Sales Predictive Analytics
Predicting CLV (Customer Lifetime Value)
CLV - or the value of a customer in terms of profit/revenue over the long-term is a key business intelligence metric. CLV can determine the right amount of incentives or discounts to provide an existing customer for customer retention or their initial acquisition. Machine learning can accurately predict this value based on demographic data and past purchase or engagement behavior
Semiconductor Industry Use Cases
Process tool excursion monitoring & FDC
Process tool multi-variate time series sensor data is used to perform real-time FDC (Fault Diagnosis & Classification) and/or anomaly detection during the running of a specific process recipe. In the case of anomaly detection only normal process tool data is required for model training; for FDC some labelled fault data is also required. In prediction (inference) mode, the model outputs the real-time probability the tool is in an excursion state, and classifies the fault state and its probability.
Predictive maintenance / Condition-based maintenance
Deep learning models are trained to predict (after each process recipe run) how close a process tool is to requiring a standard PM or maintenance action, and also to predict imminent failures of a non-standard/rare type. Supervised learning or anomaly detection can be used.
Metrology Proxy
Process tool time series sensor data is used to perform supervised learning predictions against labelled metrology data (such as etch rate, film thickness, film resistivity, doping levels, etc.). The model then becomes a real-time measure of the expected value of the metrology result, allowing excursions from POR to be caught before yield or parametric bin hit occurs.
Yield Management
Unit Process Monitoring Applications
For each Unit Process Tool monitored above, a "tool state vector" is created which can be added to the Yield Control data available to your Fab. By using both traditionally available data & this "tool state" data, better machine learning models can be created to use for predicting yield, parametrics, and other outcomes in your Fab. For each step in your run card a prediction can be made per wafer (or lot) giving you the most accurate insights available in the industry to your end-of-line analytics, and also enabling better decisions on scrap and the new necessity of wafer starts to meet demand. Even machine learning models trained only on traditional data (no tool state vectors) will allow a significant step toward better Fab-wide yield management.