AI Agent Optimizes LEO Satellite Communication: Adaptive Modulation Technology Achieves Over 25% Performance Improvement

Project Core Information

• Original Author: houcinemessad
• Source: GitHub Project Modulation_AIAgent_for_LEO_-LRFHSS-LoRa-
• Publication Date: June 6, 2026

Core Achievements

A generative AI agent built using LangChain and RAG technologies dynamically switches between LR-FHSS and LoRa modulation schemes by real-time analysis of satellite geometric parameters (elevation angle, Doppler shift), achieving the following in LEO satellite communication scenarios:

1. Packet error rate reduced by over 25%
2. Decision latency controlled within 500 milliseconds

This project provides a reference AI solution for dynamic channel optimization in LEO satellite communication.

Low Earth Orbit (LEO) satellites have the advantages of low transmission latency and high bandwidth potential, but their fast-moving nature (about 27,000 km/h) brings two major challenges:

1. Severe Doppler shift fluctuations: Relative motion causes carrier synchronization difficulties
2. Continuous change in signal incident angle: Significant signal attenuation/fading at low elevation angles

Traditional fixed modulation schemes are difficult to adapt to dynamic environments, while LoRa (long-range, low-power consumption, SF7-12) and LR-FHSS (frequency hopping enhanced anti-interference, DR8/9 high-speed mode) each have applicable scenarios, requiring dynamic switching to balance reliability and efficiency.

Architecture Design

• Based on LangChain framework: Provides tool calling and memory mechanisms
• Integrates RAG technology: Builds a communication scenario knowledge base and supports similar case retrieval

Input and Output

• Input: Satellite elevation angle (reflects atmospheric penetration depth), Doppler shift amount
• Output: Binary decision (LR-FHSS DR8/9 or LoRa SF7-12)

Training and Data

• Dataset: 1000+ communication scenario CSVs covering the entire transit process (rise to fall)
• Training strategy: Supervised pre-training (historical optimal decisions) + online reinforcement learning (explore better strategies)
• Role of RAG: Retrieve similar cases in unknown scenarios to reduce labeled data dependency

Core Components

1. Real-time data stream processing: Receives telemetry data, extracts and preprocesses geometric parameters
2. AI decision engine: Encapsulates LangChain agent and RAG retrieval logic
3. REST API interface: Provides standardized access (query recommended configurations, decision explanations)
4. Real-time dashboard: Monitors satellite trajectory, channel quality, decision history

Performance Optimization

• Model quantization and pruning: Adapt to edge devices for fast operation
• Cache precomputation: Prepare candidate decisions in advance to handle predictable satellite position changes
• Overall decision latency ≤500ms (including RAG retrieval overhead)

Key Indicator Improvements

1. Packet Error Rate (PER): Reduced by over 25% compared to fixed modulation schemes, improving data transmission success rate
2. Decision latency: Completes channel input to modulation decision within 500ms
3. Energy consumption: On-demand modulation switching saves terminal energy consumption by 20-30% (significant for IoT devices)

Tests covered simulation environments and real LEO links, verifying the practicality of the solution.

Open Source Value

Provides an AI optimization example for LEO satellite communication, supporting cumulative swarm intelligence (evolving from massive terminal experiences)

Transferability

• Extend to other modulation technologies: such as 5G NR MCS selection, DVB-S2X ACM
• Adapt to dynamic scenarios: High Altitude Platform Communication (HAPS), UAV communication

Industry Benefits

• Operators: Improve spectrum efficiency and user capacity
• Terminal manufacturers: Easily integrate AI capabilities via REST API

LEO satellite communication is moving towards large-scale commercialization, and dynamic channel optimization is a core challenge. This project integrates LangChain, RAG, and reinforcement learning to achieve a balance between performance and interpretability.

The 25%+ PER improvement and sub-second latency prove the practical value of AI technology. In the future, with dataset expansion and model evolution, it is expected to play a greater role in satellite internet construction and help realize the vision of global seamless connectivity.