Informed Machine Learning and Visual Analytics for Virtual Spinning Processes

This doctoral thesis aims to establish the role of machine learning and visual analytics in the industrial production of nonwovens, aiming for quality and resource optimization through economical and sustainable production. To accomplish this, it addresses various challenges faced by numerical simulation models that act as virtual twins of physical processes. The thesis emphasizes the benefits of integrating engineering, numerical analysis, and artificial intelligence within an industrial application. The melt spinning process is selected as a key use case, being a primary and critical step in nonwoven production. In this context, the thesis investigates and addresses the bottlenecks encountered by a virtual spinning simulator, which relies on boundary value problem solvers to model the dynamics of the spinning process through a system of differential equations with boundary conditions. The thesis makes three main contributions. First, it introduces a machine learning pipeline to automate the optimization of BVP solver settings, enhancing convergence and computation speed. Second, it presents a physics-informed machine learning approach to accelerate BVP continuation problems by providing high-quality solution approximations. Third, it integrates visual analytics to interpret simulation results, evaluate ML predictions, quantify uncertainty, and improve the interpretability of physics-informed neural networks—thereby fostering trust and usability. This thesis leverages the complementary capabilities of numerical simulations, machine learning, and visual analytics in the context of scientific machine learning to address complex industrial spinning processes. While melt spinning serves as a compelling and practical demonstration of these techniques, the methods are also applied to the simulation of fiber deposition in nonwoven manufacturing, showcasing their general applicability. The thesis concludes by discussing the broader industrial impact and future directions in ML-augmented simulations and human-centered visual analytics.