Leibniz MMS Days 2025
March 26 - March 28, 2025
Rostock-Warnemünde

A poster session will take place on March 26, 2025 from 17:45 in the Lecture Hall with the following posters (list to be continued, submission of further contributions is still possible):

Computational Material Science, AI

Reproducible plasma modelling with COMSOL Multiphysics

Markus Becker (INP Greifswald), joint work with Marjan Stankov

Commercial simulation tools like Comsol Multiphysics offer flexible capabilities for conducting numerical experiments. However, when numerous equations must be implemented, the modelling process can become time-consuming, prone to errors and hard to reproduce, e.g. for verification. Modelling low-temperature plasmas also faces these challenges, as it often needs to consider large sets of balance equations for the plasma species. To address this, the Matlab-Comsol toolbox MCPlas was developed. The latest version of MCPlas presented here enables the automated implementation of fluid-Poisson plasma models in Comsol using JSON-schema-based input files that define the species, their reaction mechanisms, and other relevant parameters. By using MCPlas for different application cases it is demonstrated how MCPlas can support the efficient and reproducible implementation of state-of-the-art plasma models in Comsol.

Attention-Based Physics-Informed Neural Networks for Solving the Spatially Dependent Electron Boltzmann Equation

Ihda Chaerony Siffa (INP Greifswald)

Accurate modeling of weakly ionized, non-thermal plasmas is crucial for optimizing plasma processing techniques used in various applications, including semiconductor manufacturing and material treatment. A key challenge lies in efficiently solving the electron Boltzmann equation (EBE) describing the electron kinetics. Traditional numerical methods for solving this equation can be difficult to develop. This work investigates the application of physics-informed neural networks (PINNs) to solve the spatially one-dimensional EBE within the two-term approximation. To overcome convergence challenges, we introduce an attention-based neural network, which effectively mitigates trivial solutions and enhances overall accuracy. We demonstrate the performance of our approach by means of numerical experiments on argon and neon plasmas subjected to homogeneous electric fields. Good agreement is achieved between the PINN results and those obtained from established numerical methods.

Embedded Representation Learning of Acoustic Emission Events for In-Situ Crack Detection in Laser Powder Bed Fusion

Denys Kononenko (IFW Dresden)

Laser Powder Bed Fusion (LPBF) technology is one of the most promising additive manufacturing methods due to its capability to produce complex geometries with high precision and near-net shape. Its advantages, such as design flexibility and minimal material waste, have led to widespread adoption in aerospace, medical, and automotive industries. However, LPBF can be susceptible to various process-induced defects, including cracks, lack of fusion, and porosity. These defects compromise the mechanical integrity of finished parts, hindering broader implementation of this otherwise transformative technology. In response to this challenge, in-situ monitoring and early detection of defects such as cracks is essential to ensure part quality and reliability. In this work, we develop a novel approach for differentiating acoustic emission (AE) events based on learned embedded representations, with the goal of identifying cracks during LPBF. Our method starts by extracting short acoustic emission events from the continuous AE signal, which are then used to train an autoencoder. The autoencoder learns low-dimensional latent representations of the AE events, capturing subtle patterns indicative of defects. These latent embeddings enable clustering of events into meaningful groups, facilitating the detection of crack-related signatures. Moreover, the system can serve as a foundation for unsupervised data collection to train classification models, further improving real-time crack detection capabilities. The proposed framework paves the way toward robust, in-situ quality assurance in LPBF processes, providing manufacturers with a powerful tool for defect mitigation and improved part reliability.

Machine learning based scale bridging for permeability prediction of fibrous structure

Denis Korolev (WIAS Berlin)

The manufacturing of fiber-reinforced polymers (FRP) requires fiber reinforcement by infiltrating fibers with resin under pressure. Predicting resin flow corresponds to modeling at multiple scales: microscale (flow between impermeable fibers), mesoscale (flow between permeable rovings), and macroscale (mold filling). Due to high computational costs of simulations, these scales are usually treated separately, requiring data exchange—downscaling structural details and upscaling permeability. This process, known as scale bridging, introduces necessary simplifications, but the computational costs still remain high. Scientific machine learning (SciML) offers a promising approach to streamline this process. This study presents a hybrid machine learning framework for permeability prediction in fibrous textiles, integrating surrogate models and physics-informed neural networks (PINNs) to enhance efficiency and accuracy across scales.

Computational and Geophysical Fluid Dynamics (CFD/GFD)

The role of entrainment in axisymmetric tropical cyclones

Tom Dörffel (IAP Kühlungsborn)

The intensification of tropical cyclones (TCs) results from the transport of conserved angular momentum, at least in an axisymmetric context. While there is general agreement on the role of moist cloud convection in driving the system, its precise contribution to intensification remains unclear. Additionally, the mechanisms by which convection facilitates angular momentum transport are still not well understood.
Two prominent but seemingly contradictory explanations for this phenomenon exist in the literature: the Conditional Instability of the Second Kind (CISK) and Wind-Induced Surface Heat Exchange (WISHE). Although these models offer different perspectives, we propose that they represent limiting, asymptotic scaling regimes of the same underlying physical process.
To reconcile these differing views, we use matched asymptotics to combine the three distinct regimes suggested by CISK and WISHE, thus providing a unified framework. Our analysis shows that the transport of angular momentum plays a crucial role in ensuring consistency with the asymptotic matching principle.
Interestingly, this work uncovers a new, previously undocumented pathway for angular momentum transport that may serve as a plausible mechanism for TC intensification. A key element of this process is the special role of the top-of-boundary-layer (BL) inflow, which is closely linked to the entrainment of convective cloud towers.
Through this combined approach, we offer a fresh perspective on TC intensification dynamics, confirming the validity of CISK and WISHE within their respective scopes and reconciling them into a more general theory.

