Method of generating a training set usable for examination of a semiconductor specimen and system thereof

A system to generate a training set usable for classifying defects in a semiconductor specimen, the system comprising a processing and memory circuitry configured to:

• generate a synthetic image usable for classifying defects;
• and
• include the generated synthetic images into the training set,
• wherein generating the synthetic image comprises:
• using “real world” training samples to train an auxiliary DNN (Deep Neural Network) to obtain clusters in a latent space;
• identifying in the latent space one or more clusters corresponding to user-specified requirements of a class of a defect to be presented in the synthetic image;
• and
• using the identified one or more clusters to generate the synthetic image by applying the trained auxiliary DNN to a point selected in the latent space in accordance with the identified one or more clusters, thereby generating the synthetic image with defects corresponding to the required class.

A non-transitory computer readable medium comprising instructions that, when executed by a computer, cause the computer to perform generating a training set usable for classifying defects in a semiconductor specimen, the generating the training set comprising:

• generating a synthetic image usable for classifying defects;
• and
• including the generated synthetic images into the training set,
• wherein generating the synthetic image comprises:
• using “real world” training samples to train an auxiliary DNN (Dee Neural Network), wherein the auxiliary DNN is trained to obtain clusters in a latent space;
• identifying in the latent space one or more clusters corresponding to user-specified requirements of a class of a defect to be presented in the synthetic image;
• using the identified one or more clusters to generate the synthetic image by applying the trained auxiliary DNN to a point selected in the latent space in accordance with the identified one or more clusters, thereby generating the synthetic image with the defect corresponding to the required class.