Implicit neural representations for turbomachinery geometry

An implicit neural representation (INR) is a framework for encoding signals into a neural network. One of these signals includes an implicit field, such as a signed distance field (SDF),to represent geometry in a compact and resolution-agnostic manner. A benefit of this method over more traditional approaches, such as surface meshes and point clouds, is the uniform memory footprint in the form of the neural network’s parameters. This means INRs are well-suited for encoding geometry into another deep learning model in a process known as meta-learning. This representation has potential benefits for the engineering field, for example, in generative deep learning models for inverse design. This work contributes to developing the INR specifically for engineering geometries, which have unique requirements and features, including high surface accuracy and a contrast between sharp features and smooth regions. This paper discusses best practices and considerations for achieving accurate INRs for turbo machinery applications. Furthermore, a new adaptive sample scheme is proposed for training INRs, which samples training data using an octree structure.