Efficient Domain Generalization in Wireless Networks with Scarce Multi-Modal Data

In 6G wireless networks, multi-modal ML models can be leveraged to enable situation-aware network decisions in dynamic environments. However, trained ML models often fail to generalize under domain shifts when training and test data distributions are different because they often focus on modality-specific spurious features. In practical wireless systems, domain shifts occur frequently due to dynamic channel statistics, moving obstacles, or hardware configuration. Thus, there is a need for learning frameworks that can achieve robust generalization under scarce multi-modal data in wireless networks. In this paper, a novel and data-efficient two-phase learning framework is proposed to improve generalization performance in unseen and unfamiliar wireless environments with minimal amount of multi-modal data. In the first stage, a physics-based loss function is employed to enable each BS to learn the physics underlying its wireless environment captured by multi-modal data. The data-efficiency of the physics-based loss function is analytically investigated. In the second stage, collaborative domain adaptation is proposed to leverage the wireless environment knowledge of multiple BSs to guide under-performing BSs under domain shift. Specifically, domain-similarity-aware model aggregation is proposed to utilize the knowledge of BSs that experienced similar domains. To validate the proposed framework, a new dataset generation framework is developed by integrating CARLA and MATLAB-based mmWave channel modeling to predict mmWave RSS. Simulation results show that the proposed physics-based training requires only 13% of data samples to achieve the same performance as a state-of-the-art baseline that does not use physics-based training. Moreover, the proposed collaborative domain adaptation needs only 25% of data samples and 20% of FLOPs to achieve the convergence compared to baselines.

💡 Research Summary

The paper addresses the pressing challenge of domain generalization for multimodal machine‑learning (ML) models in 6G wireless networks, where training data are scarce and the operating environment is highly dynamic. Domain shifts—both covariate shifts (changes in the input distribution) and concept shifts (changes in the input‑output mapping)—are frequent in wireless scenarios due to moving obstacles, weather conditions, and hardware variations. Conventional data‑driven training, which relies on large labeled datasets, quickly overfits to spurious modality‑specific features and fails when the test distribution differs from the training distribution.

To overcome these limitations, the authors propose a two‑phase learning framework that is both data‑efficient and robust to unseen domains.

Phase 1: Physics‑Based Training
A physics‑based loss term (L_{\text{phy}}) is added to the standard supervised loss (L_{\text{data}}), forming a total loss (L_{\text{total}} = L_{\text{data}} + \lambda L_{\text{phy}}). The physics loss penalizes predictions that violate known channel‑propagation laws (e.g., 3GPP path‑loss, free‑space loss). By constraining the hypothesis space to the subset (\Theta(\tau)={\theta\in\Theta \mid \overline{L}_{\text{phy}}(\theta)\le\tau}), the authors prove (Theorem 1) that the required number of training samples scales with (\ln|\Theta(\tau)|). Consequently, a model that respects physical laws needs far fewer samples to achieve a target error (\epsilon_1). Empirically, the physics‑augmented model reaches the same root‑mean‑square error (RMSE) as a baseline using only 13 % of the training data.

Phase 2: Collaborative Domain Adaptation
During deployment, each base station (BS) may encounter a new domain (P_{\text{test}}^k). The framework measures domain similarity between BS (k) and every other BS (j) using a distance metric (e.g., KL‑divergence) on their respective data distributions. BSs with high similarity contribute their model parameters to a weighted aggregation, producing an adapted initialization (\theta_k^{0}). The adaptation then proceeds with a few target‑domain samples (m_{\text{te}}) and a small number of epochs, while still minimizing the combined loss (including the physics term). This collaborative step reduces both the amount of required target data and the computational load. In experiments, the collaborative adaptation converges with only 25 % of the target samples and 20 % of the FLOPs required by non‑collaborative baselines, even under severe covariate and concept shifts such as rain, novel vehicle shapes, and antenna relocation.

Dataset Generation
To evaluate the approach, the authors build a novel dataset by integrating the CARLA autonomous‑driving simulator with a MATLAB‑based mmWave ray‑tracing engine. CARLA provides synchronized LiDAR, RGB, radar, and GPS streams for urban vehicular scenarios, while MATLAB computes the corresponding mmWave received signal strength (RSS). This pipeline yields a multimodal dataset that captures realistic correlations between visual/radar cues and wireless channel characteristics, filling a gap in publicly available wireless‑ML data.

Results and Contributions

1. Data Efficiency – The physics‑based loss reduces the required training data by a factor of ~8 (13 % of baseline).
2. Robust Generalization – By embedding physical constraints, the model maintains performance across both covariate and concept shifts without explicit access to target domains.
3. Collaborative Adaptation – Domain‑similarity‑aware model aggregation enables rapid adaptation with minimal target data and computational effort.
4. Dataset Release – The CARLA‑MATLAB dataset is made publicly available, providing a benchmark for future multimodal wireless research.

Overall, the paper delivers a theoretically grounded and practically validated solution for domain‑generalizable multimodal learning in wireless networks. Its two‑phase strategy—physics‑driven regularization followed by knowledge‑sharing among base stations—offers a viable path toward robust, data‑efficient 6G services such as beamforming, blockage prediction, and channel estimation in highly dynamic environments.