Semicon West 2016
Acknowledgements o Stephen Tobin o Samsung Austin Semiconductor, Machine Learning o Jason Malik o Samsung Austin Semiconductor, Metrology o Dr. Dragan Djurdjanovic o University of Texas, Mechanical Engineering
Outline Introduction Undersensitivity: We need more data! Oversensitivity: Whoa, too much data! Incongruity: Wow, too many formats! IP Security: Hey, whom can I trust? Summary
Daily Fab Concerns Data sampling rate is not 100% Sometimes not even close! Sample set analyses are not real time Sometimes not even close! Primary Goal/Question: Right now, is my tool/process running such that there will be no adverse yield or throughput impact on the next wafer?
Yield-Up Complexity Short Ramp Time Typical Yield Improvement Dynamics Yield Random Systematic Errors Parametric Quarters (1990s) Months (Today) Need more actionable information in real time now
Metrology Complexity 3D Structure Planar to Vertical Transistor Architecture Gate Source Drain Planar MOSFET Vertical FinFET Need more 3D structural information now
Material Complexity IC Periodic Table Need more monitoring and control info now Today, nearly two thirds of the non-radioactive elements are used in every chip! 70s 80s- 90s
UP Defect Complexity Photomask Cleanliness Particulates Ultrapure Water needs <200pcs/liter @ >50nm Max defect size on mask ~7% of wafer CD Size spec needs to track with node shrinks 14nm: 1.0nm 10nm: 0.7nm (7Å) Mask Mask Wafer Wafer Need more and better inspection information now
Process Complexity: Double Patterning Spacer Type: Litho-Etch-Litho-Etch: Ref: Brian Wang Ref: Paul Zimmerman Previously took only one lithography step Need more inline SPC results information now
Immersion Technology 1.0 The IC Lithographer s Paradox 0.5 1.5 g-line 436 Resolution = Minimum Feature Size i-line 365 k 1 NA k 2 (NA) 2 = Depth of Focus Numerical Aperture KrF 248 Sub-Wavelength! ArF 193 F 2 157 1.35 Lord Rayleigh Sn 13 85 90 95 00 05 Need a working lithography solution now V UV DUV 1 DUV 2 VUV EUV 10 10 15
EUV: Only 1-2 years away since 2005 o Pros o Cons o Resolution o Source Power Inadequate o Cheaper than Octuple Patterning o Resist Maturity (RLS) o Mask Defectivity (No Pellicle) Are we there yet?! o Tin Drop Generation Weak o Vacuum (High Maintenance) o High Cost of Consumables o Subwavelength at 10nm Source: Google Images Need to get off my soapbox now
Data Analysis in the Past: Isolated Focus Test Wafers Input Settings -Mach Constants -FDC Parameters Process Recipes Output Values -SPC Results -Log Files -Sensor Data Individual tools and processes only
Massive Amounts of Data Ignored Only small quantities of data could be collected, analyzed, and acted upon
Equipment Complexity: TMI (too much info) But more information Sensory Overload! was ignored (>100,000 than sensors) used!
Goal: 100% of Data Analyzed and Applied Use all data to produce actionable info for: o like tools/processes o prior and following tools/processes o facilities conditions
# Sensors Too Much For Humans To Digest! High Commercial Jet Life-Death Important Not Life-Death Important Semiconductor Process Med Bars of Soap Low 10 0 10 1 10 2 10 3 10 4 10 5 10 6 10 7 10 8 log # Products Toward high end of Sensors AND Products!
Back To The Future! Year # Variable Interactions Sensors Per Wafer Analysis Method # People Needed to Review All Data 1975 1-2 200 Review All Data 1 1985 1-2 500 Test Hypotheses 3 1995 1-2 5000 Model Based Problem Solving 25 2005 1-3 50,000 Model Based 250 Problem Solving 200 1 Today 1-4 >1,000,000 Machine Learning 5000 Reduces the set of variables to the significant few
Compatibility & Security Issues o Legacy systems not designed to collaborate o Unequal security levels vs. o Some data are proprietary to each party: Vendor [chamber+frame] User [fab+enterprise] Customer o Attached directly to Mfg floor control systems o High risk to production schedule impact o Data are stored in different formats o Product throughput/output planning & analysis o Yield/Quality analysis and control o Financials We have lots of data, but much is not yet in a format that can be used as information.
Summary & Help Needed o Undersensitivity o More sensors/data on critical tools, processes, etc. o Oversensitivity o More data compute horsepower and memory o More efficient analysis techniques (move only select data) o Incongruity o More capability to combine & analyze unstructured data o IP Security o More clarification of what data are truly proprietary o More sharing of what data remains
Thank You!