Like all machine learning models, PhysicsAI models are
most accurate when the design being predicted is similar to the designs used for
training.
Figure 1. Visual Representation of Similarity Score
Similarity score:
1.0
Equivalent to training sample
0.0
As far from the nearest training point as any two training points are
from each other
< 0.0
Far from a training point
Proceed with caution
Figure 2. Similarity Score for Curves
When you make a prediction, PhysicsAI will quantify how
similar the input design is to the training data in the form of a similarity score.
In HyperMesh, the similarity score is displayed in the
top-right corner of the prediction window.Figure 3.
Interpret Similarity Scores
A similarity score of 1.0 indicates that the input design is the same as one of the
training points. This is the maximum possible value.
A similarity score of 0.0 indicates that the input design is as different from the
nearest training point as the two farthest training points.
A negative similarity score indicates that the input design is very different from
the training data. It’s likely that the prediction will be low-quality unless a new
model is trained with data sufficiently similar to the predicted designs.
Note: Datasets with very little geometric variation, especially when shape changes are
the same magnitude as mesh discretization variances, may result in abnormally low
negative similarity scores.
Missing Similarity Scores
There are several reasons why you might not see a similarity score:
Your training or input designs either do not have shell elements or do not
have extracted solid faces. Similarity scores are currently only supported
in these scenarios.
Your training data contains less than two samples.
You are using PhysicsAI in a client other than
HyperMesh.
There is very little variance in the training data and nearly all inference
cases would seem like extreme novelties.
There is insufficient geometric variation in the training data.
