automotive, aerospace, and manufacturing, where precision and efficiency are paramount. Future research could focus on exploring more progressive machine learning techniques, including deep learning models, to further enhance the accuracy and efficiency of batch identification in complex machining environments. Additionally, expanding the study to include a wider variety of materials and cutting conditions would refine the model for broader industrial applications, additional optimizing performance across diverse machining operations. Integrating sensor data such as cutting force, vibration, and temperature could provide a more comprehensive understanding of the machining process, leading to more robust and adaptable models.

To ensure the successful implementation and scalability of the future system, the following management strategies are recommended:

Pilot Testing: Conduct pilot tests in collaboration with industry partners to authenticate the model’s presentation in real-world machining scenarios.

Data Integration: Develop infrastructure for seamless integration of sensor data into the machine learning model, ensuring real-time data acquisition and processing.

Model Optimization: Implement continuous learning mechanisms to update and refine the model based on new data and feedback from machining operations.

Training and Support: Provide training programs for operators and engineers to effectively utilize the system and interpret its outputs.

Collaboration and Feedback: Establish feedback loops with stakeholders to gather insights and drive iterative improvements in the system’s design and functionality.

By adopting these management strategies, the proposed machine learning-based alloy material batch identification system can be effectively deployed, leading to enhanced precision and efficiency in machining operations.

5. ACKNOWLEDGMENTS

Princess Nourah bint Abdulrahman University Researchers Supporting Project number (PNURSP2025R237), Princess Nourah bint Abdulrahman University, Riyadh, Saudi Arabia. Ongoing Research Funding program (ORF-2025-608), King Saud University, Riyadh, Saudi Arabia.

