Session: 26-02 AI-Powered Design and Optimization

Submission Number: 177031

Point Cloud RVQ-VAE for Joint Compression of 3D Gas Turbine Blade Geometry and Multi-Field Creep Simulation Data 

The multidisciplinary geometry optimization of gas turbine blades relies heavily on computer-aided engineering (CAE) simulations to evaluate diverse failure mechanisms, with creep deformation being one such discipline posing unique data challenges. Due to its nonlinear, time-dependent nature, creep simulations require storing 3D field outputs at numerous timesteps across thousands of designs, often generating terabytes per design of experiments (DoE) study. Because each design is independently remeshed to accommodate geometric variations (e.g., cooling channels, ribs), datasets lack consistent point correspondence and exhibit topological diversity, overwhelming conventional storage and hindering long-term reuse.

To address this, we propose a machine learning method for compressing and reconstructing turbine blade geometry and creep fields without re-simulation. Building upon our prior 2D variational autoencoder (VAE) approach, we introduce a fully 3D residual vector quantized variational autoencoder (RVQ-VAE) that uses point clouds. Each point contains 3D coordinates (𝑥, 𝑦, 𝑧) but also the corresponding multi-component physical field data from creep simulations. A key challenge lies in jointly reconstructing both blade geometry and associated multi-field creep simulation outputs. To address this, we employ a joint loss formulation that simultaneously optimizes geometric fidelity and physical field accuracy, enabling coherent, high-quality reconstructions across both domains.

Our method compresses 561 independently remeshed blade designs into compact latent representations, achieving > 99.99% per-sample storage reduction while preserving geometry and physics with mean normalized 𝐿₁ errors below 0.6% (nMAE < 0.006) and R² > 0.99 across all fields.

Presenting Author: Sazeed Sufian Ali Siemens Energy Global GmbH &amp; Co. KG

Presenting Author Biography: B.Sc Transport and Traffic Engineering - Technical University Berlin
M.Sc Aerospace Engineering - Technical University Berlin

Since 2019 at Siemens Energy and since 2020 part of the Tools and Data Applications group/department.
Current: PhD Student at Siemens Energy, working on data reduction methods of 3D simulation data i.a. with the help of machine learning.

Authors:

Sazeed Sufian Ali Siemens Energy Global GmbH & Co. KG
Vikas Shivprasad Yadav Leibniz Universität Hannover (LUH)
Behnam Nouri Siemens Energy Global GmbH & Co. KG
Johannes Steiner Siemens Energy Global GmbH & Co. KG
Abdulla Ghani Leibniz Universität Hannover (LUH)