Moriba Jah, University of Texas Austin
https://doi.org/10.64861/MDIM9558
Keywords: Space Domain Awareness, Space Situational Awareness, AI for SDA, multi sensor fusion, multi source fusion, knowledge graphs, graph analytics, agent based AI, large language models, high performance computing, big data analytics, hypothesis based fusion
Abstract:
Space Domain Awareness (SDA) faces growing technical pressure from increasing catalog size, heterogeneous sensing modalities, higher observation cadence, and tighter decision timelines. At the same time, many emerging analytic approaches in artificial intelligence, graph analytics, and large scale computing remain at low or early technology readiness levels and are not yet systematically evaluated in SDA contexts. This paper describes the technical architecture, experimental methodology, and early integration results of SDA TAP Lab Texas, a university-embedded research laboratory designed to mature low-TRL SDA science and technology and to connect academic research with industry innovation pathways aligned to U.S. Space Force analytic needs. The emphasis is on methods, infrastructure, and evaluation design rather than operational performance claims._x000D_
The laboratory framework is built around three coupled technical elements: large scale numerical computing, knowledge graph based data integration, and agent oriented AI workflow orchestration. High performance computing resources are used to support computationally intensive SDA analytics, including large ensemble orbit propagation, batch hypothesis generation, residual field mapping, cross sensor association sweeps, and parameter sensitivity studies. These workloads are structured to explore hypothesis spaces broadly instead of collapsing early to single track or single model solutions, enabling controlled experimentation with multi hypothesis reasoning and falsification driven filtering._x000D_
Results of numerical workloads are encoded into a typed knowledge graph that functions as the central integration layer. The graph schema represents space objects, observation sets, sensors, dynamical models, state hypotheses, derived features, analytic transforms, and evaluation metrics as distinct node and relation types with explicit provenance attributes. Each analytic step is recorded as a transformation relation linking inputs, methods, and outputs. This structure supports reproducibility, audit, and comparison of alternative analytic pathways. Multi source data fusion is expressed as graph linkage across observation, hypothesis, and model subgraphs, enabling constraint checking and compatibility scoring across heterogeneous sensor sources without requiring a single fused state estimate at each step._x000D_
Agent oriented AI components are used to orchestrate analytic workflows across compute and graph services. These agents operate under bounded tool use policies and are restricted to composing workflows from verified analytic primitives. Large language models are used in a constrained role to translate analyst questions and research tasks into structured execution plans composed of graph queries, simulation jobs, estimation runs, and validation checks. All generated plans are compiled into executable task graphs and passed through verification gates before execution. This separation ensures that generative models assist with workflow construction while numerical and statistical modules produce the actual analytic outputs._x000D_
Multi sensor and multi source fusion methods in this framework are implemented through graph based hypothesis management and compatibility evaluation. Independent sensor observations are associated with candidate hypotheses through residual compatibility measures, temporal coherence tests, and dynamical feasibility constraints. Rather than enforcing immediate track fusion, the system maintains parallel hypothesis branches with evolving support measures. New evidence updates hypothesis support and may trigger falsification, refinement, or branching. This design supports experimental comparison of probabilistic and non probabilistic fusion metrics, including compatibility scores, constraint satisfaction measures, and evidence survival counts._x000D_
The SDA TAP Lab Texas environment is explicitly structured to support low and early TRL experimentation. Academic research teams contribute new models, fusion metrics, uncertainty representations, and AI control strategies as modular components. Industry participants engage through cohort based technical challenges and prototype integrations, allowing emerging methods to be evaluated against realistic SDA problem structures and data conditions. The laboratory infrastructure provides shared datasets, compute access, graph schemas, and evaluation metrics so that alternative approaches can be compared under common experimental conditions. This structure is intended to reduce the gap between academic novelty and operational relevance while preserving independent technical assessment._x000D_
Because the laboratory stand up begins in March and the reported work covers the initial integration phase through early experimentation cycles, results presented here are preliminary and focus on system behavior and measurement methodology. Initial experiments evaluate three technical properties: computational scaling, fusion traceability, and hypothesis management behavior. Computational scaling is measured by wall clock time, resource utilization, and throughput for ensemble propagation and hypothesis expansion workloads as a function of hypothesis count and observation rate. Fusion traceability is evaluated by reconstructing analytic outputs from graph provenance records and verifying reproducibility of derived features and compatibility scores. Hypothesis management behavior is evaluated using controlled multi sensor scenarios in which conflicting and complementary observation sets are injected to test branching, survival, and falsification behavior._x000D_
Preliminary findings indicate that graph centered hypothesis representation enables side by side evaluation of competing fusion strategies and supports full analytic replay from stored provenance. Ensemble and batch analytic workloads demonstrate predictable scaling across distributed compute resources for tested problem sizes. Agent mediated workflow orchestration reduces manual pipeline assembly while preserving bounded and inspectable execution graphs suitable for technical audit. Early multi source fusion experiments show that constraint and compatibility based linkage can maintain multiple viable hypotheses across heterogeneous sensors without premature convergence._x000D_
Evaluation metrics defined in the framework include cross sensor compatibility measures, hypothesis survival depth under sequential evidence, constraint violation rates, compute cost per hypothesis, workflow latency, and analytic replay fidelity. These metrics are stored as first class graph attributes, enabling meta analysis of analytic performance across methods and experiments. This supports systematic comparison of low TRL approaches and structured maturation toward higher readiness levels._x000D_
The contribution of this work is a technically explicit architecture and experimental framework for maturing low TRL AI, graph based, and large scale computing methods in SDA. By combining high performance computation, knowledge graph integration, and bounded agent oriented AI orchestration within a university and industry collaborative setting, the SDA TAP Lab Texas approach provides a repeatable pathway for evaluating emerging SDA analytics against realistic multi source data and mission relevant problem classes. Early integration results demonstrate feasibility, traceability, and scalability properties that support continued technical maturation and comparative evaluation.
Date of Conference: September 15-18, 2026
Track: Space Domain Awareness