6. BIBLIOGRAPHY

• ^[1] LIN, H., LI, B., WANG, X., et al., “Automated defect inspection of LED chip using deep convolutional neural network”, Journal of Intelligent Manufacturing, v. 30, pp. 2525–2534, 2019. doi: http://doi.org/10.1007/s10845-018-1415-x.
  » https://doi.org/10.1007/s10845-018-1415-x
• ^[2] MITTAL, S., KHAN, M.A., ROMERO, D., et al., “Smart manufacturing: characteristics, technologies and enabling factors”, Proceedings of the Institution of Mechanical Engineers. Part B, Journal of Engineering Manufacture, v. 233, n. 5, pp. 1342–1361, 2019. doi: http://doi.org/10.1177/0954405417736547.
  » https://doi.org/10.1177/0954405417736547
• ^[3] ABADI, M., BARHAM, P., CHEN, Z., et al., “TensorFlow: a system for large-scale machine learning”, In: Proceedings of the 12th Symposium on Operating Systems Design and Implementation, pp. 265–283, Savannah, GA, USA, 2016.
• ^[4] SIMONYAN, K., ZISSERMAN, A., “Very deep convolutional networks for large-scale image recognition”, In: Proceedings of the International Conference on Learning Representations, pp. 7–9, 2015.
• ^[5] DENKENA, B., BERGMANN, B., WITT, M., “Automatic process parameter adaption for a hybrid workpiece during cylindrical operations”, International Journal of Advanced Manufacturing Technology, v. 95, n. 1–4, pp. 311–316, 2018. doi: http://doi.org/10.1007/s00170-017-1196-y.
  » https://doi.org/10.1007/s00170-017-1196-y
• ^[6] KITA, Y., ISHIKAWA, H., MASUDA, T., “Guest editorial: machine vision applications”, International Journal of Computer Vision, v. 122, n. 2, pp. 191–192, 2017. doi: http://doi.org/10.1007/s11263-017-0990-1.
  » https://doi.org/10.1007/s11263-017-0990-1
• ^[7] DENKENA, B., BERGMANN, B., WITT, M., “Material identification based on machine-learning algorithms for hybrid workpieces during cylindrical operations”, Journal of Intelligent Manufacturing, v. 30, n. 6, pp. 2449–2456, 2019. doi: http://doi.org/10.1007/s10845-018-1404-0.
  » https://doi.org/10.1007/s10845-018-1404-0
• ^[8] GOPPOLD, C., URLAU, F., PINDER, T., et al., “Experimental investigation of cutting performance for different material compositions of Cr/Ni-steel with 1 µm laser radiation”, Journal of Laser Applications, v. 30, n. 3, pp. 031501, 2018. doi: http://doi.org/10.2351/1.5013284.
  » https://doi.org/10.2351/1.5013284
• ^[9] KUO, C.-F.J., HUANG, Y.-J., SU, T.-L., et al., “Computerized color distinguishing system for color printed fabric by using the approach of probabilistic neural network”, Polymer-Plastics Technology and Engineering, v. 47, n. 3, pp. 264–272, 2008. doi: http://doi.org/10.1080/03602550701866808.
  » https://doi.org/10.1080/03602550701866808
• ^[10] HERNÁNDEZ-ORALLO, J., DOWE, D.L., HERNÁNDEZ-LLOREDA, M.V., Measuring cognitive abilities of machines, humans and non-human animals in a unified way: towards Universal Psychometrics, Technical report, Monash University, Mar. 2012.
• ^[11] JEMIELNIAK, K., KOSMOL, J., Tool and process monitoring state of art and future prospects, Institute of Mechanical Engineering and Automation, Technical University of Wroclaw, Scientific Papers, v. 61, pp. 90–112, 1995.
• ^[12] FRITZ, M., HAYMAN, E., CAPUTO, B., et al., The KTH-TIPS database, 2019. http://www.nada.kth.se/cvap/databases/kth-tips/download.html, accessed in March, 2019.
  » http://www.nada.kth.se/cvap/databases/kth-tips/download.html
• ^[13] KOTHURU, A., NOOKA, S.P., LIU, R., “Audio-based tool condition monitoring in milling of the workpiece material with the hardness variation using support vector machines and convolutional neural networks”, Journal of Manufacturing Science and Engineering, v. 140, n. 11, pp. 111006, 2018. doi: http://doi.org/10.1115/1.4040874.
