A new study published in Engineering explores how artificial intelligence (AI), deep learning (DL) and foundation models can address the core challenges of terahertz ultra-massive multiple-input multiple-output (UM-MIMO) systems, a key enabling technology for sixth-generation (6G) wireless communications. Led by researchers from the Hong Kong University of Science and Technology, Massachusetts Institute of Technology and Tsinghua University, the work outlines three systematic research roadmaps to tailor AI algorithms for terahertz UM-MIMO, bridging the gap between AI research and terahertz communication system design.

Terahertz UM-MIMO, which deploys over 1024 antennas in the 100 GHz to 10 THz spectrum, is critical for meeting 6G’s stringent requirements—including data rates up to 1 terabit per second, ultra-wide bands of 3 THz and sub-100 microsecond latency. However, the technology faces three core hurdles: computational complexity from massive system scale and short coherence times, modeling difficulty due to hybrid far-near field effects, spatial non-stationarity and wideband beam-squint effects, and measurement limitations caused by the array-of-subarrays (AoSA) architecture and hardware impairments that hinder accurate channel state information (CSI) acquisition.

The study’s first roadmap centers on model-driven DL, which integrates domain knowledge by replacing only bottleneck modules in traditional signal processing frameworks with learnable neural components. It details four key steps: defining algorithmic frameworks (fixed point networks for iterative tasks, neural calibration for non-iterative ones), selecting near-optimal basis algorithms, designing task-oriented and empirical Bayesian loss functions, and developing generalized neural architectures like graph neural networks (GNNs) and hypernetworks. Case studies show this approach boosts near-field beam-focusing and data detection performance while cutting computational complexity by over 100 times in some scenarios.

The second roadmap introduces CSI foundation models, a unified generative AI framework that leverages the wireless channel as the common basis for all transceiver modules. Trained to estimate channel score functions from either clean channel data or raw received pilot signals, these models act as versatile priors for data augmentation, posterior sampling for inverse problems like channel estimation, sequential sampling for CSI compression and joint sampling for cross-module design. Site-specific adaptation via low-rank fine-tuning and integration with model-driven DL further enhance their practicality for diverse deployment environments.

The third roadmap explores untapped applications of pre-trained large language models (LLMs) in terahertz UM-MIMO systems, including physical-layer parameter estimation, optimization from natural language problem descriptions, offline combinatorial search for network planning, autonomous network management via LLM agents and telecommunication protocol understanding. The study notes LLMs’ cross-modal and reasoning capabilities offer unique value, while highlighting the need for efficiency optimizations to address terahertz systems’ scale and latency constraints.

The research also identifies open issues for future exploration, such as hardware-friendly model-driven DL design, scaling CSI foundation models for multi-base station deployments and resolving LLMs’ sample complexity and inference latency challenges in terahertz networks. It emphasizes that these AI-driven approaches are backward-compatible with traditional MIMO systems, making them a flexible solution for both 6G and existing wireless communication infrastructures.

The paper “AI and Deep Learning for Terahertz Ultra-Massive MIMO: From Model-Driven Approaches to Foundation Models,” is authored by Wentao Yu, Hengtao He, Shenghui Song, Jun Zhang, Linglong Dai, Lizhong Zheng, Khaled B. Letaief. Full text of the open access paper: https://doi.org/10.1016/j.eng.2025.07.032. For more information about Engineering, visit the website at https://www.sciencedirect.com/journal/engineering.