In this work, a surrogate modeling approach based on graph neural networks (GNNs) is presented to rapidly predict the homogenized stress-strain response of polycrystalline microstructure volume elements (MVEs). Each grain is represented as a node in a graph, with features such as crystallographic orientation, grain size, and aspect ratio, while grain boundaries are encoded as edges. This graph-based representation enables the GNN to capture both local and long-range interactions that govern the macroscopic mechanical response. A synthetic dataset of MVEs was generated using DREAM.3D, covering a wide range of microstructural variations in grain size distributions, morphological anisotropy, and four crystallographic texture classes. Full-field crystal plasticity (CP) simulations performed on these MVEs provided the ground-truth stress-strain data used to train and validate the GNN model. The results demonstrate high correlation between GNN predictions and CP simulations, with strong agreement across different loading conditions and an inference-time speedup of approximately 55,000 × relative to full-field CP, achieved up to the reported accuracies and excluding the one-time training and data-generation overhead. The trained surrogate is then subjected to a comprehensive explainability analysis using integrated gradients, providing grain-level attribution maps that identify the microstructural features most influential to the predicted stress response. Population-level attribution statistics, together with cross-validation against GradientShap, trace the dominant attribution to the statistical representativity of each volume element, whose response dispersion scales inversely with the square root of its grain count. Overall, the proposed framework offers an accurate, efficient, and interpretable tool for microstructure-informed mechanical property prediction in heterogeneous polycrystalline materials.

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.

Rolling is a metal forming process where slabs are passed through rollers to produce strips with specific dimensions and mechanical properties. This process is performed in hot or cold formats. In hot rolling, the workpiece is initially heated above its recrystallization temperature. During the hot rolling process, plastic deformation occurs as the material’s thickness decreases and elongation takes place along the longitudinal axis of the workpiece. Due to the incompressibility of plastic deformation, the material also expands in the transverse direction, a phenomenon known as spread or lateral flow. Modeling spread is crucial for sustainability considerations and meeting customer expectations regarding the quality of the final product. Current prediction methodologies, such as the accurate but slow Finite Element (FE) method or the fast but inaccurate analytical metal forming analysis, are impractical for optimal control. To tackle these challenges, hybrid frameworks have emerged as a promising alternative. The present work aims to develop a fast and accurate model for predicting spread in hot rolling. Specifically, machine learning improves analytical models by leveraging data from a high-fidelity FE model. Initially, we review analytical models for spread, which address key aspects of the problem’s physics. To generate the ground truth (GT) space, an automated FE model for hot strip rolling is created. Moreover, the model’s sensitivity to both process and material parameters is investigated. In the Analytical Predictor Machine Learning Corrector scheme, the analytical models generate initial predictions of GT. In the correction step, a data-driven machine learning model is used to refine these predictions by compensating for deviations from high-fidelity FE simulations. The proposed hybrid framework improves the accuracy of the existing analytical models while preserving their computational efficiency.

In the last decades, there has been an increase in the number of successful machine learning models that have served as a key to identifying and using linkages within the process-structure–property-performance chain for vastly different problems in the domains of materials mechanics. The consideration of physical laws in data-driven modelling has recently been shown to enable enhanced prediction performance and generalization while requiring less data than either physics-based or data-driven modelling approaches independently. In this contribution, we introduce a simulation-assisted machine learning framework applied to the solid-state layer deposition technique friction surfacing, suitable for solid-state additive manufacturing as well as repair or coating applications. The objective of the present study is to use machine learning algorithms to predict and analyse the influence of process parameters and environmental variables, i.e. substrate and backing material properties, on process behaviour and deposit geometry. The effects of maximum process temperatures supplied by a numerical heat transfer model on the predictions of the targets are given special attention. Numerous different machine learning algorithms are implemented, optimized and evaluated to take advantage of their varied capabilities and to choose the optimal one for each target and the provided data. Furthermore, the input feature dependence for each prediction target is evaluated using game-theory related Shapley Additive Explanation values. The experimental data set consists of two separate experimental design spaces, one for varying process parameters and the other for varying substrate and backing material properties, which allowed to keep the experimental effort to a minimum. The aim was to also represent the cross parameter space between the two independent spaces in the predictive model, which was accomplished and resulted in an approximately 44 % reduction in the number of experiments when compared to carrying out an experimental design that included both spaces.