  » https://doi.org/10.1115/1.4040874
• ^[14] KRAMER, N., “In-process identification of material properties by acoustic emission signals”, CIRP Annals, v. 56, n. 1, pp. 331–334, 2007. doi: http://doi.org/10.1016/j.cirp.2007.05.076.
  » https://doi.org/10.1016/j.cirp.2007.05.076
• ^[15] LUTZ, B., KISSKALT, D., REGULIN, D., et al., “Evaluation of deep learning for semantic image segmentation in tool condition monitoring”, In: Proceedings of the 2019 18th IEEE International Conference on Machine Learning and Applications (ICMLA), pp. 2008–2013, Boca Raton, FL, USA, 2019, doi: http://doi.org/10.1109/ICMLA.2019.00321.
  » https://doi.org/10.1109/ICMLA.2019.00321
• ^[16] MAYR, A., KIßKALT, D., MEINERS, M., et al., “Machine learning in production: potentials, challenges and exemplary applications”, Procedia CIRP, v. 86, pp. 49–54, 2019. doi: http://doi.org/10.1016/j.procir.2020.01.035.
  » https://doi.org/10.1016/j.procir.2020.01.035
• ^[17] PENUMURU, D.P., MUTHUSWAMY, S., KARUMBU, P., “Identification and classification of materials using machine vision and machine learning in the context of Industry 4.0”, Journal of Intelligent Manufacturing, v. 31, n. 5, pp. 1229–1241, 2020. doi: http://doi.org/10.1007/s10845-019-01508-6.
  » https://doi.org/10.1007/s10845-019-01508-6
• ^[18] SCHNEIDER, G., “Machinability of metals”, In: Schneider, G. (ed), Cutting tool applications, v. 67, Materials Park, OH, ASM International, pp. 2–10, 2002.
• ^[19] VEJDANNIK, M., SADR, A., “Automatic microstructural characterization and classification using probabilistic neural network on ultrasound signals”, Journal of Intelligent Manufacturing, v. 29, n. 8, pp. 1923–1940, 2018. doi: http://doi.org/10.1007/s10845-016-1225-y.
  » https://doi.org/10.1007/s10845-016-1225-y
• ^[20] PEDREGOSA, F., VAROQUAUX, G., GRAMFORT, A., et al., “Scikit-learn: machine learning in python”, Journal of Machine Learning Research, v. 12, pp. 2825–2830, 2011.
• ^[21] SRIVASTAVA, N., HINTON, G., KRIZHEVSKY, A., et al., “Dropout: a simple way to prevent neural networks from overfitting”, Journal of Machine Learning Research, v. 15, n. 1, pp. 1929–1958, 2014.
• ^[22] DENKENA, B., BERGMANN, B., WITT, M., “Material identification based on machine-learning algorithms for hybrid workpieces during cylindrical operations”, Journal of Intelligent Manufacturing, v. 30, n. 6, pp. 2449–2456, 2019. doi: http://doi.org/10.1007/s10845-018-1404-0.
  » https://doi.org/10.1007/s10845-018-1404-0
• ^[23] TONG, X., LIU, Q., PI, S., et al., “Real-time machining data application and service based on IMT digital twin”, Journal of Intelligent Manufacturing, v. 31, n. 5, pp. 1113–1132, 2020. doi: http://doi.org/10.1007/s10845-019-01500-0.
  » https://doi.org/10.1007/s10845-019-01500-0
• ^[24] RASCHKA, S., “MLxtend: providing machine learning and data science utilities and extensions to Python’s scientific computing stack”, Journal of Open Source Software, v. 3, n. 24, pp. 638, 2018. doi: http://doi.org/10.21105/joss.00638.
  » https://doi.org/10.21105/joss.00638
• ^[25] DEMIR, H., “Classification of texture images based on the histogram of oriented gradients using support vector machines”, Istanbul University Journal of Electrical & Electronics Engineering, v. 18, n. 1, pp. 90–94, 2018. doi: http://doi.org/10.5152/iujeee.2018.1814.
  » https://doi.org/10.5152/iujeee.2018.1814
• ^[26] WITTEN, I.H., FRANK, E., HALL, M.A., Data mining: practical machine learning tools and techniques, Burlington, MA, Morgan Kaufmann, 2011.
• ^[27] YAO, X., ZHOU, J., LIN, Y., et al., “Smart manufacturing is based on cyber-physical systems and beyond”, Journal of Intelligent Manufacturing, v. 30, n. 8, pp. 2805–2817, 2019. doi: http://doi.org/10.1007/s10845-017-1384-5.
