Papers by Keyword: Machine Learning

Abstract: In harsh and aggressive environments, steel reinforcement corrodes, leading to a loss of rebar strength and spalling of concrete due to internal stresses caused by the swelling of corrosion products. Therefore, in order to increase the lifespan of a structure, noncorrosive reinforcement is recommended, which includes Glass Fibre Reinforcing Polymer (GFRP) bars. These bars also offer several other advantages over steel, which include higher tensile strength, low weight and cost-effectiveness. These bars exhibit a distinct bond with concrete due to linearly elastic behaviour and different surface deformation patterns. Several empirical equations have been established to analytically predict the bond strength of these bars. This study finds out that even though these empirical models provide useful insights, they may have limitations in predicting bond strength with significant accuracy; therefore, it is imperative to come up with more rigorous data-driven prediction models. This study presents the application of an eXtreme Gradient Boosting (XGBoost)-based machine learning model which predicts the bond strength with significant accuracy, exhibiting a 0.876 coefficient of determination and a 2.319 root mean square error on the full set of data, which concludes improved predictive capability compared to traditional empirical equations.

Abstract: This paper aims to extend the evaluation of the process of electric discharge machining by analysing the discharges. Therefore, a method for detecting and classifying discharges was developed. To detect different discharge types, experiments were conducted with varying of technology parameters, such as peak current or duty factor. During the experiments, the voltages and the currents were measured via an oscilloscope. For the classification, an unsupervised machine learning method was applied, to cluster and classify the detected discharges and compare them with the measured material removal rate and the measured tool wear rate.

Abstract: Despite the growing use of biopolymers in automotive, packaging and structural applications, predictive modelling of their elastic–viscoplastic deformation remains limited. In this work, a micromechanically based constitutive model is proposed to describe the micro‑ to macroscopic behaviour of a semi‑crystalline PLA matrix reinforced with short hemp fibers. The formulation relies on a multiplicative split of the deformation gradient into elastic and viscoelastic–plastic parts, with elasticity governed by fiber and crystalline phases and time‑dependent deformation localized in the amorphous phase. High fiber content and strong fiber–matrix bonding enable the suppression of lattice crystalline anisotropy, leading to a compact model with a reduced number of internal variables. The model is calibrated and validated using uniaxial tensile tests on pellet‑extrusion 3D‑printed specimens with controlled porosity and plasticiser content, and reproduces nonlinear loading, unloading, creep and stress relaxation. In a second step, synthetic data generated by the constitutive model are used to train surrogate machine‑learning models, which are discussed as a perspective for accelerating long‑term simulations and parametric studies in forming applications.

Abstract: Accurate modeling of elastoplastic behavior is crucial for forming simulations, yet conventional constitutive laws require extensive calibration and often fail to generalize across diverse loading paths. To address this limitation, a thermodynamically informed neural-network framework is proposed for predicting one-dimensional stress evolution. The model integrates physical consistency into a data-driven formulation by coupling two neural components: one learns the state evolution, predicting increments of the internal variable, while the other approximates the Helmholtz free-energy potential, from which stresses are obtained via automatic differentiation. Synthetic datasets generated from randomized strain paths with power-law hardening were used for training, ensuring broad coverage of nonlinear responses. The model successfully reproduces monotonic, unloading, reverse, and random loading behaviors with minimal error accumulation and stable recursive inference. Owing to its incremental formulation, the framework maintains predictive accuracy beyond the trained strain range, offering a physically interpretable and data-efficient alternative to conventional constitutive models.

Abstract: Draw-in distance is a key index for evaluating the quality of sheet metal stamping. Its accurate prediction is therefore required for tool design and process control. Traditional finite element (FE) simulations, while accurate, are computationally intensive and time-consuming for iterative design optimization. In this study, a graph neural network (GNN) method is proposed to predict draw-in during sheet metal forming. A dataset was built from FE simulations with different process settings, including blank holder force and draw bead force. The GNN model uses node coordinates and edge features to describe the spatial relations in the sheet. A multi-level loss function was applied. The coordinate error and edge distance error were included. In this way, the shape of the sheet is better preserved. The trained GNN can be used as a fast model for draw-in prediction. It can also be used for inverse analysis, where the process parameters are found from a given draw-in result. This provides an efficient tool for sheet metal forming design and optimization.

