ROAM-ML: A reduced order aerodynamic module augmented with neural network digital surrogates

Computational engineering and analysis is evolving with the inclusion of accurate and efficient digital surrogate models. Recently, neural networks have been proposed as machine learning tools to complement and solve physics-based processes using data-driven or physics-informed or a combination of these. The U.S Army Technology Development Directorate (TDD) computational aeromechanics group has built a research and operational infrastructure over the past ten years to model full-scale helicopter maneuvers with Helios simulation software, which consequently has allowed subject matter experts to generate significant amounts of meaningful data and rotorcraft engineering analysis. In this work, we integrate machine learning training and inference modules to the Helios software architecture in order to mine the high-fidelity computational domain and develop optimal digital surrogates. We expose our workflow and ML-pipeline with the purpose of revealing reproducible guidelines, valuable insights, limitations and early successes. Moreover, the operational deployment and validation of these surrogates in Helios reduced order aerodynamic module ROAM has allowed us to better understand how to evolve and mature these surrogates. ROAM-Surf-ML and ROAM-Line-ML are the high and mid-fidelity neural network surrogates presented in this work, suggesting a model fidelity spectrum where multiple formulations are feasible. ROAM-Surf-ML achieves a generalization accuracy that significantly outperforms currently in-place reduced order methods without introducing additional computational costs. Our implementation efficiently integrates network inference with actuator line methods and GPU-accelerated or CPU-based Cartesian solvers.