FROM M DEEP LE P LEAR ARNI NING NG TO NE NEXT-GEN N VI VISU - - PowerPoint PPT Presentation

FROM M DEEP LE P LEAR ARNI NING NG TO NE NEXT-GEN N VI VISU SUAL ALIZA IZATION ION: : A G A GPU PU-PO POWE WERED RED DIG IGIT ITAL AL TRAN ANSF SFORMAT RMATION ION

Ingrid Tobar

Senior nior Data Scien entist tist An Anadar darko

Amit Vij

San Jose, California March 18, 2019

Preside sident nt & C Co-Founder nder Kinetica tica

OB OBJE JECTIVE: CTIVE: AT ATTE TEMPT MPT TH THE IMP E IMPOS OSSIBLE SIBLE

Visualize and interact with a very high fidelity 3D representation of the Del elaware ware Ba Basin in for hydrocarbon exploration

3

120 mi 80 0 mi mi The e Delawar aware e Basin sin is s ro roughl ghly th the e si size of Mass ssachu achuse setts tts… and d 3x th the e heigh ght t of th the Empi pire re Sta tate te Buildi ding ng 4,000 00 ft

9M 9M to 90B 90B

Point Reservoir Model

Age genda nda

• About Anadarko
• GPU-Enabled Tech and Projects
• About Kinetica
• GPU-Accelerated Visualization

5

About

• ut Anada

adarko rko

Cor

• rporate
• rate St

Stra rategy tegy

ENHANCING SUCCESS IN DWGOM EXPANDING LOWER 48 FOOTPRINT ENABLING DIGITA TAL OPERATIONS

Advanced geophysical analytics to enable exploration with tiebacks to existing infrastructure High density Lower 48 subsurface characterization to provide optionality “Intelligent” control and edge computing in Drilling, Completions and Production

DJ BASIN DELAWARE BASIN DEEPWATER GOM

ANADARKO STRATEGIC FOCUS AREAS

TOTAL L VOLUME UME

700 700

MBOE/D CAPEX EX

$4.8 .8B

FOCUS US AREA EAS

DJ Basin, Delaware Basin, DWGOM

FUTU TURE RE VALUE UE

Exploration & LNG

LEGEN END

6

CASH H GENERA NERATION ION

DWGOM & International Oil

AA AAET: T: A Adva vanced ced An Analy lytic tics s and Emer ergin ging g Te Tech chnology

• logy

7

# TEAM MEMBE BERS

2016

2019

50 50

Person son Core re Team

PLATFOR ORM DEPLOYME YMENT NT

2016 2019

# Platforms rms

5

Dom

• mai

ain Pla latf tfor

• rms

ms

PROJ OJECT CT PORTFOL OLIO

2016

2019

25+ 25+

Proj rojects

# Pro rojects ts

TEAM DEMOG OGRAPHI APHIC

Data ta Scien enti tist sts Engine ineer ers Geosc scientis tists ts Data taOp Ops s & Dev evOp Ops

Team Formed APC Announces Google Partnership

STAKEHOLDER ENGAGEMENT FOCUS 2018 18

Strategic Alliance with RE Energy Group

PRODUCTIZATION STRATEGY DEVELOPED 2017 2019 DEPLOYMENT AT SCALE THROUGH PLATFORMS

• Dr. Sean Gourley

appointed to Board

INCEPTION OF DATA SCIENCE SKILLS IN APC 2016

NVIDIA GPU Technology Conference Kinetica Visualization Project Kick-Off

Exploration

• ration

Identifying swe weet et spots ts where well performance is high and land entry costs are low can generate significant value to the company

Development lopment

Selecting the optimal mal we well design gn – which involves choices in numerous areas such as completion size and well spacing – requires predicting the performance for each candidate design

Opera rations tions

Monitoring and understanding asset et behavior

• r through the life-

cycle of well construction (drilling) to extraction of underground resources (production)

