DLR Foundation Models

The rapid development of generative artificial intelligence AI is fundamentally changing the world of work and information processing, automating routine tasks and also becoming relevant for complex tasks in software development and scientific research. There is also a wide range of potential applications for DLR. However, general foundation models (FMs) are currently not well suited for domain-specific scenarios due to their limitation to unspecific training data. In order to make large FMs usable for DLR applications, methods for adapting large AI models to specific tasks are being developed in this project and made available as a framework in the form of a service architecture and the necessary infrastructure. This enables the synergetic development of specialised applications. Part of this synergy project is therefore also the selection and monitoring of promising use cases from aviation, aerospace and transport, which are then trailed in application projects. The establishment of training concepts and workshops for employees in various areas of DLR is also intended to increase DLR's overall expertise in dealing with generative AI.

DLR Foundation Models
Project structure DLR Foundation Models

In this project, the DLR Institute of Data Science is working on issues relating to the adaptation of foundation models for special applications. The aim is to provide lightweight models that are specialised only in specific capabilities of general foundation models and thus require fewer computing resources and have a lower carbon footprint. To this end, approaches to quantization, knowledge distillation and model pruning are being pursued.

Special applications here are for example

    • AI assistants in space travel
    • Digital assistants in the cockpit and in air traffic
    • Robust interaction with virtual components of future transport infrastructure
    • Obtaining information from multimodal earth observation data

Project duration: 07/2024 - 12/2026