Model-based engineering approaches are now recognized as integral to design and operation of commercial as well as military Cyber Physical Systems. The fundamental premise of model-based engineering is semantically precise representations of a system at high-levels of abstraction that can be rigorously analyzed and systematically refined into build and implementations, leading towards a correct-by-construction design.
This ideal premise, however, starts to run into practical limitations as one considers the heterogeneity, scale, and interaction complexity of real-world CPS. First off, there is not a single model in a “universal language” that can describe all aspect of a CPS. In real-world CPS industrial practices multiple languages and tools are used to model software components and design, geometric design, electronics design, and system architecture. In addition to these languages and tools for constructive models of design, there are myriads of other models that are purpose built with different levels of fidelity to enable different analysis – e.g., behavior, timing, thermal, structural etc. For example, control software design is done using a model of the system dynamics, which is a low-fidelity state-space representation that abstracts many complex physical interactions. Managing and integrating across these diverse representations constitutes an enormous challenge in effective application of model-based engineering. Moreover, CPS design is not a linear process, but a sequential and iterative decision-making process with many complex trade-offs that need to be represented as a design space of choices and systematically explored.
Our prior research on the DARPA AVM/META projects attempted addressing this challenge through construction of a model integration language, that enabled a semantically precise linkage across the different constructive models of a system developed in different languages and tools. Associated language constructs enabled modeling of design space for the system, while integrated tools enabled exploration of the design space through automated projection of integration models on to engineering domain-specific models. The project established semantic foundations of the model integration language and developed a semantic backplane for model-based engineering.
The project, more importantly, identified crucial limitations of model-based approaches, specifically: (1) multi-fidelity analysis models are essential to enable deep exploration of high-dimensional design space, however, manual construction and management of such models is cost prohibitive, (2) design space representation and exploration is needed to derive optimal designs, however enumerative and constructive representation of design space are cost prohibitive, and (3) correct-by-construction design requires formal verification in addition to engineering analyses as part of the design space exploration, however, the application of formal verification to a high-fidelity design is cost prohibitive.
Data-driven and learning-based methods provide new opportunities to address these limitations – for example, multi-fidelity surrogate models can be automatically learned; generative techniques can be used to automatically construct and expand design spaces.
In this experience talk I will motivate the need, discuss challenges, and present recent results from DARPA’s Symbiotic Design project, in synergizing data-driven and model-based methods for CPS design.
Dr. Sandeep Neema is a Professor with the Department of Computer Science, and Director of the Institute for Software Integrated Systems, Vanderbilt University. He served as a Program Manager at the Information Innovation Office of DARPA from 2016 to 2022, where he developed multiple large scale research programs at the intersection of AI/ML and Cyber Physical Systems. His research interests include cyber physical systems, artificial intelligence, model-based design methodologies, and secure distributed real-time embedded systems. Dr. Neema has more than 100 peer-reviewed articles conference, journal publications, and book chapters in the areas of cyber physical systems, model-based design methodologies, and secure distributed real-time systems.
View Slides (You must register and request HCSS Community Membership to download the slides.)