Op Oper erationa ationalizing lizing Di Digi gital tal

8

Seism smic c Inter terpret retati tion

• n

Str Strati tigra graphi phic c Top Corr rrelati tion

• n

Real-Time Time Dri rilling ling

ENHANCING SUCCESS IN DWGOM EXPANDING LOWER 48 FOOTPRINT ENABLING DIGITAL OPERATIONS

GP GPU-Ena Enabled bled Te Tech ch and nd Proj rojects ects

• Envi

vironment ment

• 1.5 yr. ago: DGX

GX-1 1 8x Tesla la P-100 100 GP GPUs

• Today:

y:

□ DGX-1 8x Tesla V-100 GPUs & □ DGX-2 16x Tesla V-100 GPUs

• Proje

ject ct Sco cope

• Seismic

ismic int nterpre erpretation ation deep ne neur ural al ne network work model el for image ge process essing ing

• Data Volu

lume

• 100s GB

GB – several eral TB

• 1000s images/attr

ages/attribu ibutes tes

• Training

ining on 1% data

• Inf

nferen rence e across

• ss

full l image ge

• Fr

Framew ewor

• rk
• TensorF

sorFlow/Py low/PyTor Torch

• 2 c

concurr rren ent fault lt predicti iction

• n

models els

Se Seis ismic ic Interpretation terpretation

10

Se Seis ismic ic Interpretation terpretation

11

Se Seis ismic ic Interpretation terpretation

12

Currently, the models only use 1% of the data for training, however, inference is performed

• n the entire image.

This means that we need to dedicate significant amounts of time to training in

• rder to deliver good inferences.
• Benefits

efits

• Training

ining and nd inf nfere erenc nce

1.5 yr. ago: ~20

20 hours

Today: <10

10 hours

• Challenge

llenges

• Time

me intensiv ensive e training aining process ess

• Loadi

ding g data a into

• GPU

U memo mory ry

• Next

xt Steps

• Netwo

work rk enhancements ncements

• Workf

kflow low impro provemen vements ts

• New DGX-2

2 box:

□ 16

16x x V-100 00 GPUs Us + 512 GB GPU U Memor

• ry
• Future

ure Environm ironment: ent:

□ Google Cloud Platform (GCP)

• Proje

ject ct Scope

• Learn

rn from m identified ntified top tops propa pagate gate at basin in scale le

• Data Volu

lume

• Trainin

ining g ~25GB GB

• Inferen

rence: e: Size ze varies ies,

• n the fly
• Fr

Framew ewor

• rk
• CNN in TensorF

sorFlow low

• Envi

vironmen ment

• De

Dev/Train /Train (on n prem.) m.):

□ DGX-1 8x Tesla V-100 GPUs

• Inferen

rence/U e/UI I (on cloud) d):

□ GCP V-100 and T4 GPUs

St Stra ratigraph tigraphic ic To Top Cor

• rrelation

relation

13

Seed Well Seed Well Seed Well Seed Well

18,000 000 ft

Challenges Next Steps Benefits

St Stra ratigraph tigraphic ic To Top Cor

• rrelation

relation

14

• Short term
• Bet

etter er net etworks ks as CNN runs

• ‘Self-tuni

tuning’ mechanism

• Lo

Long term

• Move

e workfl rkflow w to cloud

• ud
• New T4 GPU for

r in infer eren ence ce

□ In

n GC GCP since nce Jan n 2019

• Faster

er train ining ing wit ith new w GPU chip ips

• GPU

PU Qua uadr dro

• P6

P6000: 000: co coup uple e weeks

• DGX

DGX-1 1 8x Tes esla P-100 00: : 1.5 5 – 2 days

• DGX

DGX-1 1 8x Tes esla V-100 00: : < < 24 24 hours

• Acc

ccele elerat ated ed basin in eval valuation tion proce cess ss

• Massiv

ive e data vo volu lumes mes

• Rap

apid idly gr growin ing g ge geo da data

• Pic

icked ed and d in infer eren ence ce wel ells

• Expe

pert t pi pickin ing g (label elin ing)

