Jorge O’Farrill, Modern Technology Solutions Incorporated; Nathan Highsmith, Modern Technology Solutions Incorporated; Isaac Nelson, University of Alabama; Elise Theriot, University of Alabama
Keywords: AI/ML, attention, transformers
Abstract:
Artificial intelligence and machine learning (AI/ML) are powerful tools for efficiently performing the target characterization needed for Space Domain Awareness (SDA). For SDA we must characterize objects as well as their intent. This characterization is typically done using remote sensing across multiple phenomenologies. In this paper we will focus on the generation of electro-optical and infrared (EO/IR) data. This is usually done with physics-based models that are computationally intensive and may necessitate the movement of large amounts of data to the ground across communication layers and through ground entry points. These paths are becoming more burdened as we proliferate assets in space. Unlike physics-based models, AI/ML systems that can exploit hardened GPU architectures are candidates to mitigate the conflict between proliferation and robust SDA by allowing some SDA functions to move to the edge. Although AI/ML systems offer attractive solution they can be prone to hallucinations which can lead to inaccurate results. We have attempted mitigate these hallucinations by adapting the generation of our training data using simple reinforcement learning within out training pipeline. This data generation pipeline and the model we develop within are the focus of our paper. We will detail the data generation process, specify the model architectures and explore the model’s accuracy.
We have created a physics-aided deep neural network (DNN) for generating physically accurate electro-optical and infrared (EOIR) signatures. This model predicts both thermal emissions and reflections for a given target shape and material laydown for a desired sun/earth/observer/target geometry. The model is an extension of a thermal model we trained to support our Space Force’s Overhead Persistent Infrared Modeling and Analysis Center (USSF/OMAC). Our previous models were limited to encoding only thermal signatures and producing reflected signals using ray tracing. Our new model expands on this by encoding the physics and geometry of reflections for a given target. The new model consists of two specialized DNN architectures: an attention-based deep neural network for solving the underlying thermal physics of objects in space and a standard deep neural network for modeling reflective optical signatures. The attention-based DNN efficiently learns spatial and temporal thermal dependencies. THis architecture is similar to the that used by the transformers in popular Large Language Models (LLMs). The reflection model captures complex bidirectional reflectance distribution functions (BRDFs) and target shape to reconstruct realistic EO signatures using a deep convolutional neural network. These networks are trained using a high-speed physics-based data pipeline that uses our physics-based signature generator HiRTSS (High Rate Thermal and Signature Solver), which generates synthetic training data by simulating thermal dynamics and optical reflections under varying illumination conditions.
By leveraging intelligent data generation and state-of-the art architectures, our model is robust; it maintains accuracy across a wide range of geometric configuration. It is also very fast with speeds of 100-1000x that of HiRTSS. It could be a powerful tool for enhancing SDA including detecting changes in target configurations or classifying different classes of satellites. This class of models also has the potential to be deployed on space assets (the edge) since they can run on modest GPUs.
Date of Conference: September 16-19, 2025
Track: Machine Learning for SDA Applications