Transcript Relationship Between in-situ Metrology and ex

Relationship Between in-situ
Information and ex-situ
Metrology in Metal Etch
Processes
Jill Card, An Cao, Wai Chan, Bill Martin, Yi-Min Lai
IBEX Process Technology,
A division of NeuMath, Inc
Outline
●
Background
What we want for APC
The current situation in IC fabrication
●
Project Overview
Product design
Data collection
Model structure
●
Results
Background
Ideal Semiconductor Fabrication:
Processes running on target
Continuous process monitoring and control at the tool level
Impending scrap events immediately detected and prevented
Advanced Fault Detection
Reliable Root Cause Analysis
Heads-up for tool failures
Pinpoint problems and advise maintenance actions
High Yield by coordinating different steps and processes
Current Fabrication Situation
5
Chart “Violates”
6 Lot Goes on Hold
Yellow Light On
MT Takes Action
4
Data to Process
Tool SPC Chart
2 Lot Moves to Measurement Tool
Delay!
1
Lot is
Processed
3 Lot is Measured
Production line may be running for
5 lots with scraps before scraps are detected
– at a cost of $$$ per lot.
Solution?
5
Chart “Violates”
Tool SPC Chart
In-situ data is readily
available, no delays
ex-situ data enhances
the model
NN Model
Predicted
ex-situ
2 Lot Moves to Meas. Tool
1
Lot is
Processed
3 Lot is Measured
The Proposal
Suppose
We can build a map between in-situ
information and ex-situ metrology, then we can
use in-situ data to predict the wafer quality directly,
thereby avoiding the metrology delay.
Direct benefits
Real time monitoring of wafer quality
Predictions available for every single wafer
Avoid delay in detection of major scrap events
Take advantage of increasing availability of in-situ
data, e.g. sensor data.
Potentially reduce ex-situ measurement cost
Experiments
We seek answers to these questions:
Can we accurately predict ex-situ information using
in-situ results?
If yes, is there a relationship that can be easily
interpreted?
Data Collection
●
●
●
Production data from Metal Etch process
4 months of data, total = 30K records.
About 1.3K records have ex-situ
information collected.
Modeling one critical etch step
Inputs includes feed-forward metrology
information from the previous steps.
Neural Network (NN) Models
• Neural Network modeling was
chosen because the relationship
between in-situ and ex-situ metrology
is hard to formulate mathematically.
• NN learns the rules from the dataset
itself, no prior knowledge is required.
• IBEX Dynamic Neural Controller
[commercial software package] was
used.
• Separate neural network models are
built for each ex-situ metrology
measurement.
Model Inputs vs. Outputs
TCP RF Forward Power
Bias RF Forward Power
Temperature Upper Sense
Temperature Bottom Electric Sense
Temprature Turbo Manifold Sense
Tempature Vat Valve Sense
Chamber Pressure
Chamber Clamp Pressure
Chamber ESC Voltage
TCP RF Reference Power
TCP Match Tune Cap
TCP Match Phase Error
TCP Line Impedance
TCP Match Load Cap
Bias Match Load Cap
Bias Match Tune Cap
Bias Match Peak Voltage
Bias RF Ref Power
Chamber Ref Manometer Pressure
Chamber Pressure Valve Angle
Chamber Clamp Flow
Chamber End Point Channel A
Chamber End Point Channel B
Chamber ESC Current Leak
DICD_Mean
DICD_Std
Inter Layer Dielectric Deposition
Post CMP Thickness
Post CMP Thickness Nonuniformity
Percent Open Area
DFT_DICD
Outputs
FICD_mean
FICD_std
FICD_slope
DefectDensity
Results
We sought to answer these questions:
1. Can we predict ex-situ information with insitu results, accurately?
Yes!
2. If yes, is there an easily-determined
relationship?
Model Accuracy
Outputs
Accuracy Records used
FICD_mean
0.53
1254
FICD_std
0.92
1254
FICD_slope
0.93
1225
DefectDensity
0.80
99
Note:
Prior metrology is important!
Prediction Fitting Curve
FICD_Mean
0.70
0.65
0.60
0.55
0.50
0.45
0.40
0.35
0.30
12/25/03
01/14/04
02/03/04
02/23/04
Measured
03/14/04
04/03/04
Predicted
Accuracy = 0.53, r2=0.95
04/23/04
05/13/04
Accuracy Depends on
Limits Setting
Recipe 13
0.57
0.55
0.53
0.51
0.49
0.47
0
50
100
150
200
250
Measured
Predicted
Target_Lower_Limit
Target_Limit
Target_Upper_Limit
Soft_Upper_Limit
Accuracy = 0.95
300
Accuracy for A Different Recipe
FICD_Mean - Recipe 9
0.58
0.53
0.48
0.43
0.38
0
10
20
30
40
Measured
Predicted
Target_Lower_Limit
Target_Limit
Target_Upper_Limit
Soft_Upper_Limit
Accuracy = 0.61
50
Prediction Fitting Curve
FICD_Slope
90
89
88
87
86
85
84
83
12/25/03
01/14/04
02/03/04
02/23/04
03/14/04
Measured
Accuracy = 0.93
04/03/04
Predicted
04/23/04
05/13/04
Prediction Fitting Curve
FICD_STD
0.030
0.025
0.020
0.015
0.010
0.005
0.000
12/25/03
01/14/04
02/03/04
02/23/04
03/14/04
Measured
Accuracy = 0.92
04/03/04
Predicted
04/23/04
05/13/04
Prediction Fitting Curve
DefectDensity
7
6
5
4
3
2
1
0
12/25/03
01/14/04
02/03/04
02/23/04
Measured
03/14/04
04/03/04
04/23/04
Predicted
Accuracy = 0.80. Limited number of observed
records may affect the model accuracy.
