Digital Twins through the lens of machine learning

Research on digital twins is heterogeneous and covers numerous different real-world tasks in various domains. At the same time, machine learning has repeatedly demonstrated its potential as disruptive technology and has impacted many research fields. However, currently these two disciplines are somewhat disconnected, and key enabling machine learning technologies with high potential impact for digital twins remain niche topics. This review article presents digital twins from a machine learning perspective to encourage interdisciplinary research and to permit knowledge exchange between these two fields. We discuss the conceptual differences between digital twins in Earth sciences, healthcare and industry. Unlike previous reviews, we focus on machine learning-specific aspects such as data landscape, learning assumptions and objective functions for each domain. We systematically review 50 recent machine learning-based digital twins. We give a general, learning theory-based definition of digital twins to serve as starting point for theoretical research on machine learning for digital twins. Then, we describe all major aspects of the cyclical flow of information in digital twin systems on a domain-agnostic level, including data acquisition (measurement), pre-processing, data management, simulation, forecasting, control, intervention testing, and human-computer interaction.