The precise bending of sheet metal structures is crucial in various industrial and scientific applications, whether to modify deformation in an existing component or to achieve specific shapes. Laser peen forming (LPF) is proven as an innovative forming process for sheet metal applications. LPF involves inducing mechanical shock waves into a specimen that deforms the affected region to a certain desired curvature. The degree of deformation induced after LPF depends on numerous experimental factors such as laser energy, the number of peening sequences, and the thickness of the specimen. Consequently, comprehending the complex dependencies and selecting the appropriate set of LPF process parameters for application as a forming or correction process is crucial. The main objective of the present work is the development of a data-driven approach to predict the deformation obtained from LPF for various process parameters. Artificial neural network (ANN) was trained, validated, and tested based on experimental data. The deformation obtained from LPF is successfully predicted by the trained ANN. A novel process planning approach is developed to demonstrate the usability of ANN predictions to obtain the desired deformation in a treated region. The successful application of this approach is demonstrated on three benchmark cases for thin Ti-6Al-4V sheets, such as deformation in one direction, bi-directional deformation, and modification of an existing deformation in pre-bent specimens via LPF.

Within the fields of materials mechanics, the consideration of physical laws in machine learning predictions besides the use of data can enable low prediction errors and robustness as opposed to predictions only based on data. On the one hand, exclusive utilization of fundamental physical relationships might show significant deviations in their predictions compared to reality, due to simplifications and assumptions. On the other hand, using only data and neglecting well-established physical laws can create the need for unreasonably large data sets that are required to exhibit low bias and are usually expensive to collect. However, fundamental but simplified physics in combination with a corrective model that compensates for possible deviations, e.g., to experimental data, can lead to physics-based predictions with low prediction errors, also despite scarce data. In this article, it is demonstrated that a hybrid model approach consisting of a physics-based model that is corrected via an artificial neural network represents an efficient prediction tool as opposed to a purely data-driven model. In particular, a semi-analytical model serves as an efficient low-fidelity model with noticeable prediction errors outside its calibration domain. An artificial neural network is used to correct the semi-analytical solution towards a desired reference solution provided by high-fidelity finite element simulations, while the efficiency of the semi-analytical model is maintained and the applicability range enhanced. We utilize residual stresses that are induced by laser shock peening as a use-case example. In addition, it is shown that non-unique relationships between model inputs and outputs lead to high prediction errors and the identification of salient input features via dimensionality analysis is highly beneficial to achieve low prediction errors. In a generalization task, predictions are also outside the process parameter space of the training region while remaining in the trained range of corrections. The corrective model predictions show substantially smaller errors than purely data-driven model predictions, which illustrates one of the benefits of the hybrid modelling approach. Ultimately, when the amount of samples in the data set is reduced, the generalization of the physics-related corrective model outperforms the purely data-driven model, which also demonstrates efficient applicability of the proposed hybrid modelling approach to problems where data is scarce.

A finite element model is employed to perform a sequentially coupled thermo-mechanical analysis for enabling rapid process simulations of temperature fields, residual stresses and distortions for the production of additively manufactured parts via laser metal deposition. Experimental identification of characteristic process features such as temperature distribution, melt pool dimensions and bead geometries were used for the initial built-up and calibration of the model. The addition of material during process simulation is realised through reactivating inactive elements during the transient heat transfer analysis and through reactivating a combination of inactive and quiet elements during the mechanical analysis. The travelling heat source is geometrically bounded to precisely control the volume of its energy distribution. The results of the transient heat transfer analysis are sequentially coupled to a mechanical analysis for obtaining information on the resulting residual stresses and deformation. Based on the good agreement between numerical and experimental results of the thermal analysis, conclusions on the corresponding residual stress distributions and deformation are made. It is shown that the model represents an efficient tool for process prediction regarding thermal history, residual stresses and finalpart deformations. Finally, the model is utilised to identify parameters and conditions of the process that lead to reduced residual stresses and deformations of the investigated additive part.