A numerical framework for multiphase flow and heat transport in porous media using Julia computing architecture

Ernesto Meneses Rioseco (Georg-August-Universität Göttingen)

The current energy crisis requires innovative approaches and efficient engineering solutions. Depending on the heat demand, surface infrastructure and local resources availability, the optimal energy mix may considerably contribute to the decarbonization of the energy sector. In particular, we focus in this work on the synergy between natural gas and geothermal energy production from conventional gas fields (porous rock gas traps). By the optimal placement of geothermal wells and most favorable selection of their operational schemes, the observed water breakthrough in gas wells should be substantially delayed, hence augmenting the recovery factor of gas wells and simultaneously producing geothermal energy. Employing the programming framework Julia, we propose an efficient 1/2/3D formulation of multiphase flow and heat transfer in porous natural gas deposits. Employing the finite-volume method implemented in the open-source library VoronoiFVM.jl, we test different formulations of the multiphase flow like the pressure-pressure and pressure-saturation in order to select the most appropriate approach. The computational framework integrates key components such as a solver for coupled nonlinear PDEs, automatic differentiation to assemble Jacobians for the Newton solver, 1/2/3D direct and iterative linear system solvers as well as the WIAS-PDELib Julia packages for pre- and post-processing. Based on realistic geological and geophysical data, we setup compute several scenarios as to how to understand the physical plausibility of our results. We present here our most recent results.

Further MMS Topics at Large

Parameter scan algorithms for particle physics and stress resilience

Mauricio A. Diaz (LIR Mainz)

Within particle physics phenomenology, parameter scan (PS) algorithms are needed to systematically explore the multi-dimensional parameter space of a new physics model and to identify configurations that produce a set of observables matching experimental values. PS methods must address several computational challenges, including high-dimensional parameter spaces, the computational cost of numerical evaluations, and a large number of experimental constraints. To address these challenges, this work introduces a Batched Multi-Objective Constraint Active Search algorithm. Using Gaussian Process as surrogates, we developed a volume-based sampling strategy that maximises diversity and dense filling of satisfactory configurations. The algorithm is available as a modular python library (hep-aid), extending its applicability beyond particle physics. In the future, we plan to apply the search algorithms to computational models of stress resilience, i.e. the ability to function under high stress.

Detection of clinical mastitis in dairy cows by automatic image classification of a visual trap detector

Lukas Minogue (ATB Potsdam)

The detection of clinical mastitis in dairy cow is still a major hurdle for many farms around the world. The classical way of detection by visual inspection of the milk often comes with additional time-consuming work for the farmer. To tackle this and add a reliable method in finding potential mastitis cases, automation of these processes is essential. In this study, we present an image classification algorithm using simple image processing tools and machine learning to score images of a visual trap mastitis detector (Ambic Vision 2000) that was installed in the tubes of the milking system. This algorithm offers an objective classification using visual indicators independent of the human eye and helps with future evaluation of detectors. Furthermore, it shows the potential for automation of the mastitis detection process with a visual trap detector as part of typical milking systems.

Handling Mathematical Research Data

VIVO-based plasma knowledge graph for improving the discoverability of patent information in plasma science and technology

Markus Becker, Ihda Chaerony Siffa (INP Greifswald), joint work with Hidir Aras (FIZ Karlsruhe)

Patents are a vital repository of technical knowledge, yet their complexity often hinders effective utilization in scientific research. This study introduces the knowledge graph Plasma-KG for research in low-temperature plasma science. Plasma-KG is implemented using the open-source software VIVO. Via VIVO’s SPARQL API it is linked with a patent-centric knowledge graph (PKG) leveraging a semantic data model to enhance the accessibility and interconnectivity of patent information. This facilitates semantic annotation of patent texts with context-specific information, e.g. used devices, media, targets and methods. The knowledge graph-based approach supports comprehensive analysis and knowledge reuse, enabling direct links between patents and related research data and scientific literature. While tailored for plasma science and technology, the VIVO-based knowledge-graph approach exemplifies a scalable methodology for creating community-driven knowledge graphs in diverse research domains.

Funding for this research is provided by the Deutsche Forschungsgemeinschaft (DFG), Project Number 496963457.

Hands-on session on the MathModDB model database

Jochen Fiedler (ITWM Kaiserslautern)

In many applied sciences, mathematical models are crucial for representing complex phenomena yet are often unstructured and, therefore, sometimes hard to find and compare. The Mathematical Research Data Initiative (MaRDI) addresses this challenge by developing a comprehensive ontology for FAIR (Findable, Accessible, Interoperable, and Reusable) storage of models. Implemented in the MathModDB Knowledge Graph, the ontology standardizes model features—such as equations, assumptions, and application contexts—and supports straightforward model insertion via a guided interface. In this hands-on session, we demonstrate how to use the implemented MathModDB, i.e. how models are connected, how additional knowledge is generated, and how models can be input.