Optimizing Engineering Design with Conditional GANs

rolls-royce · Automotive · 2024 · source Agentic: no · promotional bias: med · source-backed · extracted with prompt 0.5.6 (schema v2)

Generated from extraction.json — do not edit by hand. Every claim cites verbatim source evidence; stated means the quote directly asserts the claim, implied means it supports it indirectly.

Problem

Domain: Engineering design-space exploration

Rolls-Royce needed to reuse legacy simulation data to explore innovative engineering designs without repeatedly running traditional geometry modeling and simulation processes.

Constraints:

• Engineering design must handle multi-objective constraints.
• Aerospace workflows require data governance and compliance.
• Traditional parametric models limit design-space exploration.
• Training workflows needed faster time-to-model and improved accuracy.

Architecture

Rolls-Royce and Databricks optimized conditional GAN training with Mosaic AI tools, AutoML, Managed MLflow, Ray hyperparameter studies, and Unity Catalog governance.

Diagram confidence: reconstructed (reconstructed from prose).

flowchart LR
  legacy-simulation-data["Legacy Simulation Data"]
  unity-catalog["Unity Catalog"]
  cgan-training["cGAN Training"]
  design-exploration-output["Design Exploration Output"]
  automl-and-mlflow["AutoML and MLflow"]
  ray-hyperparameter-studies["Ray Hyperparameter Studies"]
  legacy-simulation-data --> unity-catalog
  unity-catalog --> cgan-training
  cgan-training --> automl-and-mlflow
  cgan-training --> ray-hyperparameter-studies

Components

Data flow

Legacy simulation data is governed in Unity Catalog, cGAN training explores design concepts under multi-objective constraints, AutoML and MLflow speed iteration, and Ray hyperparameter studies improve model accuracy.

Why this architecture

Conditional GANs let the team explore multi-objective design spaces from legacy simulations while Databricks tooling handled lifecycle, tuning, and governance.

Models used: Conditional GANs · Databricks Mosaic AI · AutoML · Managed MLflow · Ray

Alternatives considered (roads not taken)

Traditional geometry modeling and simulation for each concept

Rejected because: The project aimed to assess concepts without requiring that process each time.

Parametric models alone

Rejected because: The goal was to overcome parametric-model limitations in design-space exploration.

Patterns

eval-driven-development (model-optimization)

Ray hyperparameter studies and managed experiment tooling improved model accuracy and iteration.

safety-policy-enforcement (governance)

Unity Catalog supports robust governance for aerospace compliance.

Tradeoffs

Reuse legacy simulation data

Chose cGAN-based design exploration over New geometry modeling and simulation for every concept.

Rationale: Reusing simulation data speeds assessment of innovative concepts.

Gotchas (what bit them)

Engineering design optimization must respect multiple objectives at once.

Root cause: Designs involve tradeoffs such as weight reduction versus efficiency.

Resolution: Use conditional GAN training over multi-objective constraints.

Aerospace model development needs governed data lineage.

Root cause: Simulation data and model outputs operate in a compliance-sensitive industry.

Resolution: Use Unity Catalog for governance.

Evaluation

The case reports faster time-to-model, improved accuracy through hyperparameter studies, and governance through Unity Catalog.

Metrics: Time-to-model · Model accuracy · Governance/compliance

Outcomes

• AutoML and Managed MLflow improved time-to-model.
• Ray hyperparameter studies improved model accuracy.
• Unity Catalog supported aerospace data governance.

No concrete volume or latency metrics are reported.