Research
Below are broadly the research areas that we pursue at our group:
Channel Estimation and Prediction
This research area focuses on novel methods for modeling and predicting wireless channels by representing the radio environment with learned 3D primitives. The approach enables fast and highly accurate synthesis of MIMO channel matrices, facilitating real-time channel state information (CSI) estimation, reducing pilot overhead, and improving performance in dynamic mobile networks.
Neural Gaussian Radio Fields
Modeling complex signal propagation using explicit 3D Gaussian primitives. This method allows for the rapid synthesis of MIMO channel matrices, dramatically accelerating inference and training to enable real-time applications while significantly reducing the need for pilot signals.
Continual Learning
To ensure model robustness in real-world scenarios, continual learning strategies are employed. This allows a model trained in one environment to adapt to new cells and changing conditions without catastrophic forgetting, achieving handover-robust CSI prediction for mobile users.
Non-Linear Precoder Design
This research area involves the design of novel non-linear precoding algorithms for uplink multi-carrier non-orthogonal multiple access (NOMA). The approach jointly optimizes the successive interference cancellation (SIC) decoding order, user power, and subcarrier allocation to maximize network sum-rate, moving beyond heuristic methods to find globally optimal solutions for resource allocation.
Neural minPMAC
Development of a GPU-accelerated neural network that acts as a surrogate for traditional optimization solvers. This network learns the direct mapping from channel conditions to optimal precoder parameters, replacing slow iterative calculations with real-time inference for practical deployment in 5G and 6G RAN schedulers.
CUDA-accelerated minPMAC
To complement learning-based approaches, custom, GPU-accelerated solvers are explored for scenarios requiring provable convex optimality. This research focuses on parallelizing complex optimization steps on CUDA cores to achieve massive speed-ups over traditional CPU-based solvers.
Digital Twins for Wireless Networks
The primary objective of this research is creating a powerful, simulation-ready digital twin for 6G network development. High-fidelity nGRF channel models and intelligent precoder designs are integrated into a unified, GPU-accelerated stack. This synergy enables end-to-end, closed-loop testing and optimization of AI-driven 6G concepts, allowing researchers to experiment with novel strategies under realistic, dynamically generated network conditions. This "train-in-twin, deploy-in-RAN" paradigm accelerates the entire research and development cycle for next-generation wireless systems.
