Machine Learning

With technology nodes continuing to shrink, wafer structure – with its increasing levels of ambiguities – has become more complicated.

Nova is continuously developing deep machine learning capabilities enabling the company’s tools to learn without human programming. This multidisciplinary area involves computer and mathematic sciences, statistics, data mining, and information theoretic approaches. Nova’s innovative and unique approach to Machine Learning models enables the development of state-of-the-art solutions specifically tailored to the needs of the semiconductor industry.

Nova’s Machine Learning algorithms extract valuable information from input data regardless of training sample availability. In addition, they simultaneously extract accurate information, while maintaining the smallest possible computation load and training set size.

Nova FIT®

Highlights and Benefits

Automated Procedures

Training Set Availability

Short Setup Time

High Stability

Automated Procedures

Training Set Availability

Short Setup Time

High Stability

Automated Procedures

Machine leaning model is fully automated, as opposed to full model construction that requires expert optimization work

Training Set Availability

Model requires only an input dataset, which is usually present on customer sites

Short Setup Time

Nova’s algorithms utilize advanced computational methods to minimize training time

High Stability

Models are easily updated, due to automated training, after the detection of baseline changes, while results are kept stable over time

­­Machine Learning technology

Why Machine Learning?

As technology nodes continue to shrink, wafer structure – with its increasing levels of ambiguity – has become more complicated. Though physical modeling is the measurement method of choice, a model-less approach such as machine learning may provide a more rapid solution.

How it Works?

Machine learning receives input data and corresponding responses, and then learns the relationship between them. The learning process is an optimization problem that minimizes errors between the predicted output and reference response. The output is an estimator used to predict the future response of unseen data. Suggesting a correct tunable mapping of input into output is the root of the process: Nova’s vast experience with model-based approaches helps devise these mappings.

By perusing the most accurate prediction, the algorithm developer needs to determine the type of algorithm (e.g. neural network, support vector machine, PCA) and the optimized set of hyper-parameters. Deep knowledge of Nova’s tools enables the company to characterize expected tool noise characteristics at an unprecedented level of accuracy.

Since training, validation and testing are done at the customer’s site, Nova’s algorithm needs to be fully automated. The hyper-parameter selection algorithm is generic and versatile enough to produce correct predictions on data coming from different technology processes and measurement tools. The solutions are also designed to be robust to variations in training set sizes. These requirements are achieved while maintaining the accuracy demanded by customers.

You Might Also Be Interested In

Behind Every Chip there is Algorithmic Brilliance

Advanced algorithmics, machine learning, and “crazy” ideas: Nova believes in the strategic impo...

Demonstrating a DNA of InNOVAtion

There are Infinite ways to innovate and Nova is known to cultivate an atmosphere of innovation, enc...

A Conversation with Olga Krasnykov- Proud of Software Modeli...

Olga Krasnikov, Nova’s Modeling R&D Application Manager has been working for Nova for more th...

Press Releases

Global Logic Manufacturer Selects Nova’s Latest Materials Metrology Platform

Global Logic Manufacturer Selects Nova’s Latest Materials Metrology Platform

12 Oct, 2021
Read More
IBM Research and Nova Jointly Awarded the “Best Metrology Paper” at SPIE Advanced Lithography Conference

IBM Research and Nova Jointly Awarded the “Best Metrology Paper” at SPIE Advanced Lithography Conference

26 Apr, 2022
Read More

Our Publications

Filter By Years
2023
2022
2021
2020
2019
Authored by: Authored by: Daniel SchmidtiD ,a,* Manasa Medikonda,a Michael Rizzolo,a Claire Silvestre,a Julien Frougier,a Andrew Greene,a Mary Breton,a Aron Cepler,b Jacob Ofek,c Itzik Kaplan,c Roy Koret,c and Igor Turovetsc
Authored by: Daniel Doutt*a, Ping-ju Chena, Bhargava Ravooria, Tuyen K. Trana, Eitan Rothsteinb, Nir Kampelb, Lilach Tamamb, Effi Aboodyb, Avron Gerb, Harindra Vedalac
TSV stress evolution mapping using in-line Raman spectroscopy
April 27 2023
@
SPIE Advanced Lithography + Patterning
Authored by: Stefan Schoeche, Daniel Schmidt, Marjorie Cheng, Aron Cepler, Abraham Arceo de la Pena, Jennifer Oakley
Authored by: A. Moussa, J. Bogdanowicz, B. Groven, P. Morin, M. Beggiato, M. Saib, G. Santoro, Y. Abramovitz, K. Houtchens, S. Ben Nissim, N. Meir, J. Hung, A. Urbanowicz, R. Koret, I. Turovets, G. F. Lorusso, A.-L. Charley
Authored by: Benjamin Hickey, Wei Ti Li, Sarah Okada, Lawrence Rooney, Feng Zhang
Authored by: Benjamin Hickey, Wei Ti Li, Sarah Okada, Lawrence Rooney, Feng Zhang
Authored by: Julia Hoffman, Sarah Okada, Lawrence Rooney, Bruno Schueler, Ganesh Vanamu