• Tim

ime e in intensiv sive e train ining ing pr proce cess ss

• CNN train

inin ing

• GPU tec

ech h ad advan ances es

TGS, Wood Mackenzie PetroView

• Proje

ject ct Sco cope

• Dr

Drilling illing Ops: : $M d decisions isions

• Analytics

lytics and DL models ls process ess real-time time streaming reaming log data a & & other er non-stre streaming aming data

• Rig states → De

Deriv ive e opera rational tional KPIs of

• f

drill llin ing g ops at very high resolution solution

Re Real-Time Time Dr Dril illing ling

15 Real-Time Drilling

3D Plot – Plan vs. Actual Actual Plan Offset Well

3D Traject ector

• ry

Plan vs actual and

• ffset trajectories

Side View

Target Line Landing Projection Plan Actual Upper Window Lower Window Projection

Real-Time Drilling

Current Deviation

Directi ction

• nal

al Guidan ance ce

Plan vs actual charts, directional efficiency and guidance, projections

ROP Footage Time Footage % On Bottom Time Off Bottom Time Connection Time Connection Time

Operati rationa nal KPIs s at high resolut ution

• n

Footage, Time, ROP, connection statistics

• Data Volu

lume

• Trainin

ining g dataset aset ~5GB GB

• Inferen

rence: e: Real-time time streaming reaming

□ Sliding window partition data stream □ Runs online 24/7

• Fr

Framew ewor

• rk
• RNN:

: > t trainin aining, g, no paralleliz allelization ation

• CNN:

: Curren rrent t model, el, TensorF sorFlow low

• Envi

vironment ment

• Dev/Train

/Train (on prem.): m.):

□ DGX-1 8x Tesla V-100 GPUs

• Inf

nferen rence e (on n cloud) ud):

□ Google Cloud ML Engine

Re Real-Time Time Dr Dril illing ling

16

Benef efits its

• Very

y light ht for inferenc erence e

• High

gh res. . KPIs Is to evaluat uate drilling illing performanc

• rmance

e and correc rect t traject ajectories

• ries

Challenges llenges

• Real-time

time inf. . requir ires es fast t response ponse

□ <100

100 millise lisecon cond for each inference erence

□ >1 sec

c → heavy traf affic ic, , poten enti tial al jams ms

Next xt St Steps

• Shor
• rt-term

erm

□ Offline

ine model el using ng histor

• ric

ical al data

□ Divergenc

rgence e Detec tect: : Real eal-time time vs offline line

• Long-term

erm

□ More

e comple lex models, els, more re data

THE ACTIVE ANALYTICS PLATFORM Amit Vij

Pres esid iden ent & Co & Co-Founde

• under

Kin inet etic ica

18

“NVIDIA and Kinetica have enabled us s to do th

• do the impo

imposs ssible ible — re render der a h a hig igh h fid idelity elity, , 3D vie D view w of

• f an

an oi

• il

l ba basi sin n usi sing ng 100 billion data points at scale.”

Sanjay Paranji, CTO at Anadarko Petroleum Corporation

ACTIVE ANALYTICS PASSIVE ANALYTICS

9M

Low Fidelity / Small Sample Set

90B

High Fidelity / Full Data Set + Streaming

AF AFTER ER GPU PUS BEFORE ORE GPUS

Hig igh h Fi Fide delity lity 3D 3D Vi Visualization sualization

20

Kin inetica etica Ac Acti tive ve An Analy lytics tics Pl Platf tform

• rm

21

UNIFIED ED DATA MODEL, EL, DEVELOPER LOPER APIS AND TOOLS LS STREAMING AMING ANALYTIC YTICS HISTORICAL AL ANALYTIC YTICS ML ML-POWERED ERED ANALYTIC YTICS LOCATIO TION INTELLIG LLIGEN ENCE ENTERPRISE SE GRADE GPU-AC ACCELE ELERATED TED CLOUD-READ ADY

In Yo Your r Eco cosy syst stem em

22

CUSTOM M APPS APPS BUILT-IN IN REPORTIN TING OFF THE SHELF BI & REPORTI TING ANALYTIC YTICAL AL MICROSER ERVICE VICES

HISTORI RICAL CAL ANALY LYTIC ICS STREAM REAMIN ING ANALY LYTIC ICS ML ML-POWE WERED RED ANALY LYTIC ICS LOCATIO ION INTEL ELLIGENCE LIGENCE DATA PIPELINEs | ETL STREAMS

