Fraunhofer Institute for Mechanics of Materials IWMMaterials with tailored, reproducible properties, manufactured efficiently. The project developed AI methods for data-driven materials design that identify optimal microstructures based on desired materials properties and automatically control the corresponding manufacturing process.
Tailored Materials Properties through Microstructure Optimization: Machine Learning Methods for Modeling and Inverting Structure-Property Relationships and Their Application to Sheet Metal Materials
Completed research project
Project description
Market demand and challenges
The demand for customized, high-performance materials remains strong. However, the targeted adjustment of materials properties via microstructure is complex: The causal chain from process control through microstructure to properties is high-dimensional and nonlinear. Traditional approaches reach their limits here. What is needed are methods that can model and invert this relationship in a data-driven manner - that is, calculate the optimal manufacturing process based on the desired property profile.
Project approach
The project developed a holistic, ML-based optimization approach spanning the entire process–structure–properties chain. At its core is a Siamese multitask learning network that models and inverts structure–property relationships, thereby identifying microstructures that match desired properties. This approach is complemented by reinforcement learning algorithms for adaptive process control, as well as active learning methods for the efficient generation of training and design data. All methods were validated using the example of texture optimization in sheet metal materials.
Contribution to addressing the challenges
For the first time, the developed methods make it possible to automatically determine a set of suitable target microstructures based on a desired property profile and to specifically control the manufacturing process to produce them. By considering multiple equivalent target structures, the process can always be guided toward the structure that is easiest to achieve. Active learning strategies significantly reduce data requirements. The generic approaches are transferable to any materials class and process.
Fraunhofer IWM’s work packages in the project:
Focus area 1: Data foundation and microstructure characterization
Based on representative samples of microstructure-property data points for sheet metal materials, Fraunhofer IWM created validated datasets for cold-rolled steels to serve as the training and evaluation basis for all downstream ML models.
Focus area 2: ML models for structure-property relationships and materials design
Fraunhofer IWM implemented and trained the Siamese multitask learning network for modeling and inverting structure-property relationships and developed a model that generates a diverse set of manufacturable microstructures for given target properties.
Focus area 3: Efficient data generation through active learning
Using an active learning approach with an integrated region filter for the efficient exploration of the microstructure-property space, Fraunhofer IWM designed a sampling strategy that covers larger ranges of extreme properties with significantly fewer simulations than standard methods such as Latin Hypercube Design.
Based on the project results, Fraunhofer IWM can offer industrial companies the following research and development services:
1. Data-driven modeling of structure-property relationships
The goal is to generate digital representations of material systems that companies can quickly analyze for materials design, quality prediction, and process optimization.
2. Inverse materials design: microstructures for desired properties
Based on a specified property profile, a targeted selection of manufacturable microstructure candidates that meet a company’s requirements is identified.
3. AI-supported adaptive process control with reinforcement learning
Fraunhofer IWM develops and tests reinforcement learning-based algorithms for model-free, adaptive control of manufacturing processes. Industrial companies benefit from a learning-based process control system that automatically adapts to fluctuations and significantly reduces scrap and the number of iterations required during process design.
4. Structure-guided process path optimization
Companies receive data-driven process recipes that are specifically tailored to the desired materials properties - without the need for time-consuming experimental parameter searches.
5. Active-learning-based exploration of materials and process spaces
Fraunhofer IWM employs advanced active-learning strategies to efficiently explore the relevant property space of a given material and/or process with minimal simulation and experimental effort, as well as to identify regions of extreme properties.
6. Transfer of methods to new materials classes and processes
Fraunhofer IWM adapts the generic ML methods developed in the project to the specific materials systems and manufacturing processes of the industrial partner - ranging from dual-phase steels and aluminum alloys to additively manufactured components.
