Publications

Bearing steels are complex materials composed of an iron matrix and a well defined and precise amount of several alloying elements. In order to improve sustainability and circularity, there is a tendency to increase the utilisation of scrap material for their production. The variability of the composition of scrap material has a direct impact on the properties of the final steels: There is less control on their composition due to the possible presence of larger amounts of tramp and alloying elements. One way to study the effect of tramp elements is by using universal machine learning interatomic potentials. These types of potential render the investigation of multi-element systems possible. They permit the study of interactions between iron atoms in the matrix and multiple concurrent tramp and alloying elements, a feature that is currently not available in classical potentials. In this work, we present a benchmark of four state-of-the-art universal machine learning interatomic potentials (Crystal Hamiltonian Graph Neural Network (Deng et al 2023 Nat. Mach. Intell. 5 1031–41) (v0.2.0 and v0.3.0), Materials 3-body Graph Network (Chen and Ping Ong 2022 Nat. Comput. Sci. 2 718–28), Multiple Atomic Cluster Expansion (Batatia et al 2022 Advances in Neural Information Processing Systems vol 35 pp 11423–36)) and SevenNet (Park et al 2024 J. Chem. Theory Comput. 20 4857–68), and study their applicability to the simulation of systems relevant to steels. For pure Fe, all potentials accurately predict the equilibrium lattice parameter, but the accuracy varies for other properties. For most solute–solute and solute–vacancy interactions all interatomic potentials tend to capture the general trends though there is a disparity in the predicted magnitudes. While currently ‘off-the-shelf’ universal machine learning interatomic potentials fail to predict some key properties, some of them show significant potential to serve as starting point for further training and refinement.

Link: https://www.doi.org/10.1088/1361-651X/adb483

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.

Link: https://doi.org/10.4028/p-Vj1faR

Shareable tables generated by Foundational Models (FM) are increasingly adopted in industry and scientific applications. With the advent of such data, governance of it becomes an ever more pressing matter. Towards this, watermarking is shown to be a promising approach. While existing watermarking approaches successfully embed watermarks in natural language, they fall short for tabular data due to the susceptibility of token-based representations to post-processing and the arbitrary column ordering. We propose a novel watermarking scheme for tabular data, Tabular Token(less) Watermark (TabTokWak), which embeds a watermarking signal on FM-generated tabular value and its tokenized representation. To be invariant to table generation order, our token watermark uses a seeded pattern to derive its watermark signal. In our experiments on five real-world datasets and a representative tabular FM, we demonstrate that TabTokWak enables effective data governance with high detection efficiency and minimal impact on data quality. Therewith, empirically demonstrating TabTokWak’s detectability and its robustness with a favorable quality trade-off. 

Link to article: https://doi.org/10.1007/978-981-92-1300-9_23

Code online: https://gitlab.ewi.tudelft.nl/dmls/publications/tabtokwak.

In this study, we develop a Bayesian data assimilation framework that combines a mean-field model of static recrystallization (MiReX) with a Sequential Importance Resampling (SIR) particle filter to estimate key material parameters from controlled synthetic experiments. MiReX, originally developed as a microstructurally based extension of Johnson–Mehl–Avrami–Kolmogorov kinetics, is used as a forward model in which the uncertain quantities include the grain-boundary mobility parameters (prefactor and activation energy), a stored-energy coefficient, an Avrami-type exponent, and an interface length scale. Synthetic recrystallized-fraction measurements are generated at two isothermal holding temperatures using a reference parameter set and are perturbed with Gaussian noise to mimic experimental uncertainty. Starting from broad uniform prior ranges, the particle filter propagates an ensemble of MiReX trajectories in time, updates particle weights using a Gaussian likelihood, and applies systematic resampling combined with Liu–West kernel regularization to reduce particle degeneracy while preserving posterior variance. The posterior obtained after assimilating the first temperature dataset is used as the prior for the second dataset, enabling sequential multi-temperature calibration. The synthetic experiments show that the framework recovers the reference parameters within credible intervals and provides tight uncertainty bounds on the predicted recrystallization kinetics. These results demonstrate that combining a physically based mean-field recrystallization model with sequential Monte Carlo methods provides a robust route for probabilistic parameter estimation and uncertainty quantification in microstructure evolution models.

