machine learning applied to early formation …...advanced gas interpretation hc flags simplified...

28
© F Bataller, Petrophysics Specialists, Repsol E&P, December 2019. MACHINE LEARNING APPLIED TO EARLY FORMATION EVALUATION USING MUDLOGGING DATA LESSONS LEARNED, LIMITATIONS AND FUTURE APPLICATIONS Francisco J. Bataller © F Bataller, Petrophysics Specialists, Repsol E&P, December 2019. Data Science in Petrophysics Seminar London, 5 th December 2019

Upload: others

Post on 28-Dec-2019

4 views

Category:

Documents


1 download

TRANSCRIPT

Page 1: MACHINE LEARNING APPLIED TO EARLY FORMATION …...Advanced Gas Interpretation HC Flags Simplified Multimin. Model is required since Well-2 does not have full WL logs, s s, l. WORKFLOW

© F

Ba

talle

r, P

etro

phy

sics

Spec

ialis

ts, R

epso

lE&

P, D

ecem

ber

20

19

.

MACHINE LEARNING APPLIED TO EARLY

FORMATION EVALUATION USING

MUDLOGGING DATALESSONS LEARNED, LIMITATIONS AND FUTURE

APPLICATIONS

Francisco J. Bataller

© F Bataller, Petrophysics Specialists, Repsol E&P, December 2019.

Data Science in Petrophysics SeminarLondon, 5th December 2019

Page 2: MACHINE LEARNING APPLIED TO EARLY FORMATION …...Advanced Gas Interpretation HC Flags Simplified Multimin. Model is required since Well-2 does not have full WL logs, s s, l. WORKFLOW

© F

Ba

talle

r, P

etro

phy

sics

Spec

ialis

ts, R

epso

lE&

P, D

ecem

ber

20

19

.

INTRODUCTION

Can we predict PhiT and SwT using ML techniques only withMudlogging data and LWD RT GR?

Mudlogging data is always available and sometime it arrives earlierthan LWD.

MVP to explore the potential of a data driven approach (Machine Learning) to utilize this dataset for formation evaluation.

No FLAIRNo LWD ResistivityBasic Case: Only Mudlogging data + Real Time GR

Understand limitations, potential improvement and potentialapplications.

The aim is not to replace WL or LWD logs or petrophysical analysis. It is to obtaininformation earlier and when logs are not available.

Page 3: MACHINE LEARNING APPLIED TO EARLY FORMATION …...Advanced Gas Interpretation HC Flags Simplified Multimin. Model is required since Well-2 does not have full WL logs, s s, l. WORKFLOW

© F

Ba

talle

r, P

etro

phy

sics

Spec

ialis

ts, R

epso

lE&

P, D

ecem

ber

20

19

.

INTRODUCTIONWhy now?

- Expertise in Advanced Mudlogging (Repsol’s CustomAdvanced Gas Analysis App in Geolog, PDGM)

- Increasing knowledge in Data Analytics- Easy Data Analytics within daily use software (Facimage, Geolog, Paradigm)- Abundance of data- Automatized drilling rigs (less influence of the “driller’sbreak”)- Perfect opportunity in exploration drilling campaign (3 wells with the same rig, logs, etc)

Beda and Tiwary, 2011

Page 4: MACHINE LEARNING APPLIED TO EARLY FORMATION …...Advanced Gas Interpretation HC Flags Simplified Multimin. Model is required since Well-2 does not have full WL logs, s s, l. WORKFLOW

© F

Ba

talle

r, P

etro

phy

sics

Spec

ialis

ts, R

epso

lE&

P, D

ecem

ber

20

19

.

HOW MUCH DATA SHOULD WE USE?

Initial Repsol-SLB talk about ML

project

1st ModelAttempt (SLB-

REPSOL) (Brazil pre-salt)

2nd ModelAttempt (SLB-

REPSOL)(Algeria)

POC (SLB-REPSOL)(Algeria)

PHASE 1?? Initial thoughts

100+ wells 5? wells 6 wells 2 wells

Libya: 15+? WellsUK: 15+? wells

Libya: 15+? wellsLibya: 15+? wellsLibya: 15+? wells

Data suggested/used for previous Machine Learning R&D projects (with SLB) for Phi and Sw prediction from Mudlogging data

Page 5: MACHINE LEARNING APPLIED TO EARLY FORMATION …...Advanced Gas Interpretation HC Flags Simplified Multimin. Model is required since Well-2 does not have full WL logs, s s, l. WORKFLOW

© F

Ba

talle

r, P

etro

phy

sics

Spec

ialis

ts, R

epso

lE&

P, D

ecem

ber

20

19

.