EDGE CRM ENTERPRISE SYSTEMS OPERATIONAL SYSTEMS REAL-TIME DATA HISTORICAL DATA

GPU-ACCELERATED DATABASE

AC ACTIVE E ANALYTI TICS PLATFORM TFORM DATA LAKE ENTERPRISE DATA WAREHOUSE

GPU GPU-Ac Accelerated celerated Da Data taba base se

23

PARALLEL ORCHESTRATION ENGINE

In In-Memor Memory OLAP P Databa base se

• Distrib

tributed uted

• Column

umnar ar

• Vector
• rized

zed

• SQL 92 Compli

pliant nt Wide e Range e of Analyt ytical cal Techn hniqu ques es

• Fu

Full text xt sear arch

• Time

me series es analysis sis

• Locati

ation

• n intell

elligen gence ce

• Grap

aph analyt ytics cs Enter erpr prise se Scale e Tiere ered d Stor

• rag

age

• Access

ess entire e data a corpus us

• Opti

timi mized d perfor

• rmance

mance

Kin inetica etica TI TIERE RED D ST STOR ORAG AGE

24

Data movement across tiers automatically managed by Kinetica

1 month of data 6 months of data 2 years of data 7 years of data

GPU Memory RAM System m Memory Disk k / SSO HDFS FS / S3

NEW W IN 7.0

Loc

• cation

tion Inte tellig lligenc ence e and Vi Visu sualiza alization tion

25

GEOSP SPATI TIAL AL APIS GRAPH ANALYTI TICS COMPLE LEX X GEOSPATIAL TIAL OPERATIO TIONS GEOSP SPATI TIAL AL SERVER ER

DISTRIBUTE TED GPU PIPELIN LINE 3D TILES ES

What We’re Doing Today

26

GP GPU-Acc Accelerated elerated Vi Visu sualiza alization tion

• Two-St

Stage age Reser servoir voir Modeli ling ng Strategy egy

• Challenge

llenges

Mot

• tiva

ivation tion

28

• Lar

arge ge AOI

□ Basin

in ext xten ent

• Subset

t Data

□ Reduced

uced attr trib ibut utes es

• Ren

ende derin ing g & & pr proce cess ssin ing g on CPU Coarse Resolution Model

• 3D Reg

egio ion of Inter eres est

□ Reduced

uced covera erage ge

• Mode

del at Scale

□ Repeat

eat process ess in n neigh ghbori boring g areas as

High Resolution Model

• Tim

ime-consumi consuming ng

□ Geologists

logists & reser ervoir

• ir

modelers elers perform

• rming

ing tasks sks in sequenc ence

• Incl.

. relevant t in info. . in in de decis isio ion-maki making

□ Ne

New workflo rkflows ws require uire viz & render erin ing g of

• f

massi ssive e data a volumes mes

• Incom
• mpa

pati tibil bilit ity y bet etween een mode dels ls

□ Condit

ditional ional to differen erent t inputs uts (defin ined ed by AOI/R I/ROI) OI)

Ap Appro roach ach

29

Buil ild Hig igh-Res es Model Vis isuali lize ze Data in in 2D Vis isuali lize ze Data in in 3D

• Siz

ize e & & Dim imen ensions sions

• 90 bil

illio ion po poin ints

□ ~4k slices

es (layer ers) s)

□ ~24 mill.

• ll. pts/la

s/layer er

• >

> 10,000 0,000 sq. mi. i.