Link: https://doi.org/10.4028/p-M2Pzus

The 29th International ESAFORM Conference on Material Forming (ESAFORM 2026, Thessaloniki, Greece, from 27 to 29 April 2026) is a joint scientific event of the oldest and most prestigious Mechanical Engineering academic schools in Greece, the Aristotle University of Thessaloniki and the National Technical University of Athens, respectively. The conference publication contains sixteent chapters the topicaly related with 16 Mini-Symposia around the following topics: additive manufacturing, composites forming processes, extrusion and drawing, forging and rolling, formability of metallic materials, friction and wear in metal forming, incremental and sheet metal forming, innovative joining by forming technologies, optimization and inverse analysis in forming, machining, cutting and severe plastic deformation, material behaviour modelling, new and advanced numerical strategies for material forming, non-conventional processes, polymer processing and thermo-mechanical properties, sustainability in material forming, and advanced metal casting. The edition will be useful for a wide range of specialists in structural materials forming processes.

Link: https://doi.org/10.4028/b-VxX0gh

In this study, a continuous data space driven computational mechanics framework that enhances the original data-driven methodology is presented. The traditional data-driven approach relies on finite, discrete datasets and direct projections between the dataset and the constraint space dictated by the problem settings. The use of a discrete dataset entails accuracy problems as the dataset is inherently subject to noise and outliers. Moreover, finding the optimal datapoint in the dataset is computationally demanding especially for large sized datasets and geometrically higher dimensional problems. The proposed framework addresses these limitations by leveraging a continuous data space constructed via Gaussian mixture modeling to replace the discrete dataset. The continuous data space can be interpreted as a data density field that can be used to assess the likelihood of a point capturing the underlying material behavior. Hence the effect of noise and outlier on the solution accuracy is reduced. By eliminating search queries in the discrete dataset, the computational cost is significantly reduced as well. Moreover, the data-to-constraint space projection is conducted efficiently via a search direction deduced from the continuous data space. Numerical experiments demonstrate the capability of the framework to enhance accuracy and computational efficiency.

Link: https://doi.org/10.1016/j.euromechsol.2025.105697

This work presents a data analysis-based, group-aware framework for predicting quality indicators with anomaly detection in non-i.i.d. datasets that exhibit short temporal dependencies. The design is motivated by statistical diagnostics of temporal autocorrelation and intraclass variance, which highlight the need for causal temporal encoding and group-level decomposition. The framework integrates a residual-boosted regressor, a group-aware anomaly detector, and a calibrated fusion scheme that balances precision and recall. Evaluation is conducted on real production data from hot strip mill operations, with coiling temperature prediction serving as a case study. A key contribution is interpreting coiling temperature dips, previously treated as outliers, as proxies for surface anomalies, thereby enabling their explicit detection. Results demonstrate consistent gains over physics-based and tabular machine learning baselines, confirming that the framework provides more reliable quality-risk indication for decision support in industrial predictive-control workflows.

Link: https://doi.org/10.1016/j.rineng.2026.110483

This study focuses on a critical aspect of implementing prognostics and health management (PHM) for assets: the creation of a descriptive dataset. In real-world applications, dealing with sparse and unlabelled big data is common, particularly in industries like production lines where complex subprocesses are monitored by multiple sensors. Moreover, selective application of quality control means that much of the data lacks information about end properties, making datasets provided by manufacturers unsuitable for PHM frameworks. This work aims to bridge the gap between raw production data and PHM frameworks, focusing on steel manufacturing management. In the context of steel manufacturing, compromised surface quality, characterized by thicker oxide layers chipping during milling, has been observed. We propose inferring compromised coils by analyzing temperature profiles directly before the coiling station to address this. Deviations from the goal temperature profile can indicate compromised surface quality, eliminating the need for tedious oxide layer thickness measurements, which are not feasible for continuous hot strip milling processes. The available dataset comprised multiple years of production, with no direct indication of the surface quality. Exploratory clustering analysis was the first step in the lack of labels. Even though indicative of the underlying pattern of the healthy/damaged coils distinction, three shortcomings were identified. Clustering was solely based on the similarity between the temperature profiles of the coils, so no domain knowledge was included regarding the goal temperature profile. Additionally, since different steel grades have different goal profiles, the model needs to be specifically trained for each grade. Also, a soft classification between healthy and damaged can provide more detailed information about the surface quality. Coils with low-confidence classifications can be identified and treated accordingly, thereby improving PHM framework performance by providing a dataset with only high-confidence samples. To tackle these issues, an expert-knowledge-based normalization technique and feature engineering, paired with synthetic labelling, contributed to the creation of a soft neural network classifier. This study presents the reality of handling real-world data for PHM applications and highlights the need for careful and informed feature extraction. This ensures the seamless integration of PHM frameworks into real-world systems, ultimately enhancing production yield by improving end-product quality.

