Micajah Schweikert, Modern Technology Solutions, Inc.; Nathan Highsmith, Modern Technology Solutions, Inc.; Jorge O’farrill, Modern Technology Solutions, Inc.
https://doi.org/10.64861/ZNVI1416
Keywords: Electro-optical, Infrared, Glint Estimation, AI, ML,
Abstract:
Electro-optical and infrared (EO/IR) glints arise from specular reflections of sunlight from the surfaces of space objects. These transient signatures depend sensitively on object geometry, material optical and thermophysical properties, attitude state, and the relative Sun–object–observer configuration. Accurate estimation of glint timing, magnitude, and persistence enhances object identification, discrimination, and tracking, and directly supports Space Situational Awareness (SSA) in an increasingly congested orbital regime. However, generating high-fidelity reflective signatures traditionally requires computationally intensive physics-based simulations that resolve facet-level interactions, self-shadowing, multi-bounce reflections, and temperature-dependent material behavior. As geometric resolution and thermophysical coupling increase, runtime grows nonlinearly, limiting the practicality of these methods for high-throughput trade studies and operational analysis. Additionally, traditional computational methods for target characterization will struggle to scale in the increasingly congested domain of low Earth orbit (LEO). Artificial intelligence and machine learning (AI/ML) technologies offer viable solutions by providing intelligent algorithms that can process complex data efficiently, enabling enhanced Space Situational Awareness (SSA) in proliferated environments. Many AI/ML systems are capable of operating in real time and on the edge, minimizing latency and optimizing decision-making in rapidly evolving scenarios. With edge processing, AI/ML can preserve critical communication bandwidth, allowing prioritized transmission of mission-essential data. _x000D_
_x000D_
We introduce, quailfy, and detail hardware requirements using a Broad Digtial Twin (BDT) deployable on edge systems. This BDT is a physics-informed deep neural network surrogate trained on a sparse but representative set of high-fidelity simulations. Rather than replicating expensive facet-to-facet calculations during inference, the BDT learns the nonlinear mapping from geometry, material properties, and trajectory state to EO/IR reflective response. The model captures dominant physical drivers of glint behavior while reducing computational cost by 100–1000× relative to traditional solvers, enabling scalable evaluation across large design and operational trade spaces._x000D_
_x000D_
A key contribution of this work is the generalization of the model across diverse targets. Conventional surrogate models often specialize to a single object configuration, requiring retraining for each new geometry or material set. In contrast, the BDT employs continuous, learned representations of geometry and material attributes that allow it to interpolate and extrapolate across object classes. Analogous to continuous word embeddings in natural language processing—where semantically related words occupy nearby regions in latent space—our framework constructs continuous “target embeddings” that encode geometric structure and material reflectivity characteristics. These embeddings enable the network to differentiate reflective behavior across objects while recognizing shared structural or material similarities. As a result, the model does not merely memorize signature responses; it learns a structured latent representation of target physics._x000D_
_x000D_
Geometry representations encode surface topology, facet orientation distributions, and macro-scale structural groupings, while material representations capture optical parameters such as absorptivity, emissivity, and reflectivity, along with temperature-dependent behavior. By projecting these attributes into a unified latent space, the BDT disentangles geometry-driven specular effects from material-driven reflectance trends and trajectory-dependent illumination conditions._x000D_
_x000D_
This continuous embedding strategy provides two critical advantages. First, it enables efficient learning from sparse high-fidelity datasets by leveraging shared structure across targets. Second, it supports robust generalization to previously unseen object configurations whose geometry and material properties fall within the learned latent manifold. The surrogate thus functions as a mid-fidelity model that preserves physically meaningful trends while achieving orders-of-magnitude acceleration._x000D_
_x000D_
Beyond computational efficiency, the BDT facilitates anomaly characterization. Deviations between observed glints and model predictions can indicate surface degradation, configuration changes, attitude anomalies, or deployment events. Rapid prediction across hypothetical configurations allows analysts to assess candidate explanations in near real time, enhancing identification and behavioral assessment workflows._x000D_
_x000D_
By integrating physics-informed learning with continuous target representations, this work bridges the gap between computational tractability and physical fidelity in EO/IR signature modeling. The Broad Digital Twin provides a scalable, generalizable, and interpretable framework for rapid glint estimation, enabling high-dimensional trade-space exploration and strengthening SSA capabilities across the space community._x000D_
_x000D_
_x000D_
Date of Conference: September 15-18, 2026
Track: AI for SDA Applications