Abstract: Bobbin-Tool Friction Stir Welding (BT-FSW) is a solid-state joining process in which axial forces are internally balanced by the tool, eliminating the need for a backing plate and enabling the joining of hollow aerospace structures. Owing to the coupled thermo-mechanical nature of the process, weld stability and quality are governed by the interaction between process parameters and the resulting torque response, which is difficult to assess in situ using conventional sensing alone. BT-FSW experiments were performed on AA2024-T351 sheets with thicknesses of 2.4, 2.8 and 3.6 mm using a structured Design of Experiments (DoE). The 3.6 mm joints achieved approximately 90 % of the base-material strength, while the 2.4 and 2.8 mm joints reached about 80 % and 85 %, respectively. These mechanical results were used as ground truth to train machine learning regression models for steady-state torque predictions. By augmenting nominal process parameters with force-derived features, the proposed soft-sensing framework achieved strong agreement between predicted and measured torque, demonstrating that compact, physics-based feature engineering enables reliable prediction under limited experimental data conditions.

Abstract: When it comes to predicting quality-relevant outcomes of rotary draw bending procedures, like springback and geometric errors, machine learning algorithms have demonstrated encouraging results. However, the challenges associated with understanding these models’ predictions still restrict their actual application in industrial contexts. A web-based 3D visualization designed to help with interactive exploration and explainability of machine learning predictions in rotary draw bending is evaluated through an expert-centered study. Based on an earlier Random Forest regression model, the visualization lets users change important process parameters and view the projected tube geometry and springback in real time. Sixteen experts participated in a structured online survey that combined openended comments, subjective agreement scores, and interactive parameter modification tasks. Results show that while multi-objective optimization remained difficult, participants with different degrees of machine learning knowledge and tube-bending experience were generally able to identify appropriate parameter settings in single-objective problems. Subjective assessment and qualitative feedback from the participants also highlight that the visualization could be used to assist in understanding model behavior and in early process design and training situations. Overall, our study suggests that experts in tube bending applications find benefit from the interactive 3D visualization of the predicted geometry and as a useful interface for exploring machine learning models’ predictions.

Abstract: Previous research shows that predicting width deviation inherits a central importance in hot rolling processes, so that the pass planning in the hot strip mill (HSM) can be optimized. These predictions can be enabled using machine learning, complementing analytical formulations of width spread. For reliable production, it is important for the plant operator to be able to control the geometry with high accuracy across the entire plant. Therefore, the width must be accurately known throughout the entire HSM. This paper aims on the prediction of width deviation in early product stages during the roughing mill processes, where the major deformation takes place, and thus also has the most significant influence on the width spread. Therefore, this paper takes industrial data into account which is also used for the roll passing planning. To achieve a prediction during rough rolling for the width after exiting the mill (future state of the strip width), various machine learning algorithms were implemented and tested. The prediction results are evaluated against an inline width measurement, where the XGB model performs best with a Root Mean Squared Error (RMSE) of 1.11 mm. Subsequently, feature importance analyses are used to examine which features are relevant for the prediction result and to elaborate which significance process-and geometry data has on the same strip.

Abstract: Machine learning (ML) algorithms have been studied in literature as an inverse method to predict material constitutive parameters. However, these approaches are often dependent on the mesh discretisation settings applied during numerical simulations, and then difficulty model adaptation to experimental digital image correlation (DIC) subsets. Although a recent study explores the use of an interpolation-based approach to achieve experimental adaptation from numerically-based trained ML models, the proposed methodology lacks evaluation using experimental data. As a follow-up, this study proposes a new evaluation approach. Numerical data is DIC-levelled via MatchID software and then submitted to interpolation. An XGBoost algorithm is then trained on interpolated DIC data and evaluated for parameter prediction, comparing the obtained results with those obtained from the model trained on interpolated numerical data. Overall, the proposed DIC-levelling and interpolation pipeline yields an excellent predictive performance, with results comparable to those obtained when training on interpolated numerical data. The largest deviations are observed for the hardening exponent, while the remaining parameters are predicted with consistently high accuracy. These findings validate the practical applicability of the interpolation-based strategy to reduce the subset scheme dependency of ML models trained on real experimental data.

Abstract: In this work, a new methodology for the identification of the CPB06 yield criterion parameters is presented. This methodology is based on the application of Machine Learning models and the Levenberg-Marquardt optimization algorithm. The proposed methodology relies on data obtained from a biaxial tensile test with a cruciform specimen and aims to overcome some of the challenges usually faced during material characterization with the CPB06 yield criterion. The predictive performances achieved were positive overall, when comparing the yield surfaces obtained for testing cases, highlighting the potential of the proposed methodology.