Working with the NAVPACK API
Training and deplyoing AI models for engineering applications requires connecting several steps of data processing. The general workflow generally is ordered as preprocessing, model training and hyperparameter tuning, and making predictions with the models.
The customer specific use cases, however, are unique each time – from data sources and ammount to the specific vision for the integration of AI in the daily engineering work, thus, requiring a large degree of flexibility in the process.
Our NAVPACK API is built to allow maximum flexibility to seamlessly fit any existing CAx workflow, efficiently train and leverage AI models of different architectures, and making them available where our customers want to leverage the value of real-time predictions.
From hours or days to seconds...
“From hours or days to seconds” – this is the paradigm shift AI-accelerated engineering brings to the table. Instead of being held back by the traditional iterative design process and waiting for first-principle simulation results, we empower you to focus on problem-solving in engineering. The strategy? Leftshifting or maximizing the use of existing data to uncouple the hands-on engineering work from time-consuming first-principle simulations.
-
Preprocess Data
Preprocessing of the training data is paramount for successfully training high-quality machine learning and deep learning models. Within NAVPACK, all functionality necessary for doing so is easily accessible through the Snaplib library.
Snaplib is a robust library designed for efficient data storage, manipulation, and visualization. It comprises various practical modules, each tailored to different aspects of data analysis and manipulation.
Our snapshot data backend is based on the powerful VTK library, with Snaplib facilitating the filtering, converting, and mapping of a diverse array of input data — in terms of both format and size — to the form required to leverage AI.
-
Train Models
Recognizing that not one model type can cater to all use cases effectively, NAVPACK has been designed with a flexible and versatile approach. This ensures that users are equipped with the most suitable method for each unique challenge they encounter.
NAVPACK’s ML library provides a complete toolkit for training, optimizing, and evaluating a large array of machine learning and deep learning models.
The most prminent methods included are:
– POD+I and Isomap+I for parameterized design spaces
– GNN, CNN, and PCT for learning directly from shapes
-
Make Predictions
Making predictions means leveraging the trained models to get real-time insight into new designs.
The design space can be interactively explored with supported frontend tools, such as Paraview or Blender, or by running the prediction results through the customers’ standard postprocessing workflow. Alternatively, NAVPACK enables programmatic optimization.
NAVPACK offers integrated postprocessing functionality to allow direct extraction of relevant information within the prediction workflow, eliminating the need for cumbersome file I/O. Moreover, NAVPACK can be adapted to work with customized dashboards to cater to unique use case scenarios.
Want to talk tech?
Interested in more details? Get in touch and let's figure out if we have the right solution for you.