Knowledge brought to life
Knowledge brought to life
Fraunhofer IWU’s Multimodal AI Assistant for Shape Memory Alloys
Shape memory alloys (SMAs) are among the most fascinating materials in modern engineering. Once deformed, they can “remember” and return to their original shape simply by being heated. These metallic “muscles” enable silent actuators, highly flexible medical implants, adaptive aerospace components, and much more.
However, designing SMA systems is far from straightforward. It requires in-depth materials science expertise in phase transformations, crystal structures, and thermomechanical cycles. To transfer this valuable knowledge from research laboratories into industrial practice more quickly and make it more accessible, the Fraunhofer Institute for Machine Tools and Forming Technology IWU has developed a pioneering demonstrator: an interactive virtual expert.
The Virtual Expert
The newly developed SMA AI assistant is available to developers and project partners via a web interface, allowing them to engage directly in expert discussions. What sets the system apart from conventional AI chatbots is its ability to communicate knowledge multimodally.
When users ask questions about specific SMA phenomena—such as component distortion or the precise temperature range of a phase transformation—the AI does more than provide technically accurate text. It also automatically retrieves and presents relevant diagrams, microstructure images, and experimental setups from the institute’s own knowledge database.
Behind the Scenes: Multimodal RAG and an Agile IT Architecture
The system’s technological foundation demonstrates Fraunhofer IWU’s strong AI expertise in application-oriented solutions.
Multimodal RAG (Retrieval-Augmented Generation)
Unlike conventional language models that rely on static general knowledge, this chatbot works like an expert taking an “open-book exam.” For every query, it searches a secure database—a vector store—in real time for the institute’s latest research reports.
Intelligent Document Analysis
Using advanced parsing methods and tools such as Docling, the system breaks down scientific PDFs according to their structure. Text passages, tables, embedded figures, and their captions are captured separately. This allows the system to understand visual content in context, associate it with the relevant information, and display it directly in the chat when needed.
Data Sovereignty Through Open Source and Low Code
The entire back-end workflow is built on the flexible n8n automation platform. Combined with locally hosted language models, this architecture provides maximum data security. Sensitive technological expertise remains entirely on the institute’s servers and is not shared with third-party providers.
Small and medium-sized enterprises (SMEs), in particular, are often deterred by the extensive research required to introduce innovative material technologies such as SMAs. Fraunhofer IWU’s virtual assistant closes this gap. It acts as a digital mentor, making complex scientific concepts immediately understandable and easier to grasp through visualizations such as measurement curves and engineering drawings.
“With this system, we are demonstrating how artificial intelligence can accelerate the transfer of cutting-edge research into industry,” the Fraunhofer IWU project team explains. “We combine our many years of materials expertise in shape memory technology with modern software architecture to deliver genuine added value to companies—quickly, securely, and visually.”
bjoern.senf@iwu.fraunhofer.de
