Pdf Solutions, Inc. · NASDAQ:PDFS
Pdf Solutions — Technical Presentation (October 2019)
30 slides from Pdf Solutions's technical presentation, Oct 15, 2019
Slide 1: S1.3 - Machine Learning in Exensio
16th Annual PDF Solutions Users Conference S1.3 - Machine Learning in Exensio® October 15, 2019 Jeff David, VP of AI Solutions
Slide 2: Forward-looking statements
This presentation contains forward-looking statements regarding PDF Solutions’ future products and business prospects that involve risk and uncertainty. Actual results could differ materially from those discussed. You should review PDF Solutions’ SEC filings, including its annual report on Form 10-K
Slide 3: What is Machine Learning??
What is Machine Learning?? o Algorithms that iteratively learn from data to find hidden insights o Allows for automatic generation of models that can analyze massive amounts of data to deliver faster, more accurate predictions o “Learns” complex correlations between incoming data and outputs. Everyd
Slide 4: Agenda
Agenda o Why is implementing machine learning so difficult in the semiconductor industry?? o PDF is UNIQUELY positioned to deliver the promise of AI to the industry o Augmenting analytics with AI/ML to improve results and make it easier to use PDF Solutions, All Rights Reserved 4
Slide 5: Other industries have more sophisticated AI than the semiconductor industry
Other industries have more sophisticated AI than the semiconductor industry Why is that?
Slide 6: Model Pipeline: a Collection of Steps, Methods and Best-Known Practices
Active learning Model Pipeline: a Collection of Steps, Methods and Best-Known Practices that Solve the Problem Outlier detection Imputation Scaling / Normalization Deep learning Semi-Supervised Threshold selection Feature selection & Statistical tests & uncertainty engineering estimation Hyper-param
Slide 7: Model Pipeline: a Collection of Steps, Methods and Best-Known Practices
Active learning Model Pipeline: a Collection of Steps, Methods and Best-Known Practices that Solve the Problem Outlier detection Imputation Scaling / Normalization Deep learning Nature of the problem dictates what Exensio® has a large library of type of algorithm and what order of tools to leverage,
Slide 8: Some Challenges for ML in the Semiconductor Industry
Some Challenges for ML in the Semiconductor Industry o Multimodal batch trajectories due to product mix o Test program changes o Process drift and shift, tool recalibration o Changing failure modes o Small amount of training data o Lack of labels o Lack of data for emerging technologies o Lack of tr
Slide 9: Today's Struggle: Silos and Local Optimization
Today’s Struggle: Silos and Local Optimization Process / Test 1 Process / Test 2 Process / Test 3 Process / Test N Adjust Execute Adjust Execute Adjust Execute Adjust Execute Analyze Inspect Analyze Analyze Inspect Analyze Inspect Inspect Data Data Data Data 100TB unused data! Our customers want • O
Slide 10: Exensio: Global and Local Optimization
Exensio®: Global and Local Optimization Process / Test 1 Process / Test 2 Process / Test 3 Process / Test N Data Lake • Multi-step learning • Multi-step adjustment Lower cost • Uses in-situ, upstream & Improved quality historical data Higher Yield • Human light, machine learning PDF Solutions, All R
Slide 11: PDF is UNIQUELY Positioned to Deliver the Promise of AI to the Industry
PDF is UNIQUELY Positioned to Deliver the Promise of AI to the Industry
Slide 12: PDF Solutions: Spanning the Supply Chain
PDF Solutions: Spanning the Supply Chain ® TM OSAT FOUNDRY FABLESS IDM Solutions PDF Solutions, All Rights Reserved 12
Slide 13: PDF Solutions: Spanning the Supply Chain
PDF Solutions: Spanning the Supply Chain PDF has access to the RIGHT data, including new types of data PDF has access to a breadth of data across the supply chain. PDF aligns this data to make it relevant. PDF has the infrastructure built up to integrate and leverage all of this ® data to deli
Slide 14: Semantic Models: A Key Element for Applying Machine Learning
Semantic Models: A Key Element for Applying Machine Learning o The semantic data model is a method of structuring data in order to represent it in a specific logical way. It is a conceptual data model that includes semantic information that adds a basic meaning to the data and the relationships that
Slide 15: Augmenting Analytics with ML to Improve Results and Make it Easier to Use
Augmenting Analytics with ML to Improve Results and Make it Easier to Use
Slide 16: A Unified View of Semiconductor Data is Needed
A Unified View of Semiconductor Data is Needed Semi Mfg Wafer Sort Assembly Final Test System-Level Test In-Field Monitors * ** *** **** Materials Data Wafer Wire-Bond Data Reliability Data Performance Data Trending Process Data Test Field Returns Metrology Data Data Yield prediction Wafer-level gra
