Designing an Automation Pipeline for High Fidelity Design Space Exploration

The aerospace industry faces unique bottlenecks that hinder its growth. Stringent industry standards often constrain the critical need for rapid prototyping and manufacturing. These include strict safety requirements, incredibly long certification cycles, high testing and manufacturing costs, and notably slow design cycles, all of which impede quick progress.

At USC, outside the classroom, I spent most of last year researching at a lab deploying machine learning on robots to execute real-world tasks (folding, sorting, catching, etc). The contrast between robotics, which quickly adapts to technological advances, and the sluggish aerospace industry prompted a question: How can aerospace accelerate its development by leveraging optimization, machine learning, and automation?

Then I stumbled across nTop. A platform with a plugin system and unbreakable parametric geometry is novel and answers that exact question. nTop is adaptive, quick, and intuitive.

Within my first two weeks using nTop I had built a parametric 3D wing generator with automatic ribs and spars (see Figure 1). I spent most of my first month focusing on field-driven geometry and understanding core implicit modeling concepts. Everything clicked when I grasped the idea of field remapping, allowing transformations to mold shapes into intricate geometry. As the summer progressed, my nTop skills improved along with the continuously developing software. Modelling in nTop became natural very quickly.

Soon, I was exploring UAV concepts. Everything on the drone fuselage in Figure 2 is fully parametric, from the splines controlling the loft to dynamic duct sizing. Instant updates drive quick iteration, creating the ability to sweep design spaces that were previously left unexplored.

nTop is straightforward to learn and offers incredible depth for advanced use. As the software matures, implicit geometry will become the industry standard – I expect to use it throughout my career. The ability to design, modify, and test at this pace is precisely what the aerospace industry needs to push forward.

Conceptual design projects rely on fast and consistent iteration to provide a validated project framework. However, engineers have found themselves wasting valuable time fixing and altering project models between these iterations. In the traditional CAD-to-CFD workflow, particularly under the heavy time constraints often imposed within aerospace companies, this acts as a bottleneck towards providing the succinct validation necessary to win contracts. nTop's parametric modeling and automation capabilities help engineers quickly reach an optimal solution.

From before I arrived at nTop, I knew I wanted to build an automation pipeline that would allow for powerful, high-fidelity design exploration, incorporating CAE results. Engineers can automate such conceptual design workflows by leveraging seamless integration with Luminary Cloud and nTop’s powerful parametric modeling capabilities. This specific example explores an L/D vs. CL, Max trade, enabling systematic evaluation of hundreds of unique wing configurations without manual intervention.

Design Space Exploration with nTop

The foundation of this framework begins with statistical sampling. nTop’s Run Command block offers unparalleled flexibility in methodology, allowing integration with external platforms through the command line. Although Latin Hypercube Sampling (LHS) was used in this workflow, other advanced sampling and optimization algorithms are effortlessly interchangeable.

A design of experiments (DOE) Python script generates design points across five critical wing parameters utilizing LHS (tested with 200 samples). This methodology ensures a uniform distribution across the design space.

The parameter bounds define practical design ranges, creating a 5D space for sampling:

→ NACA 4-digit airfoils
→ Sweep angles from 0° to 30°
→ Taper ratios from 0.1 to 1.0
→ Winglet bend angles from 0° to 25°
→ Root chord to wingspan ratios from 0.1 to 0.75

The nTop workflow encapsulates the complete wing geometry logic, encompassing airfoil section generation, planform shaping, winglet integration, meshing operations, and unique file names for export. Each execution produces a unique CGNS mesh file with wall and farfield boundary conditions, ready for Computational Fluid Dynamics (CFD) analysis. The Run Command block initiates the CFD by executing a wrapper script, which activates the appropriate conda environment and handles subprocess execution.

Parametric Wing Geometry Generation

In the nTop GUI, input parameters are defined at the top of a file. Five parameters from within the bounds are currently selected, which govern the geometry generation.

Although it is unrealistic to efficiently sample this design space by hand, the automation process is incredibly straightforward. nTop Automate offers an intuitive Command Line Interface (CLI), allowing users to create a JSON file template for the desired .ntop file, enabling the execution of nTop with custom inputs without ever opening the app.

The inputTemplate.json structure drives the parametric wing generation. The DOE script writes custom input files for each design from the sampling, subsequently launching nTop in headless mode through subprocess calls.

