Design and Evaluation of Next-Generation Cellular Networks through Digital and Physical Open and Programmable Platforms

TL;DR

This paper presents new experimental platforms and methodologies for designing, testing, and validating open, programmable 5G/6G RAN architectures, emphasizing digital twins, GPU acceleration, and AI integration.

Contribution

It introduces the Colosseum digital twin, X5G testbed, and tools for automated scenario creation, enabling realistic, scalable testing of open RAN solutions with AI and GPU acceleration.

Findings

Colosseum digital twin closely reproduces real-world environments

X5G integrates GPU-accelerated PHY with OpenAirInterface

Real-time RAN inference enables sub-millisecond control loops

Abstract

The evolution of the Radio Access Network (RAN) in 5G and 6G technologies marks a shift toward open, programmable, and softwarized architectures, driven by the Open RAN paradigm. This approach emphasizes open interfaces for telemetry sharing, intelligent data-driven control loops for network optimization, and virtualization and disaggregation of multi-vendor RAN components. While promising, this transition introduces significant challenges, including the need to design interoperable solutions, acquire datasets to train and test AI/ML algorithms for inference and control, and develop testbeds to benchmark these solutions. Experimental wireless platforms and private 5G deployments play a key role, providing architectures comparable to real-world systems and enabling prototyping and testing in realistic environments. This dissertation focuses on the development and evaluation of complementary experimental platforms: Colosseum, the world's largest Open RAN digital twin, and X5G, an open, programmable, multi-vendor private 5G O-RAN testbed with GPU acceleration. The main contributions include: (i) CaST, enabling automated creation and validation of digital twin wireless scenarios through 3D modeling, ray-tracing, and channel sounding; (ii) validation of Colosseum digital twins at scale, demonstrating that emulated environments closely reproduce real-world setups; (iii) X5G, integrating NVIDIA Aerial GPU-accelerated PHY processing with OpenAirInterface higher layers; (iv) a GPU-accelerated dApp framework for real-time RAN inference, enabling sub-millisecond control loops for AI-native applications including ISAC; and (v) intelligent RAN applications spanning spectrum sharing, interference detection, network slicing, security, and CSI-based sensing. Overall, this dissertation provides an end-to-end methodology bridging digital and physical experimentation for next-generation cellular networks.