Work

Here's an overview of my work, mainly publications. Highest is latest.

Machine Learning for Carbon Fiber Reinforced Polymer Production - Dissertation, defended December 2023

Complete study with three different use cases for Machine Learning for (Carbon) Fiber Reinforced Polymers. (Picture above shows all examining profs and me, relieved after the the rigorosum. (From left to right: Prof. Reif, Assoz.Prof. Fauster, me, Prof. Hähner and Prof. Bauer)

Fiber reinforced polymer composites offer a range of properties that are essential for many applications where the highest performance demands are placed on high-end components. These include targeted stiffness combined with lightness, and they are also corrosion resistant. As a result, they have become an integral part in the respective products for aerospace, automotive, construction and sports equipment, among others. However, these outstanding properties come at a price: the production of composite materials is significantly more complex and thus more expensive than, for example, metal or aluminum castings. The complexity is inter alia related to a lower degree of automation, which can be traced back to strongly fluctuating starting materials, especially in the case of textile semi-finished products, and the associated manual work steps. This is the starting point of the present work. Different possibilities for the analysis and optimization of a process family for fiber composites are presented: Liquid Composite Molding. Machine Learning methods are used for this purpose. The following scientific contributions are made: Thus, (1) ways to reconstruct the flow front and detect defects at runtime based on convolutional neural networks are discussed. Furthermore, it is shown how future progress of a process could be based on the course of this injection process in order to be able to counteract if necessary. In addition, the (2) properties of the textile are generated by different variants of neural networks, CNNs, ConvLSTMs and Transformers, based on the progress of each injection process as 2D maps and stored as a digital twin. These maps can be used for post-processing verification of the components or as a quality feature for further processing or life cycle of the product. Finally, (3) Reinforcement Learning, and neural network methods are used to control variants of an injection process so that fewer defects occur. These applications show the possibilities of Machine Learning in the context of fiber composite production, and possibilities how to work (4) with deep neural networks despite data poverty are shown. This includes, above all, sim-to-real transfer learning, in which models are "pre-trained" on data from the simulation and then retrained or "fine-tuned" with a small amount of real data. Thus, deep networks can be preconditioned with large simulative data sets to be fine-tuned to the real case.

OPUS Website, PDF is available there

Control of RTM processes through Deep Reinforcement Learning (2023), presented at the International Conference on Machine Learning and Applications (ICMLA) 2023, Jacksonville, FL

Work on the control of the process based on Deep Reinforcement Learning.

Resin transfer molding (RTM) is a composite manufacturing process that uses a liquid polymer matrix to create complex-shaped parts. There are several challenges associated with RTM. One of the main challenges is ensuring that the liquid polymer matrix is properly distributed throughout the composite material during the molding process. If the matrix is not evenly distributed, the resulting part may have weak or inconsistent properties. This is the challenge we tackle with the approach presented in this work. We implement an online control using deep reinforcement learning (RL) to ensure a complete impregnation of the reinforcing fibers during the injection phase, by controlling the input pressure on different inlets. This work uses this self-learning paradigm to actively control the injection of an RTM process, which has the advantage of depending on a reward function instead of a mathematical model, which would be the case for model predictive control. A reward function is more straightforward to model and can be applied and adapted to more complex problems. RL algorithms have to be trained through many iterations, for which we developed a simulation environment with a distributed and parallel architecture. We show that the presented approach decreases the failure rate from 54 % to 27 %, by 50 % compared to the same setup with steady parameters.