  » https://doi.org/10.1007/s10845-017-1384-5
• ^[28] CHANG, C.C., LIN, C.-J., “LIBSVM: a library for support vector machines”, ACM Transactions on Intelligent Systems and Technology, v. 2, n. 3, pp. 1–27, 2011. doi: http://doi.org/10.1145/1961189.1961199.
  » https://doi.org/10.1145/1961189.1961199
• ^[29] ZHAO, Y.F., XU, X., “Enabling cognitive manufacturing through automated on-machine measurement planning and feedback”, Advanced Engineering Informatics, v. 24, n. 3, pp. 269–284, 2010. doi: http://doi.org/10.1016/j.aei.2010.05.009.
  » https://doi.org/10.1016/j.aei.2010.05.009
• ^[30] WOODS, D.D., “Cognitive technologies: the design of joint human-machine cognitive systems”, AI Magazine, v. 6, n. 4, pp. 86–92, 1985.
• ^[31] SANTOS, M.C.G., FREIRE, C.M.A., “Evaluation by EIE of film containing γ-UPS as passivation coating to electroplated ZnFe and ZnCo alloys”, Matéria, v. 18, n. 3, pp. 1459–1466, 2013. doi: http://doi.org/10.1590/S1517-70762013000400007.
  » https://doi.org/10.1590/S1517-70762013000400007
• ^[32] MALLIK, M.K., RAO, C.S., RAO, V.V.S.K., “Effect of heat treatment on hardness and wear behavior of weld deposited Co-Cr-Mo alloy”, Matéria, v. 20, n. 2, pp. 544–549, 2015. doi: http://doi.org/10.1590/S1517-707620150002.0054.
  » https://doi.org/10.1590/S1517-707620150002.0054
• ^[33] GAUTO, E., KRAMER, G.R., MÉNDEZ, C.M., et al., “Effects of the use of support electrolytes in electrochemical tests of Al-4wt.% Mg alloy exposed to bioethanol fuel”, Matéria, v. 23, n. 2, e12386, 2018.
• ^[34] GONZAGA, M.A.M., PASSOS, T.A., GOMES, R.M., “Characterization of reinforced aluminum composites with quasicrystalline particles produced by stir casting/rolling”, Matéria, v. 25, n. 2, e12531, 2020.
• ^[35] DAI, Y., HUANG, X., BAI, Y., et al., “BRepFormer: transformer-based B-rep geometric feature recognition”, arXiv, 2025. doi: http://doi.org/10.1145/3731715.3733283.
  » https://doi.org/10.1145/3731715.3733283
• ^[36] ZHANG, S., YE, F., WANG, B., et al., “Few-shot bearing fault diagnosis based on model-agnostic meta-learning”, arXiv, 2020. https://arxiv.org/abs/2007.12851, accessed in July, 2025.
  » https://arxiv.org/abs/2007.12851
• ^[37] ZOU, Q., ZHU, L. , “Boundary representation learning via transformer”, arXiv, 2025. https://arxiv.org/abs/2504.07134, accessed in July, 2025.
  » https://arxiv.org/abs/2504.07134
• ^[38] SONG, W., WU, D., SHEN, W., et al., “Meta-learning based early fault detection for rolling bearings via few-shot anomaly detection”, arXiv, 2022. https://arxiv.org/abs/2204.12637, accessed in July, 2025.
  » https://arxiv.org/abs/2204.12637
• ^[39] LIU, Y., YANG, Z., ZOU, X., “Data quantity governance for machine learning in materials science”, National Science Review, v. 10, n. 7, pp. nwad125, 2023. doi: http://doi.org/10.1093/nsr/nwad125.
  » https://doi.org/10.1093/nsr/nwad125
• ^[40] KASEEM, M., ZEHRA, T., DIKICI, B., et al., “Improving the electrochemical stability of AZ31 Mg alloy in a 3.5wt.% NaCl solution via the surface functionalization of plasma electrolytic oxidation coating”, Journal of Magnesium and Alloys, v. 10, n. 5, pp. 1311–1325, 2022. doi: http://doi.org/10.1016/j.jma.2021.08.028.
  » https://doi.org/10.1016/j.jma.2021.08.028
• ^[41] ZHU, L., DONG, J., “Ultrafast carrier dynamics of transition metal dichalcogenides and related heterostructures investigated via optical pump-terahertz probe spectroscopy”, Jiguang Yu Guangdianzixue Jinzhan, v. 62, n. 7, pp. 0700003, 2025.
• ^[42] LIU, Y., ZOU, X., MA, S., et al., “Feature selection method reducing correlations among features by embedding domain knowledge”, Acta Materialia, v. 238, pp. 118195, 2022. doi: http://doi.org/10.1016/j.actamat.2022.118195.
