Predictive Probing: A novel approach to minimize efforts ... · A novel approach to minimize...

© Fraunhofer IISB, 11-2018

Predictive Probing:A novel approach to minimize efforts at final test

SEMICON Europa 2018TechARENA: Metrology for Emerging Technologies

Dr.-Ing. Martin SchellenbergerGroup Manager Equipment & APCFraunhofer IISB, Erlangen, [email protected]

mailto:[email protected]



Fraunhofer IISB

Schottkystraße 1091058 Erlangen

www.iisb.fraunhofer.de


Predictive Probing: A novel approach to minimize efforts at final testTechARENA: Metrology for Emerging Technologies

I. Some (rather limited) History, Part I: Integrated Metrology

II. History, Part II: „Data Analytics“ enters the metrology domain

III. The art of Predictive Probing

IV. Summary






IV. Summary


In situ spectroscopic ellipsometry in a batch furnace

Layout of the batch furnace Prism-based optical system for the in situellipsometry measurement

Realtime control of oxide growth

1990+


In-line X-ray Photoelectron Spectroscopy in a cluster tool

Transfer

and

Control

Cluster Tool

Wafer

Processing

Surface

control

by XPS

1990+

In-line control of layerproperties right after processing


In situ optical emission spectroscopy in a batch furnace

Real-time control of plasma processes by integrated OES

8

2010+

Again: Prism-based optical system






IV. Summary

ado

pte

dfr

om

gart

ner

.co

m


Measurement system mounted on a

single-wafer FOUP with adapter

Pic

ture

by c

ourt

esy o

f In

fin

eon,

Dre

sden

Stocker-integrated metrology for CD control 2000+

Measure intensity

as a function of the

azimuth angle

Diffraction signature

Fault detection:

Good / Bad

-40 -30 -20 -10 0 10 20 30 400

0.05

0.1

0.15

0.2

0.25

0.3

0.35

Azimut-Winkel (°)

Inte

nsi

taet

(a.u

.)

64M-DT-L-alle-s2s23-p8p40p72: Beugungssignaturen, Pol. 1, sortiert nach WAFNR

Azimuth angle (°)

misprocessed

wafers

Fault detection and classification (FDC) based on neural networks


Relevant data are continuously collected and analyzed using a Bayesian network.

The network predicts the real time remaining until the break of the filament with an accuracy of 10-20 hours.

Thus, no “mere” preventive maintenance after a predetermined operating time or number of processes needs to be carried out, and no system failures are risked by missed maintenance steps.

Predictive maintenance in ion-implantation 2010+

Based on these forecasts, maintenance tasks can be scheduled exactly.


Virtual metrology for deep-trench etching 2010+

Relevant data are continuously collected and evaluated using a “gradient boosting tree” algorithm.

The algorithm predicts the actual depth of the trench after the etching process with a deviation of less than 4 nm compared to values obtained from physical metrology.

Regular, costly and time-consuming physical measurements can be limited.

The application of virtual metrology allows “virtual” control of every single wafer






IV. Summary


Predictive ProbingUse-case: LED manufacturing

Several process steps, e.g.:

▪ Epitaxy to create the optically active layers

▪ Doping to achieve certain electrical properties

▪ Metallization to generate contacts

▪ Layer formation to guide the emitted light

Quality control, e.g.:

▪ Particle measurement

▪ Ultrasonic measurements

▪ Photoluminescence measurements

▪ Probing at wafer level

▪ Final test

Epitaxy Processing Packaging Light Engine

Up to 50% of total cost


100% Probing – electrical and optical measurement of every single chip:

▪ Electrical and optical properties

▪ Defect chips

Time-consuming and expensiveProbing Probing result

Predictive ProbingState-of-the-art probing in LED manufacturing


Predictive Probing

▪ Reduced probing of a selected number of chips

▪ Based on data analysis – not on mere random or statistical reduction

(reduced)

Predictive Probing

Same result, but: partly measured, partly reconstructed

full probing

Predictive ProbingThe concept


Predictive ProbingConstruction of the probing map

Optical and electrical properties:

▪ Basic test grid based on analysis of historical probing data


Optical and electrical properties:

▪ Basic test grid based on analysis of historical probing data

Defects:

▪ Analyse measurements prior to probing

▪ Individually calculate specific defect test grid for every wafer

Predictive ProbingConstruction of the probing map


Predictive ProbingNew probing approach and result

Two-step Predictive Probing process:

1. Analyse prior measurements, compile probing map to determine electrical and optical properties as well as defects and probe selected LED-chips




2. Read measurements, interpolate optical and electrical values and mark defect LED-chips



Results

▪ Accurate interpolation of LED properties

▪ Defect detection accuracy meets application requirements, almost always ...



