Inconsistent confidence score while using shell and solid mesh for prediction using PhysicsAI

I am using a PhysicsAI model which is trained on linear static FE analysis for making predictions on new mesh data. The model has three components which are connected using equivalence of nodes. (For direct prediction on CAD, I am getting low confidence scores like -60 and that's why I am using mesh for prediction.) While using solid mesh (tetras) of the components I am getting good confidence scores like 0.95. But when shell mesh of the same tetramesh is used, the confidence score is coming out -ve values like -1.7. Can somebody explain why this difference is happening? Also any tips for improving predictions on direct CAD is appreciated.
Answers
-
Hi @Premchand,
The confidence score in negative infers that the input model (being used for result prediction) is not matching the models used for model training.
Since, PhysicsAI uses course mesh in the background when CAD is used as an input, the confidence score will be low & result prediction will be poor.
This limitation has been worked upon by the development team & new architecture Transformer Neural Solver (TNS) has been added in the 2025 version & is under testing. The TNS method will be less sensitive to meshing & we expect it will also improve the CAD-based predictions.
Note:
If the PhysicsAI model has been trained on results from solid mesh models, kindly use only solid mesh models as inputs for results prediction as using shell mesh models will have low confidence score in any version prior to 2024.0. Results will be slightly improved in 2024.1 (due to background architecture improvements) but not as significant as expected in the 2025 version.
Many thanks.
Kind Regards
Garima Singh
0 -
Thanks for the reply. I am using 2024.1 version. Referring to the above query, with some repositioning of the CAD inline with training model I am getting confidence scores in the range of -1 to -2. And the mesh generated by PhysicsAI is of similar size to that of my training model meshes (see attached image).
As per my understanding, even if solid meshes are given for training, shell meshes are extracted and used for training the ML model. And predictions are made only on surfaces of 3d components. Also in training, the PhysicsAI model learns the relations between geometric features and desired output (stress in my case). So if shell meshes similar to training data is given, model should ideally give a good confidence/similarity score. Please comment on this observation and correct if the above statement is wrong.
0