AI-based Generation of Simulation Models for Material Development from CT Data

In contemporary materials science, the accurate modeling of material structures is a fundamental step in predicting material properties, particularly for composite materials, which often exhibit complex internal structures. Traditional approaches to generate geometry models (GMs) for such purposes rely heavily on 3D imaging techniques like computed tomography (CT), where the material’s internal structure is imaged and segmented based on grayscale intensity. However, this method faces significant challenges. Limited image resolution and inherent artifacts can compromise the accuracy of these models, leading to discrepancies between the GM and the actual material structure. An alternative approach involves generating synthetic models based on statistical data derived from real material structures. While this method allows for deliberate variation and a better understanding of structure-property relationships, it requires a substantial number of generation parameters to accurately reflect the diversity found in real material structures. 

In the research project, KI4MaterialModeling, funded by the German federal ministry of education and research, introduces an AI-powered workflow designed to enhance GM creation from CT data, targeting both efficiency and accuracy to meet the needs of advanced material development. In this talk, we will present our research project, its workflow, AI tools, use case, and further results.

The primary objective of KI4MaterialModeling is to develop a comprehensive, AI-based approach for generating precise GMs from CT data, with an emphasis on fiber-reinforced polymer composites as a representative case study. The methodology builds on three interconnected AI-driven components: (1) the generation of synthetic CT data containing realistic imaging artifacts, (2) a specialized AI model for data enrichment that enhances CT data quality, and (3) an AI-based object recognition system to identify structural features within CT datasets. Together, these components form a cohesive workflow that advances current GM-generation capabilities.

The first stage of the workflow involves producing synthetic CT data that closely mirrors the quality and limitations of actual CT scans, including common imaging artifacts. By simulating realistic conditions, this synthetic data enables the AI model to be trained and tested in scenarios that closely resemble real-world data, which ultimately improves its robustness and adaptability. The second stage leverages AI for data enrichment, where the algorithm augments CT data by enhancing resolution and minimizing noise or distortions, thus addressing the limitations imposed by traditional imaging constraints. The final stage utilizes object recognition algorithms that can automatically detect and classify features within the CT data, enabling the extraction of high-quality structural information.

This AI-powered workflow brings several innovations to materials modeling. By enabling automated structure information extraction from CT data, the approach enhances the efficiency and accuracy of GM generation. Notably, it facilitates a hybrid modeling technique that combines real and synthetic CT data, resulting in high-resolution GMs that overcome the limitations of individual data sources. This synergetic combination enables the generation of complex hybrid models that maintain a high fidelity to real material structures while incorporating the versatility of synthetic data-based variations.

The application of this workflow in fiber-reinforced polymer composites exemplifies its potential for widespread adoption across other material classes, including filtration media, battery materials, and geological samples. These fields, all requiring high-resolution and accurate structural modeling, stand to benefit substantially from the workflow's efficiency and adaptability. The scalability and transferability of this AI-driven approach underlines its value in accelerating material innovation and optimizing the design and development of novel materials. Ultimately, KI4MaterialModeling supports a digital transformation in material science which not only enhances the competitive advantage of industries but also meets the growing societal demand for advanced, high-performance materials across diverse applications. This project represents a significant advancement in the integration of AI with composite material science, setting a new standard for precision and efficiency in GM generation.

In this talk we will the application of these techniques applied on a variety of different typical composite material setups achieved together with our project partner from Leibniz-Institute for Composite Materials (IVW). These include standard woven fabrics and non-crimp fabrics from glass and carbon fibers. We will show how these approaches help in creating statistical digital twins of real-life materials and start varying parameters from there to create future materials with the desired properties during manufacturing, usage and recycling of composites materials.