Improved turbulence models are necessary for achieving more accurate solutions in Reynolds-averaged Navier–Stokes (RANS) simulations. RANS is widely used in various engineering applications, and enhancing its accuracy is crucial for geometry design and control applications. With the increasing availability of high-fidelity datasets, machine learning (ML) techniques offer the opportunity for data-driven inference of RANS equation closure terms. By Pope's theoretical analysis, ML turbulence closure models for RANS simulations are often used alongside physics-informed preprocessing of equivariant tensor basis. In this work, we replace this established approach by designing an equivariant turbulence model with Euclidean Neural Networks (e3nn). We demonstrate that Pope's tensor basis is a special case of our model. We test the model on the periodic hills flow case dataset. Our approach significantly improves the prediction of anisotropic components of Reynolds stresses, leading to more accurate modeling of the flow field when integrated into RANS through a single injection.
10 June 2025 in Machine Learning: Science and Technology
In this work, we utilize computationally efficient molecular dynamics simulations to create a machine learning pipeline for discovery of crystalline multi-component alloys. We employ high-quality interatomic potentials to create a dataset of NiFeCr structures and apply crystal diffusion variational autoencoder to maximize their mechanical properties, i.e. bulk modulus. As part of the experiment, we utilize local search coupled with classical interatomic potentials to explore the local structure space and show that utilization of this procedure greatly improves optimization capability of the neural model. We also expand the model with an extra submodule, which attains 42% improvement on modeling the crystalline phase of the structures. Ultimately, we verify the global stability of the created structures with quantum mechanical calculation methods.