DeepSig recently participated in ORAN PlugFest Fall 2025, demonstrating a first-of-its-kind approach to wireless channel emulation that combines direct RF measurements, generative AI modeling, and hardware-in-the-loop testing. The work highlights a shift in how the wireless industry can evaluate, test, and optimize systems under realistic conditions.
The Challenge with Standardized Channel Models
For decades, wireless system testing has relied on standardized channel models, including the 3GPP Tapped Delay Line (TDL) specifications defined in TR 38.901. While these models serve their purpose for compliance testing, they have significant limitations. They are mathematical abstractions rather than learned from real deployments; they fail to capture site-specific propagation effects, and they cannot fully represent the complexity of real-world environments.
As a result, a persistent gap exists between laboratory testing and field performance. Systems that pass compliance tests may still struggle in specific deployment scenarios such as urban canyons, industrial facilities, or complex indoor environments, where propagation conditions differ substantially from those in standardized models. Neural receivers trained on a small set of TDL scenarios, in particular, face challenges generalizing to real-world conditions.
The PlugFest Test Environment
Our PlugFest demonstration brought together four key partners:
- DeepSig: Generative AI channel modeling technology
- Keysight Technologies: PROPSIM channel emulator hardware
- srsRAN: Open-source 5G CU/DU software stack
- POWDER Lab (University of Utah): A large-scale testbed developed as part of the National Science Foundation’s (NSF’s) Platforms for Advanced Wireless Research (PAWR) program and designated by the O-RAN ALLIANCE as an Open Test and Integration Center (OTIC)
The system under test was an end-to-end 5G SISO link consisting of a commercial Quectel UE connected to a gNB built from an ORAN-compliant Benetel O-RU running the srsRAN stack. Rather than over-the-air transmission, the RF link passed through the Keysight PROPSIM channel emulator, allowing precise control over propagation conditions.
AI-Generated Channel Models from Real Measurements
At the core of thedemonstration is a generative AI approach based on normalizing flows, a class of models that learn statistical distributions directly from data. Unlike traditional channel models defined by mathematical equations, our flow-based models are trained directly on measured channel data collected in real-world environments.
The double-additive-coupling layer architecture is used in our normalizing flow model. The functions F and G are neural networks that learn to transform between the measured channel distribution and a simple Gaussian distribution.
Here’s how it works:
Real-world channel measurements are captured opportunistically using commercial LTE downlink signals. The measurement system leverages cell-specific reference signals as sounding waveforms, enabling continuous channel estimation during drive tests.
A normalizing flow model is then trained on the measured data to learn the underlying statistical distribution of the channel. Through a series of fully invertible transformations, the model maps complex, real-world channel statistics to a simple Gaussian distribution.
To generate new channels, samples are drawn from the Gaussian distribution and passed through the inverse transformation, producing channel realizations that statistically match the original measurements.
These generated channel impulse responses are formatted for the PROPSIM and loaded into the emulator, where they modulate the RF signal between the user equipment and the gNB.
Drive-test route in the Ballston neighborhood of Arlington, VA. Each colored point represents channel measurements from a different cell tower (PCI). This 65-minute drive captured 145,992 channel measurements from 12 unique towers.
Why This Approach Changes Wireless Testing
The combination of real RF measurements, generative AI modeling, and hardware-in-the-loop testing opens up entirely new possibilities:
Site-Specific Digital Twins Network operators can capture measurements from deployment locations and create AI models that faithfully reproduce those conditions in the lab. This enables testing and optimization specifically for where systems will actually operate.
Massive Data Compression Our flow model compresses approximately 16.8 GB of measured channel data into a 21 MB model file—an 800x reduction. This means site-specific channel conditions can be packaged and distributed to test labs worldwide without shipping terabytes of raw measurements.
Proven Performance Benefits: In our research on neural receiver optimization, we demonstrated that fine-tuning a 5G neural receiver on site-specific channel data (whether measured directly or generated from a trained flow model) achieves a 10% block error rate at an SNR 1.85 dB lower than receivers trained only on TDL models. This translates directly to improved coverage and throughput in real deployments.
The neural receiver training pipeline shows pre-training on TDL channels followed by site-specific fine-tuning on measured or AI-generated channel data.
Unlimited Use Cases Beyond compliance testing, AI-generated channels enable: – ORAN RIC/xApp training and conformance testing with realistic, reproducible fading – Robustness evaluation of PHY algorithms including precoding, coding, HARQ, and channel estimation – RF sensing, localization, and joint communication-and-sensing ﴾JCAS﴿ data augmentation – Large-scale synthetic datasets for reproducible ML-for-wireless research – Differentiable end-to-end system optimization via backpropagation through the channel generator – Safety-critical simulation for UAV, V2X, and emergency networks in complex NLoS environments – Environmental change detection by comparing live channel conditions against model-generated baselines.
Results from O-RAN PlugFest Testing
Phase 1 hardware validation compared three scenarios: a baseline direct RF path with no emulation, a standard 3GPP TDL-B model with 400 Hz Doppler and a 100-nanosecond delay spread, and AI-generated channels trained on measurements collected in Ballston.
The results showed that AI-generated channels produced distinct, more dynamic propagation conditions than those of standardized models. The 5G system adapted its modulation, coding, and scheduling behavior in response, demonstrating that the end-to-end workflow operated as intended. The testing also identified initial differences in the signal-to-noise ratio calibration, which will be addressed in Phase 2.
These findings validate the central concept that compact AI models trained on real measurements can drive professional channel emulation hardware in an O-RAN testing environment.
What Comes Next
DeepSig is exploring a joint path with Keysight to make this capability available to customers. The long-term vision is a library of downloadable, site-specific channel-model packages trained on measurements from real-world environments. Test labs could evaluate systems against models representing dense urban areas, manufacturing floors, stadiums, or other challenging deployment scenarios.
As more locations and environments are measured, the library will continue to grow, expanding the industry’s ability to test wireless systems under realistic conditions.
Conclusion
The integration of direct RF measurements, generative AI channel modeling, and hardware-in-the-loop testing demonstrated at O-RAN PlugFest Fall 2025 marks a shift in wireless system evaluation. For the first time, test environments can reproduce the statistical characteristics of specific real-world locations using compact models that can be shared and deployed globally.
This capability does more than narrow the gap between compliance testing and field performance. It establishes site-specific learning as a foundational building block for the future of AI-native wireless systems. As the industry looks toward 6G, AI-native receivers and AI-native air interfaces will depend on models that learn from and adapt to the environments in which networks truly operate, not on idealized abstractions. Site-specific, AI-generated channel models enable that evolution by allowing systems to be trained, tested, and continually refined using realistic deployment conditions.
This approach provides a scalable and practical path for the industry to move from standardized channel assumptions to environment-aware wireless intelligence. It strengthens today’s 5G systems while laying critical groundwork for AI-driven architectures in 6G.
This work was supported in part by the National Telecommunications and Information Administration.