The high-potential of lightweight components consisting of similar or dissimilar materials can be exploited by Solid-State Joining techniques. Whereas defects such as pores and hot cracking are often an issue in fusion-based joining processes, via solid-state joining processes they can be avoided to enable high-quality welds. To define an optimal process window for obtaining anticipated joint properties, numerous time and cost consuming experiments are usually required. Building a predictive model based on regression analysis enables the identification and quantification of process-property relationships. On the one hand, mechanical property and performance predictions based on specific process parameters are needed, on the other hand, inverse determination of required process parameters for reaching desired properties or performances are demanded. If these relations are obtained, optimized process parameter sets can be identified while vast numbers of required experiments can be reduced, as underlying physical mechanisms are utilized. In this study, different regression analysis algorithms, such as linear regression, decision trees and random forests, are applied to the refill Friction Stir Spot Welding process for establishing correlations between process parameters and joint properties. Experimental data sets used for training and testing are based on a Box-Behnken Design of Experiments (DoE) and additional test experiments, respectively. The machine-learning based regression analyses are benchmarked against linear regression and DoE statistics. The results illustrate a decryption of relationships along the process-property chain and its deployment to predict mechanical properties governed by process parameters.

Solid-state joining techniques have become increasingly attractive for joining similar and dissimilar materials because it enables further optimization of lightweight components. In contrast to fusion-based joining processes, solid-state joining prevents the occurrence of typical defects such as pores or hot cracking. Machine learning algorithms are powerful tools to identify and quantify relationships between essential features along the process-property chain. In particular, different supervised machine learning algorithms can be used to perform regression analyses and establish correlations between process parameters as well as resulting properties. This can help to circumvent the demand for conducting a vast number of additional experiments to determine optimized process parameters for desired material properties. Additionally, this knowledge can be utilized to obtain a deeper understanding of the underlying mechanisms. In this study, a number of regression algorithms, such as support vector machines, decision trees, random forest and 2nd-order polynomial regression have been applied to correlate process parameters and materials properties for the solid-state joining process of force-controlled friction riveting. Experimental data generated via a central-composite Design of Experiments, serves as source of two separate data sets: one for training and one for testing the machine learning algorithms. The performances of the different algorithms are evaluated based on the determination coefficient R2 and the standard deviation of the predictions on the test data set. The trained algorithms with the best performance measures can be used as predictive models to forecast specific influences of process parameters on mechanical properties. Through the application of these models, optimized process parameters can be determined that lead to desired properties.

Machine learning tools represent key enablers for empowering material scientists and engineers to accelerate the development of novel materials, processes and techniques. One of the aims of using such approaches in the field of materials science is to achieve high-throughput identification and quantification of essential features along the process-structure-property-performance chain. In this contribution, machine learning and statistical learning approaches are reviewed in terms of their successful application to specific problems in the field of continuum materials mechanics. They are categorized with respect to their type of task designated to be either descriptive, predictive or prescriptive; thus to ultimately achieve identification, prediction or even optimization of essential characteristics. The respective choice of the most appropriate machine learning approach highly depends on the specific use-case, type of material, kind of data involved, spatial and temporal scales, formats, and desired knowledge gain as well as affordable computational costs. Different examples are reviewed involving case-by-case dependent application of different types of artificial neural networks and other data-driven approaches such as support vector machines, decision trees and random forests as well as Bayesian learning, and model order reduction procedures such as principal component analysis, among others. These techniques are applied to accelerate the identification of material parameters or salient features for materials characterization, to support rapid design and optimization of novel materials or manufacturing methods, to improve and correct complex measurement devices, or to better understand and predict fatigue behavior, among other examples. Besides experimentally obtained datasets, numerous studies draw required information from simulation-based data mining. Altogether, it is shown that experiment- and simulation-based data mining in combination with machine leaning tools provide exceptional opportunities to enable highly reliant identification of fundamental interrelations within materials for characterization and optimization in a scale-bridging manner. Potentials of further utilizing applied machine learning in materials science and empowering significant acceleration of knowledge output are pointed out.

