Recent Advances and Applications of Machine Learning in Metal Forming Processes

1. Introduction

Machine Learning (ML) is a subfield of artificial intelligence, focusing on computational algorithms that are designed to learn and improve themselves, without the need to be explicitly programmed. ML algorithms have been applied in several fields, being particularly useful for solving complex tasks that would normally require the understanding and building of either impossible or complex first-principle models, i.e., based on fundamental physical laws.

ML approaches are emerging in the area of metal forming processes, driven by the increasing availability of large datasets, coupled with the exponential growth of computer performance. In fact, there has been a growing interest in evaluating the capabilities of ML algorithms in studying topics related to metal forming processes, such as: classification, detection and prediction of forming defects; material modelling and parameters identification; process classification and selection; process design and optimization. The purpose of this Special Issue is to disseminate state-of-the-art ML applications in metal forming processes.

2. Contributions

The Special Issue is comprised of a total of ten research articles related to ML applications for metal forming processes, including: prediction of forming results [1] and their energy consumption [2]; constitutive modelling [3] and parameters identification [4]; process parameters optimization [4,5]; prediction, detection and classification of defects [6,7,8]; prediction of mechanical properties [9,10]. The following paragraphs summarize the contributions of these works.

Several Machine Learning (ML) algorithms can be found in the literature, each with their advantages and disadvantages. In this Special Issue, two papers [1,2] compared the predictive performance of various ML algorithms in two different applications. Marques et al. [1] studied the performance of different parametric and non-parametric metamodels in predicting the forming results of the U-Channel and the Square Cup forming process. For the non-parametric techniques (ML-based), the metamodels trained with Multi-Layer Perceptron, Gaussian Processes, Kernel Ridge and Support Vector Regression algorithms were more accurate than those trained with Decision Trees, Random Forest and k-Nearest Neighbors algorithms. Additionally, the parametric metamodeling techniques, Response Surface Method and Polynomial Chaos Expansion, also showed themselves to be competitive alternatives to the best ML-based metamodels. Mirandola et al. [2] also compared the performance of different ML algorithms, but for the prediction of the energy consumption in radial axial-ring rolling forming process. Eight different ML algorithms (Random Forest, Gradient Boosting, Artificial Neural Network, Ridge, Lasso, Elastic Net, Kernel Ridge and Support Vector Regression) were used to predict energy consumption based on material, geometrical, and process parameters. The trained ML models proved to be reliable, even for extrapolation predictions (i.e., outside the training data range). The best prediction accuracy was reached by the model trained with the Gradient Boosting algorithm.

Numerical simulation is nowadays an essential tool for the development and optimization of metal forming processes. The accuracy of the numerical simulations requires constitutive models capable of describing the mechanical behavior of the material. On this topic, two papers of the special issue [3,4] investigated the application of ML to constitutive modeling and material parameters identification. Lourenço et al. [3], explored and discussed the potential contributions of ML in terms of elastoplasticity constitutive modelling. This work discusses and analyses four recent advances and applications of ML: parameters identification; the enhancement of traditional constitutive models; the development of data-driven constitutive models with embedding physical and empirical knowledge; the development of constitutive models fully based on data-driven approaches. In summary, the authors demonstrated the potential of ML-based approaches to solve diverse and complex constitutive modelling problems. Cruz et al. [4] explored the performance of ML models, using shallow artificial neural networks, in two applications: (i) identification of constitutive parameters (isotropic hardening) with a three-point bending test; (ii) optimization of a process parameter (punch displacement), to obtain a desired bending angle (after springback) in press-brake air bending process. In both applications, the trained ANN models were able to solve the identification and optimization problem with reliable results.

The optimization of forming process parameters was also investigated by Palmieri et al. [5]. This work proposes a method to perform real-time control of the blank holder force during a deep-drawing process. Metamodels were initially built with kriging technique to establish a relationship between the process parameters and quality indices. Afterwards, a multi-objective optimization was performed to obtain the blank holder force that guarantees a component without defects in the presence of variability in the yield stress values and lubrication conditions. The result was a regulation curve that is useful for real-time control of the blank holder force to avoid defects in the deep-drawing component.

The application of ML algorithms to predict, detect, and classify defects was investigated by three works of this Special Issue [6,7,8]. Hao et al. [6] proposed a method to classify hot rolling strip surface defects based on the Wasserstein Generative Adversarial Network (WGAN) and an attention mechanism. A dataset of defects images was collected, and then the WGAN model was used to generate additional images (data augmentation). The image classification was performed with the model (SE-ResNet34), coupled with an attention mechanism that enables SE-ResNet34 model to focus on the most valuable information, in order to improve the classification accuracy. The proposed method revealed an excellent classification accuracy of hot rolling strip steel surface defects. Wang et al. [7] also investigated an intelligent recognition model for hot rolling strip surface defects based on convolutional neural networks (CNN), to improve the detection accuracy. The most common defects were classified in five categories (Upwarp, Black Line, Crack, Slag Inclusion and Gas Hole), and a database of defect images was built. Using this data, defect recognition models were established using CNN. The results showed that, depending on the type of the CNN, it was possible to obtain a high recognition accuracy in a short period of time. Lee et al. [8], also proposed a methodology based on CNN to predict the buckling instability of automotive sheet metal panels. A CNN model was used to establish relations between image results on indentation points, evaluated in several localizations of the panels, and the magnitude of the buckling instability. The developed method was able to accurately predict the buckling instability magnitude for automotive sheet metal panels.

Besides forming defects, the mechanical properties of the final product are also an essential aspect to control during metal forming processes. Two works from this Special Issue focused on the application of ML to predict mechanical properties [9,10]. Wu et al. [9] proposed a new prediction model based on Multidimensional Support Vector Regression, combined with a feature selection method, which involves maximum information coefficient correlation characterization and complex network clustering. This method allows for the selection of the most representative input variables to reduce the input dimensionality. The proposed model was used to predict the steel mechanical properties based on the conditions of four main processes (smelting, continuous casting, hot rolling, and cold rolling). Compared to other models, the proposed model had, simultaneously, the highest prediction accuracy and the lowest computational complexity. On the same subject, Merayo et al. [10] developed a methodology to optimize the topology of a multilayer artificial neural network, to predict the ultimate tensile strength of aluminum alloys based on the information of their chemical composition and tempering process. The methodology consists in optimizing the number of nodes of two hidden layers, to maximize the accuracy of the predictions without compromising the computational cost. The optimized artificial neural network was able to give accurate predictions.

3. Conclusions

The Special Issue covers 10 papers about the application of Machine Learning (ML) approaches to Metal Forming Processes. Based on these works, the application of ML approaches revealed itself to be a success, reaching accurate predictions and classification tasks. As Guest Editors, we are confident that the quality of the methods and results presented in this Special Issue, represent a significant contribution for the dissemination and advancement of future research of Machine Learning in metal forming processes.