SDL Materials Design

To the overview of the Simulation and Data Labs

A cooperation of TU Darmstadt
and RWTH Aachen University

Simulation and Data Lab

The SDL Materials Design uses advanced computational methods to design better materials and optimize their mechanical and functional properties.

To this end we adapt deep learning methods and numerical algorithms to our data and evaluate them on materials specific metrics. The final goal is targeting material scientists from both experimental and computational background and providing cross-sectional showcases to bring them together.

Using high resolution panoramic images from Scanning Electron Microscopy (SEM) allows to have a large field of view of the microstructure. We use convolutional neural networks (CNNs) in tasks such as semantic segmentation, object detection and classification to find and locate damages and defects.

Further, we use the segmented images of microstructure as input to generate Finite Element (FE) mesh with interface details and carry out multiphase Finite Element (FE) simulations to evaluate the elastoplastic and fracture behavior of the samples in the framework of computational homogenization.

In the future, we plan to apply 3D convolutional models to SEM images in depth of the material to unravel the full nature of microstructures and damages. Additionally, Atom Probe Tomography (APT) which has the ability to explore the chemical composition of materials at near-atomic resolution will be used. We use deep learning to find good representations for the input 3D discrete data from APT, to enable them to perform tasks such as 3D segmentation and classification. The final goal of this analysis is to detect phase transformations in materials on an atomic scale.

The SDL Materials Design benefits from the open-source visualization software OVITO, commonly used to analyze the 3D data from atomistic simulations. The plan is to extend its functionality and adapt the existing algorithms to experimental data.

Contact the SDL Materials Design!

Competencies

Atomistic and large-scale simulations
Machine Learning:
Interatomic potentials for large-scale simulations
Process structure-property correlations
Phase identification and segmentation
Multi-physics simulations and application packages (MOOSE)
Visualization and analysis (OVITO)

Current research topics:

Deep learning models for microstructural analysis of electron microscopy images
Finite element simulation on deep learning segmented microstructures as input
Deep learning models for 3D data from atom probe tomography
Robust structure identification for experimental and simulation data using machine learning

Activities

Providing an interactive web app for material scientists to use our models
Finding solutions for the community of materials scientists for large scale visualization within the OVITO software
Providing lectures on computational materials science to interested students
Gitlab/Github courses for material scientists
Organizing Hackathons for students and the community
Training Courses in 2025:
- Community Workshop : Materials Science with Advanced Data Management and Data Science Techniques
- Machine Learning for Materials Science

Gallery

Phase identification and segmentation of atom probe tomography (APT) data: ML-based methods can improve automation and reduce the need of expert knowledge. This enables more accessible usage of APT.

MD simulation of silicon monoxide formation: By means of ML interatomic potentials it is shown that silicon monoxide is a two-phase system in the solid state. This result is most relevant for applications of Si-O as energy material.