Digital twins as self-models for intelligent structures

Area of Science:

• Artificial Intelligence
• Computer Science
• Digital Twins

Background:

• Self-models in AI are crucial for self-awareness and adaptation.
• Existing architectures may lack dynamic, integrated self-representation.
• Agent-based systems offer modularity for complex AI construction.

Purpose of the Study:

• To develop a 'digital twin self-model' using an agent-based architecture.
• To represent complex AI behaviors through dynamic assembly of digital components.
• To integrate a self-model within a knowledge graph for dynamic information management.

Main Methods:

• An agent-based architecture with heterogeneous digital components was employed.
• A knowledge graph was utilized to encode the self-model and manage information.
• Retrieval augmented generation (RAG) combined a local knowledge graph with a large language model for query responses.
• A web-based user interface was developed for visualization and natural language querying.

Main Results:

• A functional digital twin self-model was demonstrated using a small-scale building example.
• The architecture supports multiple operational modes, including an offline workflow execution.
• The dynamic knowledge graph successfully encoded the self-model and facilitated information retrieval.
• Natural language queries were processed effectively using RAG.

Conclusions:

• The agent-based digital twin architecture provides a robust framework for creating self-aware AI.
• Dynamic knowledge graphs are effective for encoding and managing self-model information.
• Integration with large language models enhances the interactive capabilities of AI self-models.