Towards high-performance strategies for the integration of traditional CFD and ML strategies

The integration of machine learning (ML) with traditional computational fluid dynamics (CFD) is opening new directions for scalable, physics-aware simulation and control on modern high-performance computing platforms. Over the last three years, Dr. Lehmkuhl group efforts at BSC has focused on developing hybrid strategies that couple data-driven models with high-order CFD solvers to address challenges of accuracy, efficiency, and robustness in turbulent flow applications. A central thread of this research is the design of hybrid reduced-order modeling approaches that combine classical POD-based representations with deep learning architectures. By embedding POD coordinates into nonlinear latent spaces and augmenting them with neural closures, we extend reduced models beyond linear dynamics while preserving physical interpretability. This effort is closely linked to a Koopman-inspired framework, where β-Variational Autoencoders are employed to construct latent embeddings that promote near-linear evolution of complex turbulent flows. Together, these strategies enable robust low-dimensional forecasting, improved long-term stability, and enhanced generalization across regimes. Beyond modeling, a major thrust has been the tight coupling of CFD with deep reinforcement learning, with particular emphasis on the SOD2D framework. Through scalable, low-latency solver–agent integration, this work demonstrates how production-grade CFD codes can become learning-enabled platforms for active flow control and turbulence manipulation. Together, these developments outline a coherent vision for high-performance CFD–ML integration, bridging rigorous numerics and data-driven intelligence for next-generation simulation and digital-twin technologies.