⚡ electrical engineering

Dynamic Interference Management for TN-NTN Coexistence in the Upper Mid-Band

This paper proposes a Proximal Policy Optimization (PPO)-based reinforcement learning framework that dynamically manages terrestrial network power and antenna downtilt to mitigate interference for non-terrestrial networks in the FR3 upper mid-band.

Original authors: Pradyumna Kumar Bishoyi, Chia Chia Lee, Navid Keshtiarast, Marina Petrova

Published 2026-02-12

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Pradyumna Kumar Bishoyi, Chia Chia Lee, Navid Keshtiarast, Marina Petrova

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are at a massive, crowded music festival.

There are two different groups of people trying to enjoy music at the same time. The first group is the "Terrestrial Network" (TN)—think of them as a massive crowd of people on the ground, all talking loudly to their friends using megaphones. The second group is the "Non-Terrestrial Network" (NTN)—think of them as a group of people in a high-rise observation tower, trying to listen to a very delicate, quiet violin performance via a radio headset.

The problem? The "megaphones" from the crowd on the ground are so loud that they are drowning out the "violin" for the people in the tower. This is exactly what happens in the FR3 frequency band (the "upper mid-band" of radio waves). As we use more 5G/6G on the ground, we risk "deafening" the satellites and mobile devices that rely on space-based signals.

The Problem: The "Loud Neighbor" Dilemma

Currently, if we want to protect the people in the tower, we usually do one of two things:

The "Silence" Method (Exclusion Zones): We tell everyone on the ground to stop talking entirely if they are too close to the tower. This works, but it’s a huge bummer for the people on the ground who can't communicate anymore.
The "Precision" Method (Beamforming): We tell everyone to point their megaphones very carefully away from the tower. This is great, but it requires everyone to know exactly where the tower is and where every single person is standing at every second. In the real world, things move too fast for this to work perfectly.

The Solution: The "Smart Conductor" (PPO Reinforcement Learning)

The researchers in this paper decided to stop using "all-or-nothing" rules and instead created a "Smart Conductor" using Artificial Intelligence (specifically an algorithm called PPO).

Instead of just turning people off, the Smart Conductor manages the crowd using three "volume knobs":

The Volume Knob (Power Control): Instead of shouting, the people on the ground are told to whisper just enough to be heard by their friends, but not so loud that they disturb the tower.
The Angle Knob (Antenna Downtilt): Imagine if your megaphone had a tilt. Instead of pointing it straight up toward the sky (where the satellite is), you tilt it down toward the ground where your friends are. This keeps the "noise" directed away from the "violin" in the tower.
The Selective Mute (Sector Muting): If one specific person is being way too loud and nothing else works, the conductor tells just that one person to be quiet for a moment, rather than silencing the whole crowd.

How the AI Learns

The AI doesn't start out knowing the rules. It learns through trial and error, much like a child learning to play a video game.

It tries a setting (e.g., "Let's tilt the antennas down").
It checks the "Score" (the Reward): Did the people in the tower hear the violin? (NTN Throughput) and Are the people on the ground still able to talk? (TN Activity).
If the score goes up, the AI says, "Aha! Do more of that." If the score goes down, it tries something else.

The Result: A Happy Festival

The researchers tested this using real-world data from Frankfurt, Germany. The results were impressive:

Better Listening: The "noise" (interference) hitting the satellites was reduced significantly (by up to 8 decibels).
More Talking: Unlike the old "Silence Method" which shut down huge parts of the ground network, this AI method kept 87% of the ground stations active.

In short: The paper provides a way for ground-based 5G/6G and space-based satellite networks to share the same "airwaves" without fighting, using an AI conductor that balances the volume so everyone can hear their own music.

Technical Summary: Dynamic Interference Management for TN–NTN Coexistence in the Upper Mid-Band

1. Problem Statement

As 5G/6G adoption grows, the Frequency Range 3 (FR3)—the upper mid-band (7–24 GHz)—is being identified as a critical spectrum resource. However, this band is already heavily utilized by incumbent Non-Terrestrial Networks (NTN), such as Low Earth Orbit (LEO) satellites.

The core challenge is the mutual interference caused by dense Terrestrial Network (TN) deployments (e.g., 5G gNodeBs) toward highly sensitive NTN downlink receivers. Existing mitigation techniques face significant drawbacks:

Beamforming/Precoding: Requires highly accurate, real-time Channel State Information (CSI), which is difficult to maintain in dynamic satellite environments.
Exclusion Zones: Rely on static "muting" of base stations, which creates massive coverage gaps and degrades terrestrial service quality.

2. Methodology

The authors propose a dynamic, adaptive interference management framework that moves away from static rules toward a continuous optimization approach.

A. Optimization Framework:
Instead of just muting cells, the paper formulates a joint optimization problem to maximize NTN downlink throughput by controlling three specific TN parameters:

TN Downlink Transmit Power ( $P_{gNB}$ ): Reducing power to lower interference.
Antenna Downtilt ( $\theta_d$ ): Reshaping the vertical radiation pattern to suppress upward-pointing sidelobes toward satellites.
Selective Sector Muting ( $\delta$ ): A discrete action to mute only specific sectors of a base station rather than the entire site.

B. Proposed Solution: PPO-based Reinforcement Learning:
Because the relationship between power, tilt, and interference is non-convex, highly coupled, and non-differentiable, traditional optimization fails. The authors implement a Proximal Policy Optimization (PPO) algorithm, an actor-critic reinforcement learning method.

Agent: A centralized controller at the core network level.
Observation Space: Includes satellite elevation angles, the percentage of interfered NTN users, current NTN throughput, and TN base station activity.
Reward Function: A weighted sum designed to balance three competing goals: maximizing NTN throughput, maintaining high TN base station activity, and minimizing the number of NTN users experiencing harmful interference.

3. Key Contributions

Joint Control Strategy: Developed a unified framework that simultaneously manages power, antenna tilt, and sector muting.
Adaptive RL Framework: Introduced a PPO-based approach capable of learning optimal control policies in dynamic environments (e.g., as a satellite moves across the sky).
Modeling of FR3 Dynamics: Provided a comprehensive mathematical model for both TN (Urban Macro) and NTN (3GPP TR 38.811) propagation and antenna patterns specifically for the 12 GHz band.

4. Results and Discussion

The framework was tested using real-world base station data from Frankfurt, Germany, simulating a LEO satellite pass.

Interference Reduction: The proposed method achieved a 6–8 dB reduction in the median Interference-to-Noise Ratio (INR) compared to no-coordination and rule-based (ASCENT) methods.
Terrestrial Continuity: Unlike the ASCENT method, which saw TN base station activity drop sharply (from ~79% to ~28%) as NTN density increased, the proposed PPO method maintained a high TN activity level of over 87%.
Robustness: The method proved effective across different NTN user types, providing significant protection even for "direct-to-cell" mobile users who have much lower antenna gains than fixed gateway terminals.

5. Significance

This research is highly significant for the future of 6G and integrated satellite-terrestrial networks. It demonstrates that spectrum coexistence in the FR3 band is feasible without sacrificing the performance of either system. By replacing "blind" muting with "intelligent" adaptive tilting and power control, the paper provides a blueprint for how operators can maximize spectrum efficiency in increasingly crowded frequency ranges.