A Trustworthy LLM-Assisted Optimization Modeling Framework for Remote Sensing Satellite Downlink Scheduling.

Area of Science:

• Artificial Intelligence
• Space Systems Engineering
• Operations Research

Background:

• Remote sensing satellite downlink scheduling is complex, requiring efficient resource allocation.
• Traditional scheduling methods may struggle with dynamic or large-scale optimization problems.
• Large Language Models (LLMs) offer potential for complex problem-solving but require careful integration.

Purpose of the Study:

• To investigate the trustworthy application of large language models (LLMs) for remote sensing satellite downlink scheduling.
• To develop a robust pipeline for integrating LLM-generated optimization models into operational workflows.
• To ensure the reliability and accuracy of LLM-driven scheduling solutions.

Main Methods:

• A three-step workflow: candidate generation, benchmark-based validation, and fallback exact solving.
• Development of a global time-slicing validator to convert visibility windows into atomic intervals.
• Physically faithful checking of constraints including ground-station/satellite mutual exclusion and per-satellite download caps.

Main Results:

• A prototype instance demonstrated the feasibility of integrating LLM-based modeling into a scheduling pipeline.
• The proposed workflow successfully validated LLM-generated schedules.
• External verification and recovery mechanisms are crucial for dependable LLM integration.

Conclusions:

• LLM-based modeling can be a valuable tool for satellite downlink scheduling when implemented with rigorous validation and fallback mechanisms.
• A guarded, multi-step approach enhances the trustworthiness of LLM-generated optimization models.
• This framework enables dependable integration of AI into critical space operations.