Research
Latest academic paper digests, data charts, and research insights from the 6G and AI research community.
16 papers
Causal Inference for Root Cause Analysis in Autonomous Networks
Dr. Fatemeh Ghassemi, Prof. Mung Chiang — Purdue University
We apply causal inference methods to network root cause analysis, moving beyond correlational ML approaches that often mislead. Our causal discovery algorithm constructs a directed acyclic graph of KPI dependencies from observational network data, enabling identification of true root causes rather than symptoms. On a dataset from a tier-1 operator with 10,000 fault incidents, causal analysis correctly identifies root causes in 88% of cases, compared to 62% for correlation-based methods.
Multi-Modal AI for Predictive Network Maintenance Using Telemetry and Images
Dr. Jie Xu, Dr. Olav Tirkkonen — Aalto University / Nokia Bell Labs
We propose a multi-modal AI system for predictive maintenance that fuses network telemetry data (KPIs, alarms, counters) with visual inspection data (drone images, thermal camera feeds) from cell sites. A cross-modal attention mechanism learns correlations between equipment visual condition and network performance degradation patterns. The system predicts equipment failures 14 days in advance with 92% precision, reducing emergency site visits by 55% in a 6-month field trial with a European operator.
Federated Split Learning for Privacy-Preserving AI in Multi-Operator Networks
Dr. Kaibin Huang, Dr. Deniz Gunduz — University of Hong Kong / Imperial College London
We propose federated split learning (FSL) as a privacy-preserving AI framework for multi-operator 6G network optimization. FSL splits the neural network model between operator premises and a neutral aggregation server, with only intermediate representations (not raw data) shared. This provides stronger privacy than standard federated learning while reducing on-device computation. Applied to multi-operator spectrum sharing, FSL achieves 95% of the performance of centralized training while provably protecting each operator's proprietary data.
AI-Driven Semantic-Aware Network Slicing for 6G Services
Dr. Petar Popovski, Dr. Elli Stai — Aalborg University / NTUA Athens
We introduce semantic-aware network slicing where AI models understand the semantic importance of data flows to dynamically allocate resources. Unlike QoS-based slicing that treats all bits equally, our approach prioritizes semantically important information within each slice. A semantic extraction module identifies critical content in video, speech, and sensor data streams, enabling 3x bandwidth reduction while maintaining task-specific quality metrics. Evaluation on a multi-service 5G testbed shows 40% improvement in perceived quality under congestion.
AI Agents for Autonomous Network Operations: Architecture and Evaluation
Dr. Laurent Ciavaglia, Dr. Razvan Beuran — Rakuten Mobile / NICT Japan
We present an AI agent architecture for autonomous network operations where LLM-based agents plan and execute multi-step network management tasks. Our system decomposes high-level operator intents into sequences of actions, executes them with safety guardrails, and learns from outcomes to improve future performance. In a production trial managing 200 base stations, the AI agent successfully resolved 78% of common network incidents autonomously, reducing mean time to repair by 65%.
Neural Network-Based LDPC Decoding for 6G Ultra-Reliable Communications
Dr. Eliya Nachmani, Prof. Yair Be'ery — Tel Aviv University
We propose a neural network-enhanced LDPC decoder that achieves near-ML decoding performance for 6G ultra-reliable low-latency communications (URLLC). Our approach uses graph neural networks on the Tanner graph structure of LDPC codes, learning optimal message passing weights that outperform standard belief propagation. The decoder achieves a 0.5 dB gain at 10^-7 block error rate with only 5 iterations (versus 50 for standard BP), enabling the ultra-low latency required for 6G URLLC.
Diffusion-Based Generative Models for Synthetic Network Traffic Generation
Dr. Guillaume Chevalier, Prof. Jiayu Zhou — Michigan State University
We develop a denoising diffusion probabilistic model (DDPM) for generating realistic synthetic network traffic data. The model captures complex temporal correlations, long-range dependencies, and multi-variate relationships in network KPIs. Synthetic data generated by our model passes 95% of statistical fidelity tests and, when used for training, improves downstream ML model performance by 18% in data-scarce scenarios. This enables operators to develop and test AI solutions without exposing proprietary network data.