Finding partners for HORIZON-CL4-INDUSTRY-2025-01-DIGITAL-61: AI Foundation models in science

The increasing complexity of environmental and analytical processes in the fields of biology and chemistry requires innovative digital support systems. This project aims to develop a multi-agent system (MAS) that supports researchers through intelligent task sharing, data-driven analyses and knowledge retrieval.

At the centre of the system is a coordinator agent that acts as the central control unit. It analyses incoming requests, decomposes them into manageable subtasks, and assigns them to specialised agents. Each specialised agent is designed for a specific function, such as data analysis, mathematical computation, chemical structure recognition, or knowledge extraction. These agents register their expertise with the coordinator, enabling it to efficiently distribute tasks. The results of each agent’s computations are iteratively sent back to the coordinator, which consolidates them and determines if further processing is required before delivering a final response to the user.

The agents are powered by Large Language Models (LLMs), which serve as their reasoning and communication backbone. Additionally, the agents have tool-calling capabilities, allowing them to interact with external environments, perform complex calculations, and retrieve relevant scientific data. These tools include database access for structured information retrieval, API-based knowledge extraction, web queries for real-time data acquisition, and even neural networks that provide supplementary computational capabilities. The system also incorporates a Retrieval-Augmented Generation (RAG) framework, which enables the agents to query domain-specific knowledge bases, scientific literature, and experimental results, thereby ensuring that responses remain contextually accurate and up to date.

A key component of the system is the Human-in-the-Loop (HITL) mechanism, which allows users to provide continuous feedback. By reviewing and correcting outputs, human experts refine the system’s accuracy and efficiency, fostering a self-improving cycle of learning. Over time, this feedback-driven adaptation enhances the MAS's ability to handle increasingly complex scientific inquiries.

The project represents a new type of adaptive and collaborative architecture for use in scientific laboratories. Potential use cases include TLC (thin-layer chromatography), tumour research, and battery recycling. This approach is intended to increase the efficiency and precision of scientific processes and create an interactive research assistant.

We are currently a group consisting of the following members:

- OTH Amberg-Weiden (Natural Language Processing, Computer Vision, AI)

- Fraunhofer ISC (Laboratory Automation, Tumour Research Group, Battery Cell Recycling)

- Danish Technological Institute (Battery Cell Recycling)

- Tallinn University of Technology (Control Systems, Digital Twins, AR/VR/XR)

- Jozef Stephan Institute (Robotics)

- SynHet (start-up for TLC use case)