Download PDF (preprint)

Inferring material properties from CFRP processes via Sim-to-Real learning (2023)

Journal paper greatly extending the work from Permeabilitynets (2021). (Accepted at the International Journal of Advanced Manufacturing Technology)

Carbon fiber reinforced polymers provide favorable properties such as weight-specific strength and stiffness that are central for certain industries, such as aerospace or automotive manufacturing. Liquid composite molding (LCM) is a family of often employed, inexpensive, out-of-autoclave manufacturing techniques. Among them, resin transfer molding (RTM), offers a high degree of automation. Herein, textile preforms are saturated by a fluid polymer matrix in a closed mold. Both impregnation quality and level of fiber volume content are of crucial importance for the final part quality. We propose to simultaneously learn three major textile properties presented as a three-dimensional map based on a sequence of camera images acquired in flow experiments. The three properties are fiber volume content and permeability in X and Y direction. Finally, we show how simulation-to-real transfer learning can improve a digital twin in CFRP manufacturing, compared to simulation-only models and models based on sparse real data. The best model, trained on the most realistic simulation data outperforms the same model trained on less sophisticated simulation data by 4 percent points and 0.34 points in intersection over union, more than tripling this metric.

Download PDF (preprint)

ICMLA 2021: PermeabilityNets: Comparing Neural Network Architectures on a Sequence-to-Instance Task in CFRP Manufacturing

Here, we used real data from an RTM-like process for the first time, we predicted maps of permeability for the RTM process in collaboration with Ewald Fauster of the Montanuniversität Leoben.

Carbon fiber reinforced polymers (CFRP) offer highly desirable properties such as weight-specific strength and stiffness. Liquid composite moulding (LCM) processes are prominent , economically efficient, out-of-autoclave manufacturing techniques and, in particular, resin transfer moulding (RTM), allows for a high level of automation. There, fibrous preforms are impregnated by a viscous polymer matrix in a closed mould. Impregnation quality is of crucial importance for the final part quality and is dominated by preform permeability. We propose to learn a map of permeability deviations based on a sequence of camera images acquired in flow experiments. Several ML models are investigated for this task, among which ConvLSTM networks achieve an accuracy of up to 96.56%, showing better performance than the Transformer or pure CNNs. Finally, we demonstrate that models, trained purely on simulated data, achieve qualitatively good results on real data.

Download PDF (preprint)

ACSOS 2021: A Real-World Realization of the AntNet Routing Algorithm with ActivityBots

In this work, my students Jonas Wilfert and Niklas Paprotta used the AntNet routing algorithm with ActivityBots. It was great to use these bots, normally in use for an early bachelor student practical class. Here's the abstract:

To ease teaching self-organizing systems design, we implemented the AntNet routing algorithm for real-world application using educational robots called ActivityBot. Using line sensors and ultrasonic distance sensors, the robotic ants traverse a tiled graph printed on paper, collectively converging to the shortest path. In our descriptions, we address the challenges to face when employing such self-organizing systems on educational hardware and provide a video on YouTube: https://youtu.be/JFduHJ0o0UM

Download PDF (preprint)

ETFA 2021: Genetic Programming for Fiber-Threading for Fiber-Reinforced Plastics

In this work, my student Jonas Wilfert investigated the application of genetic algorithms on a part of the pultrusion process pipeline. Great working together with the Fraunhofer IGCV in Augsburg, namely Frederik Wilhelm. Here's the abstract:

Setting up fiber-threading for a pultrusion line is tedious, error prone and takes a long time. Between 100 and 1000 fibers have to be arranged into a two-dimensional shape, which have to be threaded between several support plates without causing crossovers. When manually planning this process based on intuition, it is hard to keep track of the complexity. This slows the process down to where it can take several hours or several days, and shortening this duration reduces the cost considerably. As planning the setup takes up a large chunk of time, we are proposing a simulation and an algorithm to automatically calculate how the fiber bundles need to be threaded from the creels through the support plates to result in the desired shape. Using a three-dimensional simulation for collision detection in conjunction with a genetic algorithm, we are able to shorten the planning of the fibers to around 10 minutes on a modern 8-core personal computer. Based on this data, further work can be done to further improve, visualize or permanently store the data in a digitized company.

