Machine Learning-based Analytics

Victory Analytics is a machine learning-based analytics tool created especially for the technology industry’s research, development, and manufacturing. Data management and visualization, feature selection and training, modeling, and optimization components make up this system. The adaptability of the software tool enables it to cover both module level in-depth data analysis such as electrical device characterization, process flow optimization, and fab tool recipe optimization and factory level yield improvement such as Fault Detection and Classification (FDC) and Statistical Process Control (SPC).

State-of-the-art Data Analysis

Applications for Victory Analytics

Manufacturing processes: Identify bottlenecks in manufacturing processes, improve quality control, and reduce costs
Device / Circuit design and optimization: Engineers make better decisions by providing them with insights into their data. Increases the operation margin by the utilization of systematic data training
Product Yield improvement: Quality control of process specification and / or target metrics and data-driven decision making

Key Features

Multivariate data analysis using a combination of advanced statistics, neural networks, and decision trees algorithm
Capabilities for managing input data, such as detecting outliers and missing data imputation
Support multi-level classification analysis driven by the classical statistics and machine learning algorithm
Advanced Design of Experiments algorithm such as computer-generated D-optimal DoE
State of art data filtering and feature selection techniques based on the decision forest algorithm.
Various sensitivity algorithms including Sobol, decision forest, etc.
Victory Analytics can conduct time series analysis including AR/MA as well as seasonality analysis
An integrated visualization tool that offers data insight
Provide a python interface that enables complete end-user automation

Data Analysis and Modeling

Victory analytics involves collecting, cleaning, and analyzing data to identify trends, patterns, and relationships. It is an iterative process that involves a sequence of steps and requires general understanding of your input and output variables. it is recommended to set up the correct workflow:

Dataset generation: Generate dataset for modeling including identifying categorical variables, duplicated inputs, etc. Transformation and subset of data if necessary
Regression term selection: Selection of predictors and regression terms for modeling
Model build: Build regular or M.L. regression model
Model analysis: Model score, sensitivity, and data trend analysis

Data Analysis and Modeling

Neural Network Modeling Window

Neural Network Analysis

Victory Analytics Neural Network is based on the supervised learning configuration and feed forward architecture. Having a non-linear activation function in neurons does makes a N.N. non-linear, which is the biggest advantage of NN regression over conventional polynomial regression method.

Next Generation Analysis with Gaussian Process Regression

Victory Analytics Gaussian Process Regression excels in fitting non-linear datasets. GPR simplifies balancing underfitting and overfitting considerations with easy-to-use parameter tuning. With Victory Analytics expect superior performance compared to Python and R Scripts.

Gaussian Process Regression

Optimization and Desirability

Victory Analysis provides various optimization algorithms. The choice of optimization options is attributed by various conditions such as non-linearity of the model, number of inputs and outputs. The optimization algorithms are:

Local Optimization Algorithms
- Levenberg-Marquardt
- Hooke-Jeeves
Global Optimization Algorithms
- Simulated Annealing
- Parallel Tempering
- Genetic Algorithm
- Differential Evolution

In conjunction with optimization algorithm, Victory Analytics provides Desirability analysis. It is a data analysis technique that quantifies the multiple output target into a single index. It is predicated on the assumption that there is an ideal set of values for the data and that any divergence from this ideal set of values is bad.