  » https://doi.org/10.1016/j.actamat.2022.118195
• ^[43] ZHAO, Q., AVDEEV, M., CHEN, L., et al., “Machine learning prediction of activation energy in cubic Li-argyrodites with hierarchically encoding crystal structure-based (HECS) descriptors”, Science Bulletin, v. 66, n. 14, pp. 1401–1408, 2021. doi: http://doi.org/10.1016/j.scib.2021.04.029. PubMed PMID: 36654366.
  » https://doi.org/10.1016/j.scib.2021.04.029
• ^[44] HE, B., CHI, S., YE, A., et al., “High-throughput screening platform for solid electrolytes combining hierarchical ion-transport prediction algorithms”, Scientific Data, v. 7, n. 1, pp. 151, 2020. doi: http://doi.org/10.1038/s41597-020-0474-y. PubMed PMID: 32439922.
  » https://doi.org/10.1038/s41597-020-0474-y
• ^[45] LIU, Y., WU, J.-M., AVDEEV, M., et al., “Multi-layer feature selection incorporating weighted score-based expert knowledge toward modeling materials with targeted properties”, Advanced Theory and Simulations, v. 3, n. 2, pp. 1900215, 2020. doi: http://doi.org/10.1002/adts.201900215.
  » https://doi.org/10.1002/adts.201900215
• ^[46] LIU, Y., WU, J., YANG, G., et al., “Predicting the onset temperature (Tg) of GexSe1−x glass transition: a feature selection based two-stage support vector regression method”, Science Bulletin, v. 64, n. 16, pp. 1195–1203, 2019. doi: http://doi.org/10.1016/j.scib.2019.06.026. PubMed PMID: 36659690.
  » https://doi.org/10.1016/j.scib.2019.06.026
• ^[47] LIU, Y., DING, L., YANG, Z., et al., “Domain knowledge discovery from abstracts of scientific literature on nickel-based single crystal superalloys”, Science China. Technological Sciences, v. 66, n. 7, pp. 1815–1830, 2023. doi: http://doi.org/10.1007/s11431-022-2283-7.
  » https://doi.org/10.1007/s11431-022-2283-7
• ^[48] SONG, Y., LI, W., MA, Y., et al., “Construction of hierarchically porous carbon spheres supported nonprecious metal single-atom electrocatalysts for oxygen reduction reaction”, Journal of Power Sources, v. 545, pp. 231913, 2022. doi: http://doi.org/10.1016/j.jpowsour.2022.231913.
  » https://doi.org/10.1016/j.jpowsour.2022.231913
• ^[49] LIU, Y., WU, J., WANG, Z., et al., “Predicting creep rupture life of Ni-based single crystal superalloys using divide-and-conquer approach based machine learning”, Acta Materialia, v. 195, pp. 454–467, 2020. doi: http://doi.org/10.1016/j.actamat.2020.05.001.
  » https://doi.org/10.1016/j.actamat.2020.05.001
• ^[50] LIU, Y., WU, L., YANG, Z., et al., “Descriptors divide-and-conquer enables multifaceted and interpretable materials structure–activity relationship analysis”, Advanced Functional Materials, v. 35, n. 26, pp. 2421621, 2025. doi: http://doi.org/10.1002/adfm.202421621.
  » https://doi.org/10.1002/adfm.202421621
• ^[51] SCHMIDT, J., MARQUES, M.R.G., BOTTI, S., et al., “Recent advances and applications of machine learning in solid‐state materials science”, npj Computational Materials, v. 5, n. 1, pp. 83, 2019. doi: http://doi.org/10.1038/s41524-019-0221-0.
  » https://doi.org/10.1038/s41524-019-0221-0
• ^[52] DIETTERICH, T.G., & quot;Approximate statistical tests for comparing supervised classification learning algorithms", Neural Computation, v. 10, n. 7, pp. 1895–1923, 1998. PubMed PMID: 9744903.
• ^[53] LUTZ, B.; KISSKALT, D.; MAYR, A.; REGULIN, D.; PANTANO, M.; FRANKE, J. In-situ identification of material batches using machine learning for machining operations. Journal of Intelligent Manufacturing, v. 32, n. 5, p. 1485–1495, 2021.

Publication Dates

• Publication in this collection
  22 Aug 2025
• Date of issue
  2025

History

• Received
  02 Dec 2024
• Accepted
  25 June 2025