The source code for the paper AI-Based Shape Design.
Basic python environment and the following package:
We conduct two kinds of experiments: data-driven design and model-based design.
-
data-driven design: The initial dataset are come from real world. The code and results for this part of experiment are in data-driven-design.
-
model-based design: The initial dataset are generated by sampling the space of geometric parameterization and carefully specifying a set of generated rules to satisfy some geometric conditions, such as non-self-intersection, naturalistic appearance, that is not easily described mathematically and accurately. The code and results for this part of experiment are in model-based-design.
A general steps of experiment are shown as follows.
When obtain an initial dataset, if the amount of data is not enough, we need to enrich the data. We use WGAN to generate more diverse data.
In our experiment, only the airfoil design exactly use WGAN for sampling. For wing design and rotor design, their shape are lofted by some airfoil sections, thus we use the airfoil data already enriched in airfoil design.
For model-based design, since the design rules are known, we just generate the sufficiency data directly.
Usually, we need to label all the data before training model. But labeling all these data is prohibitively expensive.
To alleviate this problem, we develop mean-teacher-based active learning algorithm. This algorithm can select the most informative samples for labeling and at the same time train the physical model.
In general design optimization problems, the design variable tend to be overly redundant, which lead to shape anomaly problem.
To alleviate this, we use auto-encoder to reconstruct the normal shapes. The auto-encoder can learn the intrinsic features of the normal shapes, when an abnormal shape is fed, the auto-encoder would not reconstruct the data well. Large reconstruction error means that the sample is more likely to be an anomaly. Therefore, we can embed the reconstruct error constrain into the design optimization problem.
The critical factor impacting the performance of auto-encoder training is the choice of the dimension of the latent space. We apply local PCA to compute the intrinsic dimension of the data.
When the physical model and auto-encoder are trained, we implement a gradient-based optimization algorithm incorporating the trained physical model and the auto-encoder, while adhering to the reconstruction constraint and desired properties constraints, to craft the optimal shape. This part of experiment requires pyoptsparse.