Sensor-based Quality Monitoring of Nano-manufacturing Systems

Transcription

1 Sensor-based Quality Monitoring of Nano-manufacturing Systems Satish Bukkapatnam Ranga Komanduri Oklahoma State University Stillwater, OK National Science Foundation Workshop on Research Challenges in Nanomanufacturing Systems Feb 11 12, 2007

2 Introduction Micro-electronic circuits have sub- 30nm feature size, >9 metallization layers, >10 8 nano-devices in a chip Stringent planarity of thin films needed for depth-of-focus ranges for sequential lithography process Wafer quality and yield depend on assuring uniformity and integrity of the nm thick films and structures Sensor-based monitoring of deposition, growth and planarization of nano-films and structures desirable Wong, 2002 Quality assurance is a major challenge in nano manufacturing systems

3 Recent advances in wireless MEMS/NEMS sensor technologies, and nonlinear stochastic dynamics modeling offer new opportunities localized and economical Sensing of Nano-Manufacturing Processes Monitoring of nano-film/structure deposition and growth Endpoint detection e.g., thickness and deposition rate monitoring (Buzea and Robbie 2005) Size and fault detection e.g., in-situ AFM, optical/laser scanning, pyrometer, surface photo absorption, micro-balance Monitoring of nano-film planarization and asperity removal Endpoint detection e.g., Optical, AE, electrical current, acoustic, load sensors (Bibby and Holland 1998, Saka 2004, Moore 2004, Mikhaylich 2004) Defect detection e.g., Optical, AE, electrical current, load sensors (Tang et al. 1998, Vasilopolous 2000, Surana 2000, Dennison 2004, Steckenrider 2001, Braun 1998, Chan et al. 2004, Xie et al. 2004)

4 Sensor Advancements MEMS/NEMS wireless sensor units Facilitate close proximity monitoring Obviate need for bulky dynamometers Optimized antennae for micro RF sensors Small device footprint (<1mm) Greater spatial resolution Self-powered (micro-power generators) e.g., use ambient acoustic, thermal, solar energy Facilitate durability of monitoring systems High transmission throughput protocols 1kHz-0.1MHz data transmission rates Facilitate high frequency sampling and multisensor data fusion OSU s Microchip (PIC) sensor tag

5 Modeling & Feature Extraction Advances Consideration of nonlinear dynamics important to track nano-scale variations underlying measured sensor signals Dynamics is nonlinear stochastic and signals are highly nonstationary (bursty) Signals from wireless sensors contain higher noise levels due to measurement errors and transmission related distortions Sine Wave Recurrence Plot Recent approaches motivated in weather prediction (e.g., customized wavelets, recurrence analysis) can be helpful Time Index Time Index

6 Research Methodology Modeling: Statistical analysis Dynamic characterization Wireless sensing of MRR and Ra: Oxide CMP Cu-CMP Cu-ECP

7 Experimental Approach

8 Sensor Signal Characteristics Aperiodic, Nonstationary (bursty) with continuous frequency spectrum

9 Auto Mutual Information Nonlinear Stochastic Dynamics Time Delay (Data Points) Fraction False Nearest Neighbors Embedded Dimension Correlation Sum C (d E, ε) d E = 1 d E = 9 ln (Divergence Rate) Log. Neighborhood Window (log ε) Iteration Number Experimental characterization shows that nonlinear stochastic models are necessary to capture (high dimensional chaotic) process dynamics

10 MRR Estimation in Oxide CMP MRR ζ + ψ Θ wfmean Observed Nons 11 terms + Sets + 11 terms 15 terms 18 terms Predicted 6 x ζ Ψ Θ R 2 = 85.6% 15 Terms e Run Order Observed x 10-4 Machine settings (e.g., down force, flow rate) with principal sensor features enhance MRR prediction accuracy to 90% For unknown processing times, accuracy of statistical models with machine settings alone: R 2 ~40.1%, sensor features: R 2 ~82%

11 Monitoring Slurry Characteristics in Cu-CMP Jindal et al. 2004

12 Cu-CMP Sensing Results MRR is modeled effectively using vibration sensor data. Variations in input process parameters, i.e. slurry chemistry are modeled using sensor signals. Basis for more dynamic models to predict process performance.

13 Surface Finish Monitoring in Cu-ECP Anode contact Vibration sensor Wafer Holder ADE Phase Shift MicroXam Laser Surface Profiler Slurry Reservoir Motor Drive Unit min polish, Ra = 10-50nm 2min polish, Ra = nm 1min polish, Ra = nm As received, Ra = nm

14 Real-time MRR and Ra Monitoring

15 Summary of Current Work Physically consistent and accurate models of MRR and WIWNU are possible using Genetic Algorithms ANFIS (SC) can capture the characteristics of CMP process and provide precise prediction (< 5% prediction errors) using sparse data, and optimal parameter search as needed Nonlinear and descriptive features lead to ~13% improvement in predictability compared to the use of descriptive features alone Sensor features can act as surrogates to process parameters, and consideration of nonlinear descriptors leads to an improvement in process predictability of : 10% in Oxide CMP, 38% in Cu-CMP, 50% in Cu-ECP Effects of in-process slurry chemistry variations on MRR and Ra in Cu-ECP is captured using vibration sensor signal features Cu-ECP is optimized in 2 stages: first high MRR, second low surface roughness

16 Sensing Needs in Nanomanufacturing Sensors MEMS/NEMS wireless sensors with optimized antenna design (for close proximity monitoring, leading to enhanced sensitivity) Self-powered to obviate need for battery High throughput transmission protocol for multi-sensor fusion Feature Extraction Models that capture nonlinearity and nonstationarity of process dynamics Multi-scale representations and Recurrence analysis Applications Beyond fault detection Track variations in process dynamics to detect instantaneous variations in deposition, growth, and planarization rates Acknowledgements: The generous support from NSF Manufacturing Machines and Equipment Program (CMMI , Program Director: Dr. George A. Hazelrigg) is gratefully acknowledged