A large dataset to train machine learning models for aerial vehicle design

Designing a reliable aircraft can be both challenging and time consuming, as it often involves many steps and analyses. Deep learning models could potentially help accelerate aircraft design and deployment, helping developers identify the most promising solutions or potential flaws with a specific aircraft.

To train these models, researchers will need a comprehensive dataset containing a wide range of aerial vehicle designs. However, these datasets can be difficult to compile, as many designs are protected by proprietary contracts or are difficult to source.

Researchers from SRI International, Southwest Research Institute, and Vanderbilt University recently created AircraftVverse, a massive dataset that includes thousands of aircraft designs of varying complexity. Their dataset is presented in a paper previously published arXiv It can be used to train machine learning to assist aerial vehicle designers.

In their paper, Adam D. Cobb, Anirban Roy and their colleagues wrote, “Aircraft design involves various physics domains and, therefore, involves many modalities of representation.” “Computational fluid dynamics tools for drag and lift calculations, computer aided design tools for structural and manufacturing analysis, battery models for energy estimation to scientific analytical and simulation models for evaluating these Cyber-Physical System (CPS) designs and simulation models for flight control and dynamics.”

Most existing datasets to train machine learning for computer assisted design (CAD), such as sketchgraph, deepcad And ABC The dataset mainly contains data related to individual mechanical parts. On the other hand, the dataset presented by Cobb, Roy and their colleagues includes fully developed aircraft designs that combine multiple components such as propellers, wings, motors, batteries, etc.

“The Aircraftverse contains 27,714 diverse air vehicle designs – the largest collection of engineering designs with this level of complexity,” Cobb, Roy and their colleagues report in their paper.

“Each design includes the following artifacts: a symbolic design tree describing the topology, propulsion subsystem, battery subsystem, and other design details; a Standard for Interchange of Product (STEP) model data; a stereolithography (STL) file using a 3D CAD design format; a 3D point cloud for the shape of the design; and evaluation results from high-fidelity state-of-the-art physics models that characterize performance metrics such as maximum flight distance and hover-time.”

The designs included in the AircraftVerse dataset were created using a deep learning-based approach based on general rules provided by expert aircraft designers. The researchers ran final versions of these designs through engineering models that generated metadata summarizing each of their unique features and performance.

Cobb, Roy and their colleagues write, “We also present baseline surrogate models that use different modalities of design representation to predict design performance metrics, which we provide as part of our dataset release.” We do.” “Finally, we discuss the potential impact of this dataset on the use of learning in aircraft design and CPS more generally.”

The new dataset created by this team of researchers is now publicly available onlinewith this Baseline model and underlying code, This means it could soon be used by designers and developers around the world, helping them with the design and performance evaluation of new aerial vehicles.

© 2023 Science X Network