• 100

00ft t XY (spa patia ial) res esoluti lution

• n
• 1ft Z (

Z (de dept pth) res esoluti lution

• n
• Data Inges

gestion ion

• 6TB of po

poin int da data

• Reveal

veal

• Kinetica’s viz

iz frame mewor

• rk

k as en end- user er web eb clie ient

• Layer

ers dr drape ped d over er base e map

• User-Def

Define ined d Funct ctio ions ns (UDF) F) Framewor ework

• 3D til

iles es on the e fly

• Rep

epres esen enta tati tion

• n of

reservoir

• ir mode

del

• CesiumJ

iumJS

• Geospati

tial l 3D mappin ing platform

• rm

to rende der til iles

30

Re Resu sults lts

31

• Reve

veal al dashboard wit ith model l lo loaded d and poin ints geosp

• spatially

tially refere ferenced ced

Rapid d naviga gati tion

• n

throu

• ugh

gh ~4k model

• del layer

ers

Re Resu sults lts

32

Map proper perty ty (porosi

• sity

ty) ) can be interr errogat ated ed at any point t locati tion

• n

Point nt resolut solution ion varies es accor

• rdi

ding ng to zoom

• m level

el

Re Resu sults lts

33

Select lect and filter er recor cords ds by spatial l extent ent (drawing ng a shape) pe) Point nt coun unt t up updates es when hen spatial al filter er is applied ed

Re Resu sults lts

34

Quick executi cution

• n
• f calculat

ation ions s on layer er range ge and genera erati tion

• n of

deri rivati tive e data (e.g. .g. Calculat ulate e average age poros

• sit

ity) y) Select lect a range ge

• f layer

ers (1-10) 0)

Re Resu sults lts

35

Slider der bars s enable able smoo

• oth

th moveme ement nt through

• ugh horizon

izontal tal slices es of the e model

• del

Re Resu sults lts

36

High resolut solution ion

• f mode

del mainta ntaine ined d in 3D view Ful ull model

• del

volume me is cut ut by inter ersect ecting ing cross ss- secti tional

• nal

slices es

Re Resu sults lts

37

Map proper perty ty can be interr errogat ated ed by se select ecting ing a point nt of interes erest

Ne Next t St Step eps

38

• Inter

eraction ction

• 2D &

D & 3D w D well logs gs

• More

e models els & pet etroph

• ph.

. attribut tributes es

• Comple

plex x and stored

• red

calculation lations

• Vis

isuali lizat zation ion

• High

gh-res res rendering dering

• UI Enhanceme

cement nts

• Integrat

egrated d views ws

• Well log

g displa lay and hist stogram

• gram
• Cross
• ss-sections

sections

1.

• 1. Dig

igita ital l transfo format mation: ion: Le Leverage verage GPU U techn chnology logy to derive ive co commer merci cial l

• utco

comes mes and support our co corporate te strategy egy 2.

• 2. Clo

loud and GPU technology logy im impr provement

• vements

s wil ill l all llow w us to m make ke decisio cisions ns wit ith more acc ccuracy acy, , faster er 3.

• 3. Kinetica’s GPU-

Acc ccele elerat ated ed Databa base se technology logy giv ives es us a n new w way way to vi visualize lize and in interact act wit ith massiv ive e data

Clo losin sing

39

40

POSTER TER SESSI SION ON

• Monday

day, , Mar 18, , 6:00 0 PM - 08:00 00 PM

• SJCC Upper

Conco course se

Ac Ackn knowled

• wledgments

gments

41

Richard ard Sech

Geolog

• logis

ist

Yuxi xing g Ben

Data Scientis entist

Dingzho zhou Cao

Data Scientis entist

Yanya yan n Zhang

Data Scienti entist

Ping Lu

Data Scienti entist

Seth th Bra raze zell ll

Geolog

• logis

ist

Alex x Baye yeh

Data Scientis entist

Ryd ydel l Pere reir ira

Soluti tion

• n Ar

Archi hite tect ct

(Frontend)

Shawn Axs xsom

Devel veloper

• per

(Frontend)

Zhe Wu

Devel veloper

• per

(Ingestion Infrastructure)

Chad Ju Juliano ano

Soluti tion

• n Ar

Archi hite tect ct

(Backend)

Nohyun yun Myu yung

Soluti tion

• n Ar

Archi hite tect ct

(Environment Setup)

Solongo go Erd rdenekhu ekhuyag yag

Projec ject t Manager ger

THAN ANK YOU

QUESTIONS?