Cases & Examples
Proof Of Concept
The first version of Spaider technology was successfully validated in a proof of concept (PoC) with a major Italian aerospace company. This PoC took place within the OpenItaly framework, an open innovation contest where various corporations present technology challenges for startups to solve.
Our challenge was to develop an Artificial Intelligence-based method to accelerate numerical analyses. The provided case study involved a Computational Fluid Dynamics (CFD) analysis of an air intake scoop for a hybrid-electric aircraft.
We validated the CFD analysis results and then parametrized the geometry to generate a dataset of analyses, including different geometries and corresponding results. The proposed method utilized Convolutional Neural Networks, and the developed architecture enabled us to efficiently manage the entire design process..
Database creation
Artificial intelligence requires a large amount of data; therefore, we needed to create a dataset of similar cases for training the mathematical model.
To achieve this, we parametrized the baseline geometry. The selected parametrization provided great flexibility in exploring all possible shapes of the air intake. We then automatically generated hundreds of alternative geometries, which were subsequently simulated using a Computational Fluid Dynamics (CFD) tool.
Traininig the model
Once the simulation database was ready, it was "tensorized," meaning it was reshaped to extract information and store it in an n-dimensional grid.
Since this case study was a proof of concept (PoC), we used a moderately coarse grid, concentrating most of the cells around the scoop, where fluid dynamics gradients were highest.
Training the neural network required several hours of computation on an NVIDIA GPU. As a result, the mathematical model was able to generate a blueprint of the fluid field directly!
comparing the results
The predictions of the trained neural network model were very promising.
The CFD flow fields for pressure and velocity were accurately captured in the vicinity of the scoops, with gradients and trends well respected. Furthermore, overall errors were quite low and primarily concentrated at the edges of the tensor grid.
These differences were expected due to the imposed boundary conditions on the grid; however, they can be easily mitigated by increasing the grid's extension and resolution.
The trained model was tested on various scoop geometries using multiple Euclidean distance criteria to assess shape similarity with the database and improve predictions for extreme geometries.
Some Numbers
The proposed approach can be used both to predict the expected flow field and to accelerate numerical analyses, offering significant advantages.
In a standard CFD campaign of about 600 runs (~1100 hours of simulations on a 32core unit), we achieved:
- Flow field prediction: Real-time
- Simulation acceleration factor: 3X
- Time reduction with Spaider-AI: 68.8%
- Cost reduction with Spaider-AI: 69.1%