Download PDF (preprint)

Using mesh convolutions and graph neural networks to enable predictions for RTM in 3D space

ECML 2020: FlowFrontNet: Improving Carbon Composite Manufacturing with CNNs

Carbon fiber reinforced polymers (CFRP) are light yet strong composite materials designed to reduce the weight of aerospace or automotive components – contributing to reduced emissions. Resin transfer molding (RTM) is a manufacturing process for CFRP that can be scaled up to industrial-sized production. It is prone to errors such as voids or dry spots, resulting in high rejection rates and costs. At runtime, only limited in-process information can be made available for diagnostic insight via a grid of pressure sensors. We propose FlowFrontNet, a deep learning approach to enhance the in-situ process perspective by learning a mapping from sensors to flow front “images” (using upscaling layers), to capture spatial irregularities in the flow front to predict dry spots (using convolutional layers). On simulated data of 6 million single time steps resulting from 36k injection processes, we achieve a time step accuracy of 91.7% when using a 38 × 30 sensor grid with 1 cm sensor distance in x- and y-direction. On a sensor grid of 10×8, with a sensor distance of 4 cm, we achieve 83.7% accuracy. In both settings, FlowFrontNet provides a significant advantage over direct end-to-end learning models.

Download PDF

ETFA 2020: Towards Real-time Process Monitoring and Machine Learning for Manufacturing Composite Structures

Components made from carbon fiber reinforced plastics (CFRP) offer attractive stability properties for the automotive or aerospace industry despite their light weight. To automate the CFRP production, resin transfer molding (RTM) based on thermoset plastics is commonly applied. However, this manufacturing process has its shortcomings in quality and costs. The project CosiMo aims for a highly automated and costattractive manufacturing process using cheaper thermoplastic materials. In a thermoplastic RTM (T-RTM) process, the polymerization of epsilon-caprolactam to polyamide 6 is investigated using an “intelligent tooling”. Multiple sensor types integrated into the mold allow for tracking of several process-relevant variables, such as material flow and state of polymerization. In addition to the evaluation of the T-RTM process, a digital twin helps to visualize progress and to make predictions about possible problems and countermeasures based on machine learning. In this paper, the combination of software and hardware developments is described which will help to validate an optimal process setup for an industrial CFRP demonstrator.

Download PDF

OCDDC 2018: Transfer Learning for Optimization of Carbon Fiber Reinforced Polymer Production

The main problem that keeps many areas of research from using Deep Learning methods is the lack of sufficient amounts of data. We propose transfer learning from simulated data as a solution to that issue. In this work, we present the industrial use case for which we plan to apply our transfer learning approach to: the production of economic Carbon Fiber Reinforced Polymer components. It is currently common practice to draw samples of produced components statistically and perform destructive tests on them, which is very costly. Our goal is to predict the quality of components during the production process. This has the advantage of enabling not only on-line monitoring but also adaptively optimizing the manufacturing procedure. The data comes from sensors embedded in a tooling in a Resin Transfer Molding press.

Download PDF

Reviewing

I reviewed at several venues, including:

About

My name is Simon Stieber, I received my doctorate in Computer Science at the University of Augsburg, Germany in late 2023. My professional focus is Machine Learning and Deep Learning for industrial applications. I am currently on sabbatical leave and am searching for positions starting August 2024.

Download CV (German)

CV

Group Lead "Artificial Intelligence Methods", Focus Project Management @ University of Augsburg

Augsburg, March 2021 - January 2024

Science Project Management

• CosiMo "Composites for Sustainable Mobility"
• AICUT "Automatisierte Erkennung von Prozessstörungen und Qualitätsschwankungen bei der spanenden Fertigung mittels maschinellen Lernens"

Science Project Acquisition: Networking, project proposals etc.

