A Data-Efficient Software Toolkit for the Accelerated Characterization of Store Trajectories via Operator Learning

The innovation proposed by Combustion Research and Flow Technology, Inc. (CRAFT Tech) and the University of Pennsylvania (Penn) is the development of a software toolkit supporting the analysis of Computational Fluid Dynamics (CFD) datasets of store separation scenarios from a weapons bay. This software toolkit will enable the classification of such scenarios into store releases that are acceptable or to be avoided as a function of the flowfield dynamics and store states prior to release. A distinctive attribute of the proposed innovation is that the classification capability of the most advanced approach utilized relies on an underlying predictive capability that is informed by a relatively small number of CFD datasets.  This data efficiency can be leveraged to provide an estimation of the prediction of uncertainty bounds. The proposed software toolkit entails (i) fluid-dynamics-oriented pre-processing of CFD data guided by expert physical insight such as Proper Orthogonal Decomposition (POD) and (ii) advanced problem-agnostic artificial intelligence (AI) and machine learning (ML) methodologies to perform data mining of CFD datasets. Within the context of AI/ML, dimensionality reduction approaches, e.g., Self-Organizing Maps (SOMs), and most importantly state-of-the-art operator learning approaches such as Learning Operator with Coupled Attention (LOCA) will be employed to enable fast and accurate classification and/or prediction of post-release store trajectories. A key challenge for the AI/ML approaches is the robust treatment of mathematically continuous functions of time and space (infinite dimensionality) in terms of discretized CFD solution datasets with user-specified temporal and spatial grid resolutions (high yet finite dimensionality). Unlike the vast majority of existing AI/ML techniques that focus on learning maps between high-dimensional vector spaces, the proposed operator learning methods are suitable for learning maps between infinite dimensional function spaces, allowing an analyst to accommodate functional data that have a continuous dependence on space-time coordinates. This attribute is crucial for data-efficient and resolution-independent learning from datasets that involve high-frequency or continuous measurements, such as unsteady CFD data of 6DOF trajectories. The proposed approach will be able to (i) receive continuous functions as inputs describing the flow field and pressure states in the cavity as well as the forces and moments on the store before its release and then (ii) predict continuous trajectories describing the post-release store position, velocity, forces and miss distance. In turn, this will enable the expedient classification of good or bad releases across a range of mission-relevant parameters.  A critical element of the proposed software toolkit is the capability to augment existing data, synthesizing additional cases using Artificial Neural Networks (ANNs), to ensure adequate training of the ML models.