Link: https://doi.org/10.36001/phme.2024.v8i1.3968

Recycling steel at scale is hindered by tramp elements such as Cu and Sn, which degrade material properties. Atomistic simulations using foundational machine-learned interatomic potentials (MLIPs) trained on large databases, such as Materials Project, Alexandria, and OMAT, offer a promising approach to study the effects of these impurities. However, fine-tuning these models to specific systems can lead to catastrophic forgetting–the loss of general chemical knowledge acquired during pretraining. Here, we evaluate forgetting in three foundational MLIPs: CHGNet, SevenNet-O, and MACE, by fine-tuning on a data set of bcc-based structures, with Fe atoms only. When evaluated on a subset of the Materials Project data set with a learning rate of 0.0001, the fine-tuned MLIPs of CHGNet and SevenNet-O exhibited only a minor increase in RMSE of 0.047 and 0.022 eV/atom, respectively, indicating markedly minor forgetting. In contrast, fine-tuned MACE exhibited catastrophic forgetting, despite a range of additional strategies such as layer freezing and data set replay. We attribute the catastrophic forgetting to architectural sensitivity. These results highlight the importance of fine-tuning hyperparameters, model architecture, and data set design, with fine-tuned models of CHGnet and SevenNet-O showing some potential for efficient and transferable modeling of recycled steels.

Link: https://doi.org/10.1007/s11669-025-01225-z

Link: https://www.calphad2024.kit.edu/downloads/CALPHAD2024_Book_of_Abstracts_Web.pdf#Bookhttps://www.calphad2024.kit.edu/

Cold rolling forces are strongly affected by lubrication and material properties, and maypotentially be used to estimate material property variations along the coil. The 1D slab method iscommonly used to estimate rolling forces as it is computationally inexpensive. By model definition,the pressure distribution as modelled in the slab method has a single peak or friction hill in the rollbite. However, it has been observed in several studies that multiple local peaks can appear in thepressure distribution in the roll bite. Additionally, material anisotropy also affects the contact pressuredistribution and steady state roll force. In the present work, a 2D plane strain finite element rollingmodel is used for a detailed sensitivity study of the multitude of parameters that affect the verticalpressure distribution and the steady state roll force. The considered model parameters are materialanisotropy, entry sheet thickness, roll radius, tensions, coefficient of friction and roll gap.

Link: https://doi.org/10.4028/p-I1nbWq

Synthetic data offers alternatives for data augmentation and sharing. Till date, it remains unknown how to use watermarking techniques to trace and audit synthetic tables generated by tabular diffusion models to mitigate potential misuses. In this paper, we design TabWak, the first watermarking method to embed invisible signature that control the sampling of Gaussian latent codes used to synthesize table rows via the diffusion backbone. TabWak has two key features. Different from existing image watermarking techniques. TabWak uses self-cloning and shuffling to embed the secret key in postitional information and random seeds that control the Gaussian latents, allowing to use different seeds at each row for high inter-row diversity and enabling row-wise detectability. To further boost the robustness of watermark detection against post-editing attacks. TabWak uses a valid-bit mechanism that focuses on the tail of the latent code destribution for superior noise resilience. We provide theoretical guarantees on the row diversity and effectiveness of detectability. We evaluate TabWak on five datasets against baselines to show that the quality of watermarked tables remains nearly indistinguishable from non-post-editing attacks, with a 100% true positive rate at a 0.1% false positive rate on synthetic tables with fewer than 300 rows.

Link: https://iris.unito.it/bitstream/2318/2066932/1/11455_TabWak_A_Watermark_for_T.pdf

Code online: https://github.com/chaoyitud/TabWak