Session: 26-02 AI-Powered Design and Optimization

Submission Number: 179109

A Generative AI Framework for Gas Turbine Component Design – Industrial Perspective 

Generative AI can speed up engineering design by boosting creativity, automating routine work, combining different types of information, and improving decisions on performance, manufacturability, and cost. But today’s text- and image-focused tools struggle with complex engineering needs like spatial reasoning, physics, standards, and smooth workflow integration. This points to the need for physics-informed, domain-aware AI that works with existing tools to improve safety, quality, delivery, and cost.

In gas turbine design, requirements flow from the system down to subsystems, components, and subcomponents. Each level adjusts design variables and runs multidisciplinary analyses to meet higher-level targets. When targets are missed, feedback and requirement changes ripple upward, causing repeated, time-consuming cycles. Requirement analysis becomes a bottleneck: targets can be too strict or not feasible, and discovering infeasibility often takes many loops—delaying programs and increasing cost.

To address challenges, we propose a Generative AI framework for inverse design that seeks feasible designs meeting targets with significant reduced iterations. Invertible neural networks (INN), Variational Autoencoder (VAE), Generative Adversarial Network (GAN), Transformer Model, Diffusion Model etc. are suitable candidates for inverse mapping from targets to design variables, enabling rapid feasibility assessment and early identification of requirement modifications when necessary—thereby compressing design cycle time.

A major barrier is data availability. Training powerful Generative AI models often requires large, high-fidelity simulation datasets, which are costly and time-consuming to generate in industry. Our framework mitigates this by first training forward AI/ML models on smaller, adaptively sampled, multi-fidelity datasets to learn the physics-informed design space, quantify uncertainty, and generate synthetic data to train generative inverse models. This baseline approach—learn forward physics, calibrate with UQ, synthesize data, then train a generative model—improves data efficiency and reliability for engineering applications.

Design parameterization presents another obstacle. Conventional parameterizations can constrain the design space and impede transfer learning and multi-fidelity model building when legacy parameter schemes differ. We address this by employing geometric deep learning framework —e.g., graph neural networks, point-cloud-based modeling, signed distance function-based modeling framework—for forward and inverse modeling. They are less dependent on specific parameterizations, enabling reuse of heterogeneous legacy data and exploration of novel geometries. To further reduce data requirements while  ensuring generalizability (flow, stress, etc), physics-informed neural networks and neural operators provide an attractive options.

Industrialization requires robust integration and governance. Upstream, simulation process and data management (SPDM) must contextualize disparate data (operating conditions, meshes, boundary conditions) to ensure consistent metadata and lineage, enabling reliable model training and traceability. Downstream, continuous model management is required as AI models evolve, ensuring versioning, validation, and compliance. Finally, integrating diverse AI components demands standardized context and interoperability. We explore Model Context Protocol and agentic AI to propagate intent and data across tools, creating a unified, context-aware design workflow that aligns with industry calls for better systems integration, secure data ecosystems, and safe, effective AI deployment.

The framework will be demonstrated on gas turbine component applications, including aerodynamics, structural analysis, and aeroacoustics, showing how our framework can reduce iteration loops and accelerate delivery of optimized, manufacturable designs.

Presenting Author: Sayan Ghosh GE Aerospace Research

Presenting Author Biography: Sayan Ghosh, PhD, is a Senior AI/ML Research Lead at GE Aerospace Research. He focuses on Generative AI for design, hybrid physics–ML, and probabilistic methods for gas turbines. With 14+ years in Aerospace engineering, he has led multidisciplinary teams and programs spanning Generative AI for design, probabilistic design tool, ML for material design & manufacturing, and predictive & risk analytics for services. He holds a PhD in Aerospace Engineering from the Georgia Institute of Technology.

Authors:

Sayan Ghosh GE Aerospace Research
Sandipp K. Ravi GE Aerospace Research
Anthony Degennaro GE Aerospace Research
Anindya Bhaduri GE Aerospace Research
Tang Liang GE Aerospace Research
Jason York GE Aerospace Research
Changjie Sun GE Aerospace Research
Genghis Khan GE Aerospace Research
Robert Zacharias GE Aerospace Research