Integration with Luminary Cloud CFD Analysis

Integrating with Luminary Cloud transforms each generated geometry into comprehensive aerodynamic performance data. Aside from an intuitive web-based GUI, Luminary Cloud provides an incredibly robust Python SDK, enabling complete control from simulation setup and meshing to advanced post-processing. Using the SDK, a Python script orchestrates the entire CFD workflow, encompassing:

→ Project Creation: Each design is assigned a unique Luminary project, utilizing timestamped naming conventions for systematic organization.
→ Boundary Condition Setup: Automated detection of mesh boundaries facilitates the assignment of farfield and wall conditions. The script identifies specific boundary names and configures flow physics accordingly.
→ Materials and Physics: Air properties are modeled using ideal gas assumptions with Sutherland's viscosity model. Farfield conditions are set for Mach 0.7 flight at standard atmospheric conditions.
→ Angle of Attack Sweeps: Each geometry undergoes analysis across multiple angles of attack to identify optimal performance characteristics launched in parallel. Convergence monitoring, tracked through lift coefficient changes, ensures solution quality.
→ Outputs: Surface forces are integrated to calculate lift and drag values, alongside their respective coefficients. Luminary offers an advanced output suite far beyond L/D and aerodynamic coefficients.
→ SDK Integration with nTop: After simulation, the desired outputs ( L/D and CL, Max ) get captured through stdout redirection in the subprocess wrapper, allowing nTop's Run Command to receive clean numerical results. The results can be easily adapted into the workflow where they may be needed for downstream analysis. In the context of this project, the results are returned to nTop, parsed, and placed into the output section. This integration simplifies post-processing by generating an outputTemplate.json file parsed into a CSV file. The file captures five unique input parameters, alongside the L/D and CL, Max outputs for plotting.

Post-Processing and Visualization

Not only can this automation generate data, but the Luminary SDK also enables sophisticated visualizations, including wake analysis. This script automatically identifies wing geometry bounds and constrains vortex visualization to relevant regions, producing clear wake structure images devoid of farfield clutter and high-velocity flow around the wing. Velocity field visualization is employed, as it offers superior visual clarity in the context of this model. Automated camera positioning and image capture ensure consistent documentation across all analyzed designs. This creates beautiful frames documenting each wing’s unique flow analysis. Again, all geometry creation and visualizations were completed autonomously.

After the sweep is complete, the 200 points are easily processed from a CSV file. Each point represents a uniquely generated wing configuration and maintains references to the original geometry alongside the respective input parameters and simulation results. This allows the designer to see both the overall results and respective geometry, which they can pull into nTop by inputting the “Config” name (e.g., “CONFIG_7”).

Note: Although this trade omits the nuance and detail of an actual aerodynamic study, nTop and Luminary are more than capable of industry-grade analysis. The workflow provides a framework for a novel and automated conceptual design workflow.

Beyond Wing Design: A Framework for Aerospace MDO

This modular approach extends its applicability far beyond singular wing optimization. This project was chosen for its relatively simple geometry, eliminating mesh bottlenecks and allowing full-scale exploration without industry hardware access. The nTop-Luminary pipeline could be easily adapted for other applications outside of external flow. Some simple examples could include:

→ Propeller Design: Optimization of blade twist, chord distribution, and airfoil selection.
→ Heat Exchanger Development: Fin optimization, flow passages, and thermal analysis.
→ Parametric Duct Optimization: Flow path geometry, cross-sectional area transitions, and pressure drop minimization.

This exact framework also sets up a streamlined path for MDO. Although Luminary Cloud CFD was utilized in this model, users could incorporate as many external applications as desired. If the software can run from the command line, incorporation into an nTop workflow is easy.

This framework manages design complexity through its modular components. Geometry generation, meshing, simulation, and post-processing operations function independently. This modularity allows for the interchangeability of different analysis systems and optimization algorithms, among other benefits.

Zoom image

Physics AI: An Obvious Extension

In under 6 hours, a 200-design sweep produced 3400 individual CFD runs. Each simulation processed multi-million cell meshes across the cloud infrastructure. This large volume of data can serve as a direct basis for developing Physics AI surrogate models.

Rapid iteration cycles quickly generate datasets, accelerating early-stage design workflows by replacing expensive CFD calls with fast surrogate predictions. This framework shows how automated pipelines can transform weeks of manual analysis into hours of systematic exploration.

The 3400 simulation results could provide valuable training data. Surrogate models enable thousand-point design sweeps in minutes instead of days, making interactive parameter adjustment possible with instant CFD predictions.

Each additional CFD run expands the training dataset, creating self-improving design systems where surrogate accuracy increases with every new simulation. The combination of rapid data generation and machine learning scales conceptual design from single components to complete vehicle optimization. nTop, Luminary Cloud, and NVIDIA are already pioneering what's possible with Physics AI.