2. Read measurements, interpolate optical and electrical values and mark defect LED-chips



Predictive ProbingImproving defect detection

… but not for wafers with edge voids:

▪ A small percentage of wafers show edge voids

▪ No accurate detection with traditional analysis methods possible, only work-around solutions

▪ Visible for the human eye in photoluminescence measurements, though


Predictive ProbingEdge Void Classification

Challenges:

▪ Measured brightness varies highly – so do edge void shapes and sizes

▪ Every single chip (>130,000) must be classified

photoluminescence measurements


Edge Void ClassificationSolution approach

Fully Convolutional Networks:

▪ Based on a special network architecture for computer vision

▪ Self-learning algorithm for pixel-wise classification


Edge Void Classification Some intuition about fully convolutional networks

Fully Convolutional Networks:

▪ Vanilla (regular) neural network process vectorised data

▪ Computer vision networks, by contrast, preserve spatial information by filtering the image


Filters: feature detectors that are robust against rotation, scale and translation variance

fur structure

cat eyes

cat ears

cat nose

fur colour

paws

whiskers

cat mouth

fur pattern



edge void

void

chips ok

edge void

voids

void



▪ A typical network contains thousands of filters, allowing the classification of highly variant images

▪ With increasing network depth filters are getting more complex

incr

easi

ng

net

wo

rk d

epth

https://storage.googleapis.com/deepdream/visualz/vgg16/index.html

“cat

”



▪ The network’s performance is to learn suitable filters for the given classification task

▪ Therefore the network has to be trained with a carefully assembled dataset of inputs and corresponding labels

https://distill.pub/2017/feature-visualization/

network training progress

randomly initialised filter trained filterearly training advanced training



Edge Void ClassificationFully Convolutional Networks

Network training:

▪ Input: about 100 photoluminescence measurements

▪ Labels: 3 prediction classes - chips ok / background / defect chips


Edge Void ClassificationResults

Predictive probing defect detection accuracy significantly improved

▪ Over 98.5 % of all 168,100 pixels correctly classified

▪ Independent of wafer / chip size

▪ Network knowledge transferrable to other pattern recognition tasks






IV. Summary


Predictive Probing to reducetime and cost for device test

▪ Relevant upstream data are continuously collected and analyzed to predict device properties without actually measuring them

▪ Basis: long-term/short-term historic data, e.g. from upstream measurements

▪ Plus: Fully convolutional neural network

▪ Accurate interpolation of LED properties with ~7% measured chips.

▪ Benefit: significant time and cost savings

Summary IPredictive Probing in LED final test

LED-chips not to be measured

LED-chips to be measured


Artificial intelligence (AI) will play an inevitable roleas enabler for “Metrology for Emerging Technologies”

▪ AI is the next step, amending “integration” and “data analytics”

▪ Semiconductor (SC) industry deals with a complex environment for production and quality control

▪ SC industry already gained a lot of experience in the AI domain

▪ Predictive maintenance, virtual metrology

▪ Combine analytics with domain knowledge (no “mere” informatics)

▪ Follow closely other AI domains!

▪ Especially in the area of big data analytics and deep learning

Summary IITechARENA: Metrology for Emerging Technologies


Acknowledgement

Parts of the work presented here was funded within

- IMPROVE (ENIAC Joint Undertaking)

- SEAL (EU Seventh Framework Programme)

- EPPL (ENIAC Joint Undertaking)

- INTEGREAT (German BMBF)


Thank you for your interest!

Dr.-Ing. Martin SchellenbergerGroup Manager Data AnalyticsFraunhofer IISB, Erlangen, [email protected]

mailto:[email protected]

Date post:	16-Mar-2020
Category:	Documents
Upload:	others
View:	2 times
Download:	0 times

Predictive Probing: A novel approach to minimize efforts ... · A novel approach to minimize...

Documents