Wire-based laser metal deposition enables to manufacture large-scale components with deposition rates significant higher compared to powder-based laser additive manufacturing techniques, which are currently working with deposition rates of only a few hundred gram per hour. However, the wire-based approach requires a significant amount of laser power in the range of several kilowatts instead of only a few hundred watts for powder-based processes. This excessive heat input during laser metal deposition can lead to process instabilities such as a non-uniform material deposition and to a limited processability, respectively. Although, numerous possibilities to monitor temperature evolution during processing exist, there is still a lack of knowledge regarding the relationship between temperature and geometric shape of the deposited structure. Due to changing cooling conditions with increasing distance to the substrate material, producing a wall-like structure results in varying heights of the individual tracks. This presents challenges for the deposition of high wall-like structures and limits the use of constant process parameters. In the present study, the temperature evolution during laser metal deposition of AA5087 using constant process parameters is investigated and a scheme for process parameter adaptions in order to reduce residual stress induced componential distortions is suggested.

Laser additive manufacturing (LAM) has become increasingly popular in industry in recent decades because it enables exceptional degrees of freedom regarding the structural design of lightweight components compared to subtractive manufacturing techniques. Laser metal deposition (LMD) of wire-fed material shows in particular the advantages such as high process velocity and efficient use of material compared to other LAM processes. During wire-based LMD, the material is deposited onto a substrate and supplemented by successive layers allowing a layer-wise production of complex three-dimensional structures. Despite the increased productivity of LMD, regarding the ability to process aluminium alloys, there is still a lack in quality and reproducibility due to the inhomogeneous temperature distribution during the process, leading to undesired residual stresses, distortions and inconsistent layer geometries and poor microstructures. In this study, the aluminium alloy AA5087 as wire and AA5754 as substrate material were utilized for LMD. In order to obtain information about the temperature field during LMD, thermocouple and thermography measurements were performed during the process. The temperature measurements were used to validate a finite element model regarding the heat distribution, which will be further used to investigate the temperature field evolution over time. To consider the continuous addition of material within the FE-model, an inactive/active element approach was chosen, where initially deactivated elements are activated corresponding to the deposition of material. The first results of the simulation and the experiments show good agreement. Therefore, the model can be used in the future for LMD process optimization, e.g., in terms of minimizing local variations of the thermal load for each layer.

One of the rate-limiting steps in the large-scale production of cellulose nanopaper-reinforced polymer composites is the time consuming dewatering step to produce the reinforcing cellulose nanopapers. In this work, we show that the dewatering time of bacterial cellulose (BC)-in-water suspension can be reduced by reducing the grammage of BC nanopaper to be produced. The influence of BC nanopaper grammage on the tensile properties of BC nanopaper-reinforced polylactide (PLLA) composites is also investigated in this work. BC nanopaper with grammages of 5, 10, 25 and 50 g m(-2) were produced and it was found that reducing the grammage of BC nanopaper from 50 g m(-2) to 5 g m(-2) led to a three-fold reduction in the dewatering time of BC-in-water suspension. The porosity of the BC nanopapers, however, increased with decreasing BC nanopaper grammage. While the tensile properties of BC nanopapers were found to decrease with decreasing BC nanopaper grammage, no significant difference in the reinforcing ability of BC nanopaper with different grammages for PLLA was observed. All PLLA composite laminates reinforced with BC nanopapers possessed similar tensile modulus of 10.5-11.8 GPa and tensile strength of 95-111 MPa, respectively, at a BC loading fraction v(l,BC) = 39-53 vol.-%, independent of the grammage and tensile properties of the reinforcing BC nanopaper.