• FORinfPro "Bayerischer Forschungsverbund Intelligente Fertigungsprozesse & Closed-Loop-Produktion"- accepted, starting March 2024
• GRAIL "Grasping using AI for increased fLexibility" - not accepted

PHD Student @ University of Augsburg

Augsburg, February 2018 - December 2023

Machine Learning for Manufacturing of Carbon Composites. Process Monitoring and Optimization for Resin Transfer Moulding (RTM): Project Website

• See Work for details on my dissertation and papers.
• Deep Learning in a novel field; Transfer Learning for data frugality
• Giving Lectures and Tutorials
• Establishing collaborations with industrial and academic partners to start projects, writing proposals
• Leading group of up to five student assistants in parallel (B.Sc. and M.Sc. students) (Co-)Supervision of 3 master theses, 3 bachelor theses and several seminal works. Overall: 9 Student Research Assistants, 4 Interns, 6 Theses, excluding tutors
• Installation, administration and maintenance of a custom compute cluster: 25+ hosts, including several NVidia DGX systems (A100 and V100); planning and execution of structural measures to accommodate 25+ kW of compute.
• Data Acquisition and Analysis in real life and in simulation
• Parallelized Fluid Simulation in industrial simulation software

Supervised Theses (B/M) and Internships (I)

• Creation of a HDF5-file database for data search and dataset creation, Dennis Hartmann, 2019 (I)
• Optimization of Dataloaders and other Framework-internals, Johannes May, 2019 (I)
• Setup of pipeline and initial Trainings on RTM data, Lukas Lodes, 2019 (I)
• Learning World Models for a simulated Robot Arm through Action-conditioned Frame Prediction, Monika Pichlmair, 2020 (M)
• "Automatische Planung der Faserführung einer Pultrusionsanlage durch einen genetischen Algorithmus" - Automatic planning of fiber threading for a pultrusion process with a genetic algorithm, Jonas Wilfert, 2020 (B)
• "In-situ Überwachung von Composite-Herstellungsprozessen mit tiefen neuronalen Netzen auf der Basis von 3D-Repräsentationen", In-situ supervision of composite processing with deep neural networks based on 3D-representations, Lukas Lodes, 2021 (M)
• "Optimierung der Fließfrontberechnung im Harzinjektionsprozess mithilfe von GANs" - Optimization of flow front detection within Resin Transfer Moulding aided by GANs, Nik Julin Nowoczyn, 2021 (B)
• Sim-to-Real analysis of an industrial RTM process, Matthias Sobotta, 2021 (M)
• Creation of U-Net Based surrogate Model for RTM, Simon Weichselbaumer, 2021 (I)
• Mesh-Based Graph Neural Network Predictor Model for RTM, Leonard Heber, 2021 (I)
• Sim-to-Real Transfer Learning with U-Net Based predictor Model for RTM, Lennart Eing, 2022 (I)
• "Optimierung des RTM-Prozesses zur CFK-Herstellung durch Bestärkendes Lernen", Leonard Heber, 2023 (B)

Master Thesis Student @ Continental

Lindau, April 2017 - September 2017

Crafted my Master's Thesis at Continental in Lindau. The goal is to classify various weather conditions based on camera images. Computer Vision for ADAS / Autonomous Driving.

Graduate Student Research Assistant @ University of Augsburg

Augsburg, Oktober 2016 - Mai 2017

Tutor at a course aimed at autonomous driving. Programming, mostly the integration of a lidar for a model car based on the Audi Autonomous Driving Cup (AADC) 2016 car. Tutoring and teaching master students.

Working Student @ IGEL Technology

Augsburg, September 2014 - August 2016

Helped the R&D team by optimizing the build process, making a custom diff tool in C for different versions of their software, writing a small monitoring tool for the thin clients and more.

Tools and Languages

Programming Languages & Frameworks

Python, Shell, Latex, Java, C, C++, PyTorch, SciKit-Learn, TensorFlow, Keras and more.

