A Framework for Embedded Model Validation and Guided Data Collection

The overall goal of the proposed project is to develop an integrated and intelligent framework for multi-fidelity data-driven modeling, statistical model validation, and adaptive model refinement. The salient aspects of the proposed solution are: (1) use of multi-fidelity modeling to infuse partial or prior knowledge of the systems (e.g., low-fidelity data or simplified physics models) into the data-driven modeling to alleviate the requirement for costly and hard-to-acquire data; (2) statistical model validation techniques including a suite of data sampling, partitioning, ensembling, and mixture-of-expert techniques to interrogate the models under construction against all available data and to validate the model by utilizing its uncertainty and error characteristics; (3) adaptive model refinement exploiting model uncertainties and errors to judiciously guide data collection by selecting the most critical regions in parameter space for the next-round data acquisition, ensuring model construction with minimal efforts/resources; and (4) modular software architecture to automate and enhance model validation, and facilitate seamless integration into the DoD’s simulation and test workflow. In Phase I, all key software modules within the tool will be developed. Phase II will focus on capability extension, data-driven model optimization, visualization and interaction, and extensive technology validation and insertion into missile defense system level testbed and environment. Approved for Public Release | 21-MDA-11013 (19 Nov 21)