Optimizing Engineering Design with Conditional GANs

Overview

This case study describes a collaboration between Rolls-Royce, the renowned aerospace and defense engineering company, and Databricks to implement conditional Generative Adversarial Networks (cGANs) for engineering design space exploration. The project represents an interesting application of generative AI in industrial manufacturing contexts, specifically targeting the optimization of preliminary engineering design processes.

It is worth noting upfront that this case study was jointly authored by representatives from both Rolls-Royce and Databricks, meaning there is an inherent promotional element to the content. The described benefits should be viewed with appropriate skepticism as they come from a vendor partnership perspective. Nevertheless, the technical approach and architecture provide useful insights into how generative models can be deployed in production engineering environments.

Problem Statement and Business Context

Rolls-Royce faced limitations with traditional parametric models in their engineering design processes. The specific challenge was to enhance design space exploration capabilities, allowing engineers to identify and assess innovative design concepts that satisfy specified design conditions without requiring the full traditional geometry modeling and simulation pipeline. This is a common bottleneck in aerospace engineering where simulation runs are computationally expensive and time-consuming.

The project worked with multiple data types including numerical, text, and image data. A particularly interesting aspect was the need to handle multi-objective constraints—design requirements that can conflict with each other. For example, the team was simultaneously trying to reduce model weight while also increasing efficiency, objectives that often trade off against one another. The goal was to produce solutions that are broadly optimized across multiple facets rather than optimal for just one dimension.

Technical Architecture and Approach

The conceptual architecture for the cGAN project followed a structured pipeline approach with several key components:

Data Modeling: The team set up data tables optimized for the specific use case, involving generating identity columns, setting table properties, and managing unique tuples. This foundation is critical for any ML system but especially so for generative models that require carefully curated training data.

ML Model Training: The developed ML models were trained using 2D representations of 3D results from typical simulation studies. This dimensionality reduction approach is pragmatic—working with 2D representations is computationally more tractable while still capturing essential geometric and performance characteristics. An innovative aspect was embedding knowledge of unsuccessful solutions into the training dataset to help the neural network avoid certain areas and find solutions faster. This is an interesting form of negative sampling that can improve training efficiency by explicitly teaching the model what to avoid.

Model Export and Transfer Learning: The architecture supports exporting trained models in standard formats, enabling deployment to secure environments where transfer learning can be conducted with project data that has restrictive Export Control or IP classification. This is a crucial operational consideration for aerospace companies dealing with sensitive defense-related work, demonstrating awareness of real-world deployment constraints in regulated industries.

MLOps and Production Considerations

The case study highlights several MLOps-related benefits from using the Databricks platform, though these should be evaluated critically as they represent vendor claims:

Experiment Tracking and Reproducibility: The integration of MLflow in Databricks is cited as ensuring transparency and reproducibility. MLflow provides experiment tracking, results sharing, and collaborative model tuning capabilities. For a project involving cGAN optimization where many hyperparameter combinations need to be tested, having robust experiment tracking is genuinely valuable for understanding which configurations work and why.

Hyperparameter Optimization: The project leveraged Ray for hyperparameter studies, enabling scalability through execution of more complex use cases that would not be viable through standard machines. Ray’s distributed computing capabilities allow parallel execution of multiple hyperparameter configurations, which is particularly useful for GAN training where finding stable training configurations can be challenging. The ability to run concurrent development with multiple individuals working on or having access to the model also speaks to collaboration benefits.

AutoML Capabilities: Databricks Mosaic AI tools are described as reducing model training and deployment complexity through features such as AutoML and Managed MLflow. While AutoML can accelerate initial model development, for specialized architectures like cGANs, the benefit may be more limited compared to standard supervised learning tasks. The claim of “faster time-to-model” is plausible but the actual speedup would depend heavily on the specific use case.

Data Governance with Unity Catalog: For a compliance-centric industry like aerospace, data governance is non-negotiable. The implementation of Unity Catalog is described as establishing a crucial governance framework, providing a unified view of all data assets and making it easier to manage and control access to sensitive data. This addresses real concerns around data lineage, access control, and audit trails that are essential in regulated industries.

Model Lifecycle and Deployment

The architecture describes a planned continual optimization cycle based on results, involving adjusting parameters, refining the dataset, and ultimately changing the approach to handling multi-objective constraints. This suggests an iterative development approach rather than a one-time model deployment, which is consistent with best practices in production ML systems.

The transition from experimentation to deployment is highlighted as a key benefit. Moving from experiments to production deployments is indeed a common challenge in ML projects, and having a unified platform that supports both phases can reduce friction. However, the case study does not provide specific details on how models are actually deployed into production design workflows, what inference infrastructure looks like, or how predictions are validated before being used in actual engineering decisions.

Future Directions and Limitations

The team acknowledges that given the three-dimensional nature of engines, future work will include exploring the transition from 2D models to 3D models. This is a significant technical challenge, as working with 3D data substantially increases computational requirements and model complexity. The current 2D approach is described as using 2D representations of 3D simulation results, which is a practical compromise but may have limitations in capturing full geometric complexity.

The handling of multi-objective constraints is described as a planned next step rather than a fully implemented feature. Developing algorithms or methods to balance conflicting objectives and arrive at optimal solutions is an active area of research, and the case study does not detail what specific approaches will be used.

Critical Assessment

While the case study presents compelling technical concepts, several aspects warrant careful consideration:

The quantitative results or performance metrics are notably absent. Claims about cost reduction, time savings, or model accuracy improvements are made in general terms without specific figures. This makes it difficult to assess the actual impact of the project.

The case study is essentially a vendor partnership announcement, which means the benefits of Databricks tools are emphasized while potential challenges or limitations are downplayed. Real-world ML deployments typically involve significant challenges around model validation, integration with existing systems, and change management that are not discussed here.

The use of cGANs for design generation is technically interesting but represents a relatively specific application of generative AI. The broader applicability of this approach to other engineering domains would depend on the availability of similar simulation datasets and the feasibility of the 2D representation approach for different problem types.

Despite these caveats, the project demonstrates a thoughtful approach to applying generative AI in industrial contexts, with appropriate attention to governance, reproducibility, and the practical constraints of operating in a regulated industry. The emphasis on reusing legacy simulation data to drive new design exploration is a pragmatic approach that leverages existing organizational knowledge rather than requiring entirely new data collection efforts.