Microsoft Office, PAM RTM

Languages

German (Mother tongue), English (C1), Italian (Basic), French (Basic)

Contact

• Mail
• LinkedIn
• ResearchGate
• GitHub

Datenschutzerklärung

Personenbezogene Daten (nachfolgend zumeist nur „Daten“ genannt) werden von uns nur im Rahmen der Erforderlichkeit sowie zum Zwecke der Bereitstellung eines funktionsfähigen und nutzerfreundlichen Internetauftritts, inklusive seiner Inhalte und der dort angebotenen Leistungen, verarbeitet.

Gemäß Art. 4 Ziffer 1. der Verordnung (EU) 2016/679, also der Datenschutz-Grundverordnung (nachfolgend nur „DSGVO“ genannt), gilt als „Verarbeitung“ jeder mit oder ohne Hilfe automatisierter Verfahren ausgeführter Vorgang oder jede solche Vorgangsreihe im Zusammenhang mit personenbezogenen Daten, wie das Erheben, das Erfassen, die Organisation, das Ordnen, die Speicherung, die Anpassung oder Veränderung, das Auslesen, das Abfragen, die Verwendung, die Offenlegung durch Übermittlung, Verbreitung oder eine andere Form der Bereitstellung, den Abgleich oder die Verknüpfung, die Einschränkung, das Löschen oder die Vernichtung.

Mit der nachfolgenden Datenschutzerklärung informieren wir Sie insbesondere über Art, Umfang, Zweck, Dauer und Rechtsgrundlage der Verarbeitung personenbezogener Daten, soweit wir entweder allein oder gemeinsam mit anderen über die Zwecke und Mittel der Verarbeitung entscheiden. Zudem informieren wir Sie nachfolgend über die von uns zu Optimierungszwecken sowie zur Steigerung der Nutzungsqualität eingesetzten Fremdkomponenten, soweit hierdurch Dritte Daten in wiederum eigener Verantwortung verarbeiten.

Unsere Datenschutzerklärung ist wie folgt gegliedert:

I. Informationen über uns als Verantwortliche
II. Rechte der Nutzer und Betroffenen
III. Informationen zur Datenverarbeitung

I. Informationen über uns als Verantwortliche

Verantwortlicher Anbieter dieses Internetauftritts im datenschutzrechtlichen Sinne ist:

Simon Stieber

E-Mail: info@simon-stieber.de

II. Rechte der Nutzer und Betroffenen

Mit Blick auf die nachfolgend noch näher beschriebene Datenverarbeitung haben die Nutzer und Betroffenen das Recht

• auf Bestätigung, ob sie betreffende Daten verarbeitet werden, auf Auskunft über die verarbeiteten Daten, auf weitere Informationen über die Datenverarbeitung sowie auf Kopien der Daten (vgl. auch Art. 15 DSGVO);
• auf Berichtigung oder Vervollständigung unrichtiger bzw. unvollständiger Daten (vgl. auch Art. 16 DSGVO);
• auf unverzügliche Löschung der sie betreffenden Daten (vgl. auch Art. 17 DSGVO), oder, alternativ, soweit eine weitere Verarbeitung gemäß Art. 17 Abs. 3 DSGVO erforderlich ist, auf Einschränkung der Verarbeitung nach Maßgabe von Art. 18 DSGVO;
• auf Erhalt der sie betreffenden und von ihnen bereitgestellten Daten und auf Übermittlung dieser Daten an andere Anbieter/Verantwortliche (vgl. auch Art. 20 DSGVO);
• auf Beschwerde gegenüber der Aufsichtsbehörde, sofern sie der Ansicht sind, dass die sie betreffenden Daten durch den Anbieter unter Verstoß gegen datenschutzrechtliche Bestimmungen verarbeitet werden (vgl. auch Art. 77 DSGVO).

