6G networks will be an extremely complex system, characterized by a pervasive deployment of metasurfaces and holographic surfaces, complex signal waveforms for joint communication and sensing, and the use of THz spectrum. This makes the analysis, design, and operation of future 6G networks more challenging than any previous wireless generation:

1) Optimization complexity.  Any accurate model of 6G wireless networks will be too complex to be optimized by traditional mathematical optimization techniques. Moreover, the optimization of radio resources is subject to stringent time requirements, since it must be updated in real-time as the network operates, in order to adapt to the variations of the wireless channels. This is especially true in high-mobility scenarios like vehicular networks. For the last few years, machine learning has been used as a tool to reduce the complexity of wireless network design. However, Machine Learning architectures applied in wireless networks are typically just inherited from other fields (e.g. computer vision) and are blindly applied to wireless. This results in poor scalability, low generalization capabilities, and a large performance gap compared to mathematical optimization models.
A promising way of overcoming this challenge lies in overcoming the traditional blind use of machine learning architectures inherited from other fields of science, and embedding machine learning architectures with specific context-aware information about the wireless network to be designed. Prior knowledge about 6G networks provided by theoretical models of electromagnetic propagation and wireless signaling must be embedded into machine learning architectures. This provides physics-based principles to guide machine learning architectures towards optimized design policies with reduced computational complexity. Leveraging machine learning for 6G networks requires an unparalleled paradigm shift in how machine learning is applied to wireless systems, abandoning its routine blind application inherited from other fields of science and empowering it with essential domain knowledge on 6G technologies. 

2) Data Acquisition. Machine learning methods need a large amount of training data, and physics-based machine learning is no exception. The acquisition of large datasets is problematic since it would require expensive measurement campaigns in large-scale testing sites. The way this is typically dealt with is by the use of synthetic data, i.e. computer-generated data based on a model of the wireless network to design. However, at present, no model exists for 6G networks and only preliminary studies for individual 6G technologies are available. Moreover, a traditional network simulator would not be effective, because it would only provide information on network operation in a static operating condition, defined by the operator that programmed the simulator. Instead, a dynamic simulator is needed, which modifies its behavior in real-time mimicking the behavior of the actual wireless network.
A recent breakthrough with the potential of securing enough data for machine learning architectures is digital twinning. The concept of digital twinning is that of developing a virtual replica of a physical system, called digital twin, which mirrors in real-time the behavior of the physical system. This goes significantly beyond traditional approaches to system simulation. Traditional simulators are static, in the sense that they replicate the way a system operates given the operating conditions specified by their program, and their behavior will not change unless their programming is manually updated. Instead, a digital twin is dynamic, because its behavior will automatically change based on the real-time evolution of the network parameters, without any external intervention by a human operator. A digital twin of a wireless network will have an interface with the physical network, through which it receives the time-varying network parameters (e.g. positions of the mobile nodes and objects in the environment, number of users, users’ requirements, etc.) and based on this information, as well as on a preset theoretical model of the system to emulate, it automatically adjusts to replicate in real-time in the virtual world what is happening in the physical world. A digital twin of 6G wireless networks would solve the data acquisition problem, enabling the generation of as much accurate data for machine learning training as needed, without having to perform expensive measurement campaigns and having to build large-scale testing sites. 

TWIN6G will synthesize the world’s first digital twin 6G network-level emulator, which integrates holistic models and physics-based machine learning designs for dynamic, real-time 6G networks emulation.