Machine Learning

With the continual shrinkage of technology nodes, wafer structure becomes ever more complicated with increasing levels of ambiguities. Though physical modeling is the method of choice, there are situations in which a model-less approach provides a more rapid solution. Nova is continuously developing machine learning capabilities that allow our tools to learn without human programming. This vastly multidisciplinary area involves computer & mathematics sciences, statistics, data mining, and information theoretic approaches. Both unsupervised and supervised learning algorithms are being tested in-house, and several of these have already been incorporated into our tools. These algorithms extract valuable information from input data whether training samples are given or not, and simultaneously extract accurate information, while maintaining the computation load and training set sizes as small as possible.

Highlights & Benefits

  • Automated procedure: Unlike full model construction which requires expert optimization work, Machine Learning model is fully automated.
  • Availability of training sets: Machine Learning model requires only input dataset and reference. This type of data is usually already present on customer site.
  • Short setup time: Nova’s algorithms are utilizing advanced computational methods to keep training time minimal.
  • High stability: As training is automated, Machine Learning models are easily updated following baseline changes upon detection, and results are kept stable over time.

Supervised Learning receives input data and corresponding responses and learns the relation between them. The output is an estimator which is used to predict future response of unseen data. The process of learning is an optimization problem that minimizes error between predicted output and reference response. The crux of the matter is to suggest a correct tunable mapping of input into output. Nova’s vast experience with model-based approaches helps in devising these mappings.

By perusing the most accurate prediction, the algorithm developer needs to determine the type of algorithm (for example: neural network, support vector machine, PCA, etc.) and the optimized set of hyper-parameters, while avoiding fitting the data to noise. Once again, deep knowledge of our tools, allows us to characterize expected tool noise characteristics to an unprecedented level of accuracy.

Since the training, validating, and testing are all done at the customer’s site, Nova’s algorithm needs to be fully automated. This means that the hyper-parameters selection algorithm should be generic and versatile enough to produce correct predictions on data coming from different customers, different technology processes and different measurement tools. The solutions are also designed to be robust to variations in the size of the training set. All these requirements are achieved while maintain the accuracy required by our customers.