Darüber hinaus ist der Anbieter dazu verpflichtet, alle Empfänger, denen gegenüber Daten durch den Anbieter offengelegt worden sind, über jedwede Berichtigung oder Löschung von Daten oder die Einschränkung der Verarbeitung, die aufgrund der Artikel 16, 17 Abs. 1, 18 DSGVO erfolgt, zu unterrichten. Diese Verpflichtung besteht jedoch nicht, soweit diese Mitteilung unmöglich oder mit einem unverhältnismäßigen Aufwand verbunden ist. Unbeschadet dessen hat der Nutzer ein Recht auf Auskunft über diese Empfänger.

Ebenfalls haben die Nutzer und Betroffenen nach Art. 21 DSGVO das Recht auf Widerspruch gegen die künftige Verarbeitung der sie betreffenden Daten, sofern die Daten durch den Anbieter nach Maßgabe von Art. 6 Abs. 1 lit. f) DSGVO verarbeitet werden. Insbesondere ist ein Widerspruch gegen die Datenverarbeitung zum Zwecke der Direktwerbung statthaft.

III. Informationen zur Datenverarbeitung

Ihre bei Nutzung unseres Internetauftritts verarbeiteten Daten werden gelöscht oder gesperrt, sobald der Zweck der Speicherung entfällt, der Löschung der Daten keine gesetzlichen Aufbewahrungspflichten entgegenstehen und nachfolgend keine anderslautenden Angaben zu einzelnen Verarbeitungsverfahren gemacht werden.

Kontaktanfragen / Kontaktmöglichkeit

Sofern Sie per Kontaktformular oder E-Mail mit uns in Kontakt treten, werden die dabei von Ihnen angegebenen Daten zur Bearbeitung Ihrer Anfrage genutzt. Die Angabe der Daten ist zur Bearbeitung und Beantwortung Ihre Anfrage erforderlich - ohne deren Bereitstellung können wir Ihre Anfrage nicht oder allenfalls eingeschränkt beantworten.

Rechtsgrundlage für diese Verarbeitung ist Art. 6 Abs. 1 lit. b) DSGVO.

Ihre Daten werden gelöscht, sofern Ihre Anfrage abschließend beantwortet worden ist und der Löschung keine gesetzlichen Aufbewahrungspflichten entgegenstehen, wie bspw. bei einer sich etwaig anschließenden Vertragsabwicklung.

Tracking und Logging (IONOS)

Mit welchen Technologien ermittelt IONOS die Daten? Die Daten werden entweder durch einen Pixel oder durch ein Logfile ermittelt. Zum Schutz von personenbezogenen Daten verwendet WebAnalytics keine Cookies. Die IP des Besuchers wird bei der Übermittlung eines Seitenabrufes übertragen, nach der Übermittlung direkt anonymisiert und ohne Personenbezug verarbeitet.

Welche Daten speichert IONOS von meinen Websitenbesuchern? IONOS speichert keine personenbezogenen Daten von Websitenbesuchern, damit keine Rückschlüsse auf die einzelnen Besucher gezogen werden können. Es werden folgende Daten erhoben: Referrer (zuvor besuchte Webseite) Angeforderte Webseite oder Datei Browsertyp und Browserversion Verwendetes Betriebssystem Verwendeter Gerätetyp Uhrzeit des Zugriffs IP-Adresse in anonymisierter Form (wird nur zur Feststellung des Orts des Zugriffs verwendet) Zu welchem Zweck werden die Daten erhoben? In WebAnalytics werden Daten ausschließlich zur statistischen Auswertung und zur technischen Optimierung des Webangebots erhoben.

Werden Daten an Dritte weitergegeben? Es werden keine Daten an Dritte weitergegeben.

Muster-Datenschutzerklärung der Anwaltskanzlei Weiß & Partner zzgl. Informationen zu IONOS tracking. Quelle 1&1 IONOS

Elements

i = 0;

while (!deck.isInOrder()) {
    print 'Iteration ' + i;
    deck.shuffle();
    i++;
}

print 'It took ' + i + ' iterations to sort the deck.';