05/13/04
Sensitivity Analysis
We sought to answer these questions:
Can we predict ex-situ information with in-situ results,
accurately?
Yes! We successfully predicted ex-situ
metrology from the in-situ metrology with
reasonable accuracy (ranging from 0.5 to 0.9)
If yes, is there an easily-determined relationship?
No. It requires Sensitivity Analysis.
Sensitivity Analysis
Recipe 1
Bias Match Voltage
DICD Mean
Complicated
relationship.
FICD depends
on multiple
inputs
Temp Turbo Manifold Sensor
Temp Turbo Manifold Sensor
Sensitivity Analysis
Recipe 2
DICD Mean
Sensitivity is
also recipe
dependent
Temp Turbo Manifold Sensor
Temp Turbo Manifold Sensor
Sensitivity Analysis
Recipe 2
Other ex-situ
metrologies
show similar
complicated
sensitivity
curves. An
example,
FICD Slope, is
shown.
Sensitivity of ex-situ metrology
Ex-situ metrology depends on complicated
interactions among the trace inputs and the
feed forward metrology.
Recipe-dependence
Non-linear sensitivity curves
Possible dependence on tool health situation
Sensitivity changes over time
This demands an intelligent algorithm for
better interpretation.
Output Dependency on Inputs
Variables
FICD_mean FICD_std FICD_slope DefectDensity
TCP RF Forward Power
Bias RF Forward Power
Temperature Upper Sense
X
Temperature Bottom Electric Sense
X
Temprature Turbo Manifold Sense
X
X
X
X
Tempature Vat Valve Sense
X
Chamber Pressure
X
Chamber Clamp Pressure
Chamber ESC Voltage
X
TCP RF Reference Power
TCP Match Tune Cap
X
TCP Match Phase Error
X
TCP Line Impedance
TCP Match Load Cap
X
Bias Match Load Cap
X
Bias Match Tune Cap
X
Bias Match Peak Voltage
X
X
X
X
Bias RF Ref Power
X
X
Chamber Ref Manometer Pressure
X
Chamber Pressure Valve Angle
Chamber Clamp Flow
Chamber End Point Channel A
X
Chamber End Point Channel B
X
Chamber ESC Current Leak
X
X
DICD_Mean
X
X
X
X
DICD_Std
X
Inter Layer Dielectric Deposition
X
X
Post CMP Thickness
X
X
Post CMP Thickness Nonuniformity
X
X
Percent Open Area
X
X
DFT_DICD
X
X
X
Summary
●
●
Our previous work** shows comprehensive
root cause analysis through neural model of
all metrology outputs (in-situ and ex-situ) and
controllable variable inputs.
Recommends corrective action Wafer to
Wafer
maintenance actions
setpointed recipe parameters.
** Card, et. al. Fab Process And Equipment Performance Improvement After An Advanced Process
Controller Installation, AEC/APC-Europe 2004
Recommended optimal
Repair or Recipe adjustment
Gas flow
Pressure
Temp
Conditioning Run
Wet Clean
Replace MFC
Replace Quartz
Replace Chuck
HGS
Replace vat valve
Etch Rate
Uniformity
Selectivity
Particles
Valve Angle
He Clamp Flow
Wafer Area Pres.
Ex-situ
In-situ
Next Steps
●
●
By prediction of ex-situ measures with
precision, DNC can provide root cause
analysis for tool health and process health
without reliance on ex-situ measures.
Addition of more complex sensors (RF
probe, OES) may well add the remaining
information content to complete ex-situ
characterization
Recommended optimal
Repair or Recipe adjustment
Gas flow
Pressure
Temp
Conditioning Run
Wet Clean
Replace MFC
Replace Quartz
Replace Chuck
HGS
Replace vat valve
Valve Angle
He Clamp Flow
Wafer Area Pres.
In-situ
OES
RF Probe
Conclusion
Accurate predictions of ex-situ metrology can be achieved
from in-situ information only.
Next Steps
Introduce root cause tool control algorithm for maintenance
and recipe parameter response.
Continue evaluation of complex sensors to further enhance
ex-situ metrology prediction using in-situ sources only.
Sensitivity analysis
Complex relationship to ex-situ metrology. However, if
information present, root cause optimization can follow
with no loss of precision.