Design Oracle for AI-Based CPS Design

Abstract

Automated design processes, especially using Machine Learning/AI techniques, require proposed systems to be evaluated across all relevant attributes, requirements, and concerns. Traditionally, teams create models in a set of engineering tools for design evaluation data. We describe a Design Oracle, where automated designers can submit a system as a graph of components and request a full range of evaluations, composing analysis workflows containing multiple engineering tools across multiple physical domains, fidelity, and scenarios. Manually created system models are not required. The system works across cloud architectures, supporting the use of parallel computation.The system leverages technology developed under the DARPA Adaptive Vehicle Make/META program, which created a uniform, multi-domain design and component representation, and tools for composition to engineering tool models across multiple tool types, domains, and fidelity levels [1]. The Design Oracle has been used within the DARPA Symbiotic Design for Cyber-Physical Systems, across two classes of target designs, Unmanned Air Vehicles and Unmanned Underwater Vehicles.