Slide 17: Lots of Great Modules in Exensio that Incorporate 20 Years of PDF Expertise
Lots of Great Modules in Exensio® that Incorporate 20 Years of PDF Expertise, but Requires all the Right Skills to get the Needed Results EXENSIO® Modules Rule Ensemble Engine (REE) Costs Spatial Signature Analysis (SSA) Fail Signature Detection and Analysis (FSDA) Yield Product Sensitivity Analysis
Slide 18: AIM Solutions Overview (Fabless/IDM)
AIM Solutions Overview (Fabless/IDM) Solution Description ROI Reduce engineering 1. IMD – Intelligent Material Wafer level grading & disposition with disposition time by > 50% Disposition (aka MRB) near real time execution Improves quality of disposition 200 180 Trade-Off Improves with: 160 Mo
Slide 19: AIM Solutions Overview (Foundry/IDM)
AIM Solutions Overview (Foundry/IDM) Solution Description ROI 8% Yield Improvement FDC to YMS AI modeling & prediction for 5. FDC – Fault Detection and 40% excursion reduction yield variability reduction & control plan Classification 7% Faster NPI Ramp upgrade Learning Rate 10% improvement i
Slide 20: AIM Solutions with AI Makes it Easier to Obtain Optimal Results
AIM Solutions with AI Makes it Easier to Obtain Optimal Results EXENSIO® AIM SOLUTIONS ML Pipeline Intelligent Material Disposition (IMD) ® Smart Testing (ST) Costs Early Life Failure Prediction (ELF) Fab Predictive Modeling (FPM) Fault Detection and Classification (FDC) Yield Capacity & Efficiency
Slide 21: AIM Solutions with AI Makes it Easier to Obtain Optimal Results
AIM Solutions with AI Makes it Easier to Obtain Optimal Results EXENSIO® AIM SOLUTIONS ML Pipeline Intelligent Material Disposition (IMD) Costs ® Costs Smart Testing (ST) Early Life Failure Prediction (ELF) Fab Predictive Modeling (FPM) Fault Detection and Classification (FDC) Yield Capacity & Effic
Slide 22: Augmenting Analytics with ML to Make it Easier to Use and Deploy
Augmenting Analytics with ML to Make it Easier to Use and Deploy o Our systems make it easy for our o Using ML, we automate the work customers to incorporate their performed by our customers for their knowledge to improve performance use cases Allow customer to deploy their own models ETL – Dyna
Slide 23: FDC – Results on Large Volume of Production Wafers
FDC – RESULTS on Large Volume of Production Wafers Across Multiple Process Steps Process Step True Positive True Negative Single Machine Learning Pipeline Rate Rate produced accurate model for multiple (actual = normal, (actual = abnormal, processes prediction = normal) prediction = abnormal) No
Slide 24: Collaborative Learning: Captures User Knowledge in the System
Collaborative Learning: Captures User Knowledge in the System o Final report provides the user a way to assign a new class per wafer, and retrain the model o Human interaction via Collaborative Learning allows the user to continually improve and validate wafer classifications, building user confiden
Slide 25: Smart Testing
Smart Testing *Smart Testing can be applied to any test Target Inline Test Test Decision Final Final Fab WAT ASSY Station1 Station2 Node Test 1 Test 2 Data Input Data Output Data 200 o Goal: No risky chips to field, while diverting chips from Trade-Off Improves with: Fallout (DPPM) expensive tests 1
Slide 26: Deployment Challenges
Deployment Challenges Data OSAT Data loss Training engine Foundry Security risk Inference engine OSAT OSAT Time delay Prediction PDF Solutions, All Rights Reserved 26
Slide 27: DEX Enables Edge Deployment Architecture
DEX™ Enables Edge Deployment Architecture “THE EDGE” Data OSAT Training engine Foundry Inference engine OSAT OSAT Model Edge Analytics – fast turnaround times on predictions, making ACTIONABLE predictions a reality Reduced data loss Brings advanced ML capabilities to Adaptive Testing Allows
Slide 28: Overall Benefits
Overall Benefits o Access to a breadth of data, and access to the RIGHT data o Infrastructure already built up across the supply chain o Data Science enabled in the tool – you don’t have to be a data scientist to set it up or maintain it o Our systems are designed to make it easy for our customers t
Slide 29: In Summation…
In Summation… o The semiconductor industry has been slower to adopt AI technologies than other industries o Given those challenges, PDF is UNIQUELY positioned to deliver the promise of AI to the industry o PDF’s goal is to make it easier to adopt AI – you don’t need to be a data scientist o …the AI
Slide 30: Thank You
Thank You pdf-solutions pdfsolutionsinc pdfs.inc pdf_solutions pdf_cn
PDF Solutions — Technical Presentation (October 2019) on 6ix