Suitable material solutions are of key importance in designing and producing components for engineering systems – either for functional or structural applications. Materials data are generated, transferred, and introduced at each step along the complete life cycle of a component. A reliable materials data space is therefore crucial in the digital transformation and an important prerequisite for machine learning in materials science. Therefore, the consortium NFDI MatWerk aims to develop a sustainable infrastructure for the standardized digital representation of materials science and engineering (MatWerk). The goal is to seamlessly integrate decentralized data and metadata, experimental and computational workflows, and a materials ontology to maximize interoperability and reproducibility of materials data processing. To this end, data use profiles of participant projects from different sub-disciplines are analyzed to identify the most relevant scientific scenarios within MatWerk. Similarly, the Plattform MaterialDigital (PMD) is committed to provide a prototypical infrastructure for the digitalization of materials in an industrial context implemented by decentralized data servers, semantic data schemas and digital workflows. The standards, methods, and tools developed within the PMD are deployed and consolidated within the context of currently more than 20 BMBF-funded academic and industrial research consortia. Scientific workflow environments represent a major focus area, including efforts to improve the definition and representation of digital workflows, as well as their distribution in form of a workflow store. In this presentation we will describe the overarching visions behind these initiatives, their status, and progress of dissemination with a focus on the workflow activities and the connection to machine learning applications. Following the philosophy of both consortia, specific examples will be used to demonstrate innovative and pragmatic solutions.
Graphene oxide (GO) materials, while promising for a variety of applications, can be difficult to fully understand and predict its properties due to the highly irregular atomic structure. The presence of several oxygen functionalizations across the surface can greatly change depending on multiple factors. Techniques that can help predict the structure of systems, such as x-ray absorption spectroscopy (XAS) can provide an insight into characterizing the composition and structure of materials. However, for GO experiments can be hard to properly characterize due to the multiple chemical environments in the sample. First-principle simulations can help provide a route to possibly breakdown the spectra to its individual chemical environments, but the computational cost to perform calculations, especially on structures of a large enough size can be expensive. Using a graph neural network (GNN) implementation, an initial data set of 300 molecules representative of GO with time-dependent density functional theory (TDDDFT) calculated spectra was used to train a machine learning model. This model can help predict spectra with accuracy and help identify functional groups found in GO at a much smaller computational cost.
The Helmholtz-Zentrum Hereon is operating imaging beamlines for X-ray tomography (P05 IBL, P07 HEMS) for academic and industrial users at the synchrotron radiation source PETRA III at DESY in Hamburg, Germany. The high X-ray flux density and coherence of synchrotron radiation enable high-resolution in situ/operando/vivo tomography experiments. Here, large amounts of 4D data are collected from a wide variety of samples, which is challenging to reconstruct, process, and analyze. In this talk, we will give an overview of the application of modern machine learning methods to data processing and analysis of synchrotron-radiation tomography experiments for biodegradable implant materials, such as segmentation, denoising, multi-modal imaging, phase retrieval, and digital volume correlation.