Advanced AI-ML approaches for CPS design and functional correctness of LEC in CPS operation

Abstract

In CPS design process, the presence of computationally complex slow evaluation processes such as Computational Fluid Dynamics (CFD) acts as bottlenecks in the design iteration. To overcome these bottlenecks and accelerate the design loop, I pursued two research directions: first, developing sample-efficient optimization methods and second replacing these simulation behaviors using data-driven surrogate models. For sample efficiency, by leveraging modern AI-ML-based approaches such as Bayesian optimization, Gaussian Process-based mixture model, etc to these domains, I demonstrate high sample efficiency and convergence properties. Second, through multiple experiments, I demonstrate that AI models can effectively capture and generalize complex physical behavior involved in solid-fluid dynamics and finite element-based numerical simulations. By utilizing these trained surrogates to replace physics-based simulations, design optimization loops achieve a remarkable speedup of up to ten thousand-fold. However, training these data-driven surrogates is challenging due to the requirement of generating a sheer amount of training data. To address this, I propose a noble approach of deep active learning that selects strategic samples during surrogate training. In system operation, Learning Enabled Components (LEC) deployment in system operation leads to functionally incorrect behaviors and safety violations, necessitating simulation-based verification due to the in-feasibility of formal verification in all environmental conditions. I developed simulation-based approaches to rapidly discover in-distribution rare event failures in trained LECs and also to perform point OOD detection. Empirical evaluation of the failure detection approach shows a thousand-fold speedup in the discovery of these rare failure cases than a standard Monte Carlo method and the OOD detection approach is robust to small changes in distributions. Finally, an open-source SciML tool is offered, combining CAD, CFD, Python, and AI optimization methods, enabling shape optimization in nautical and aeronautical design problems by studying solid-fluid dynamics under subsonic flow conditions.