Integrating Conversational AI Agents with Digital Twins: A Systems Engineering Approach to Complex Infrastructure Management and Predictive Decision-Making

Abstract

Background: Managing complex infrastructure increasingly requires predictive, adaptive, and human-centered systems. Traditional approaches often struggle with operational complexity, fragmented data, and high technical barriers. Methods: This study presents a TRL4 proof of concept integrating a conversational AI agent with a user-adaptive digital twin for occupancy forecasting. Users can upload their own datasets, and dynamically configure prediction models (ARIMA, SARIMA, Random Forest, XGBoost) based on input variables such as occupancy or demand drivers. The AI agent, powered by Gemini 2.5 Flash Lite, functions as an orchestration layer, translating natural language instructions into data ingestion, model execution, and query actions. While the digital twin supports additional variables (energy, water, waste), these are envisioned for future work and were not part of the current validation. Results: Functional validation confirmed the system’s capability to interpret user intentions accurately, adapt model training to the characteristics of user-provided data, and present results through convenient and comprehensible visualization methods. The integrated architecture demonstrated stable performance across multiple validation scenarios, achieving satisfactory prediction accuracy (within expected ranges for TRL 4). Conclusions: This work validates the technical and functional viability of integrating conversational AI agents with digital twins as an emergent system of systems, extending beyond conventional predictive pipelines by enabling context-specific modeling. The systems engineering approach reveals how such integration transforms reactive infrastructure management into proactive, data-driven, and human-centered decision-making processes, establishing a foundation for future developments toward higher technology readiness levels.