Neural Network Enhancements for Customized Reduced Order Models to Accelerate Nuanced COA Evaluation and Reasoning (NECROMANCER)

Planning in all its forms (e.g., early concept development and evaluation, current/future operations planning) remains a detailed, time-consuming process involving subject matter experts and rigorous analysis. While analytical tools, enabled through modeling and simulation (M&S), exist to support this process, they are still far too heavyweight—limiting the number and quality of options explored and developed. To ensure future dominance, the decision cycle for course of action (COA) development must be radically condensed. COA development involves two critical elements: COA planning and COA adjudication Currently, digital tools for COA planning and adjudication are produced as stovepipes; that is, monolithic applications that are tightly coupled and not easily separable or composable with other applications. As a result, unlocking the full potential of COA planning will continue to be restrained without advances in COA adjudication. Modern M&S tools are ideal candidates for effective COA adjudication However, high-fidelity simulations remain prohibitively expensive for COA adjudication. To address these constraints, researchers have explored simulation acceleration techniques, broadly categorized as reduced order models (ROMs). ROMs aim to preserve key behaviors of complex simulations while drastically reducing computational cost. What is needed is to accelerate COA adjudication using customized ROMs that reflect the real world at an appropriate level of fidelity and that can be executed more rapidly than current high-fidelity simulations or manual analyses. To achieve this, several challenges must be overcome. ROMs typically lose expressivity, failing to capture the full richness of original system representations. What is needed is representation that is expressive enough to fully capture the complex native data structures while enabling the learning of ROMs for system dynamics. Second, current ROM approaches bundle together entity dynamics, forcing the ROM to slow down to the “weakest link.” A method is needed to extract features describing entity dynamics while explicitly modeling the interaction structure and separability. Third, these ROMs must predict dynamics across heterogeneous, adaptive timescales to support rapid application-dependent adjudication. To address these challenges, we are pleased to introduce Neural Network Enhancements for Customized Reduced Order Models to Accelerate Nuanced COA Evaluation and Reasoning (NECROMANCER). NECROMANCER proposes a novel neural network architecture and learning paradigm to learn expressive, separable, and accelerated ROMs. NECROMANCER combines a grammar-structured autoencoder, a transformer-based entity interaction model, and an entity-structured dynamics model featuring black-box deep neural network regression and Koopman mode estimation to produce customized ROMs that can accelerate COA adjudication.