HOW MUCH DATA DO WE REALLY HAVE?An example…

6161 Datapoints 5961 Datapoints

Well-1 (only 12.25in Section, step: 0,5in, 0,1524m)Gas, oil, condensate and water bearing

Well-2 (only 12.25in Section, step: 0,5in, 0,1524m)Water bearing

Data suggested/used for previous Machine Learning R&D projects (with SLB) for Phi and Sw prediction from Mudlogging data

Initial Repsol-SLB talk about ML

project

1st ModelAttempt (SLB-

REPSOL) (Brazil pre-salt)

2nd ModelAttempt (SLB-

REPSOL)(Algeria)

POC (SLB-REPSOL)(Algeria)

PHASE 1?? Initial thoughts

100+ wells 5? wells 6 wells 2 wells

Libya: 15+? WellsUK: 15+? wells

Libya: 15+? wellsLibya: 15+? wellsLibya: 15+? wells

HOW MUCH DATA SHOULD WE USE?

Page 6: MACHINE LEARNING APPLIED TO EARLY FORMATION …...Advanced Gas Interpretation HC Flags Simplified Multimin. Model is required since Well-2 does not have full WL logs, s s, l. WORKFLOW

© F

Ba

talle

r, P

etro

phy

sics

Spec

ialis

ts, R

epso

lE&

P, D

ecem

ber

20

19

.

DATA AVAILABLE

Only Mudlogging data and Real Time GR

Explore the potential of a mínimum dataset case

- Clastic reservoir- Slope-channel (Well 3)

and turbidite sheets(Well 1 and Well 2).

- Porosities around 25%

Exploration campaign- Drilled by same Service

Co. And tools.- Additional FLAIR (GSS)

and LWD Resistivity, but not considered forbuilding the modelssince we are exploringthe mínimum case scenario.

What if we selectively train the predictive model to by “showing” only

To be drilledFor trainning

Page 7: MACHINE LEARNING APPLIED TO EARLY FORMATION …...Advanced Gas Interpretation HC Flags Simplified Multimin. Model is required since Well-2 does not have full WL logs, s s, l. WORKFLOW

© F

Ba

talle

r, P

etro

phy

sics

Spec

ialis

ts, R

epso

lE&

P, D

ecem

ber

20

19

.

GROUND TRUTH

Core Data Core XRD

Well Logs

LWD logs

SwT

PhiT

2 Wells available:Well-1 (HC bearing)

Well-2 (water bearing)

Well-1

Full Multimin. model

Well-1

SimplifiedMultimin. model

Well-2

SimplifiedMultimineral

model

Well-2

DB DeterministicInterpretation

Res Flag

Pay Flag

Cutoffs

Well-1

MDT/XPT + Advanced Gas Interpretation

HC Flags

Simplified Multimin. Model isrequired since Well-2 does nothave full WL logs

Page 8: MACHINE LEARNING APPLIED TO EARLY FORMATION …...Advanced Gas Interpretation HC Flags Simplified Multimin. Model is required since Well-2 does not have full WL logs, s s, l. WORKFLOW

© F

Ba

talle

r, P

etro

phy

sics

Spec

ialis

ts, R

epso

lE&

P, D

ecem

ber

20

19

.

WORKFLOW

LWD GR

Gas DataDrilling

Parameters

SwT

PhiT

Res Flag

Pay Flag

HC Flags

Training DatasetSpecific Parts of Well-1 12,5” section (around

46%)

Validating DatasetSpecific Parts of Well-1

well 12,25” section(around 54%) and all

Well-2

Building predictiveModel

To Predict SwT, PhiT, Reservoir, Pay and Fluid Type Flags.

Software used will be Geolog(Facimage), which is simple butpowerfull for building these typeof models.

Models available:- ANN: Artificial Neuronal Networks- MRGC: Multi Resolution Graph

Based Clustering- AHC: Ascendant Hierarchical

Clustering- SOM: Self Organizing Maps- STM: Similarity Threshold Method

SwT PhiT Res FlagPay Flag HC Flags

Applying predictive Modelin validating dataset

LWD GR

INP

UTS

OU

TPU

TS

Mudlogging Conv. Gas Data

Model Outputs

Val

idat

ing

ou

tpu

ts v

s gr

ou

nd

tru

th

Split Dataset

Page 9: MACHINE LEARNING APPLIED TO EARLY FORMATION …...Advanced Gas Interpretation HC Flags Simplified Multimin. Model is required since Well-2 does not have full WL logs, s s, l. WORKFLOW

© F

Ba

talle

r, P

etro

phy

sics

Spec

ialis

ts, R

epso

lE&

P, D

ecem

ber

20

19

.

TRAINNING 1 WELL12 ¼” section

(2838/6161Samples)

TRAINNING

Training dataset

Validatingdataset

Model for PhiT

Model for SwT

Model for Reservoir Flag

Model for Pay Flag

Regression

Regression

Clustering/Regression

Clustering/Regression

Clustering

Model for HC Flag

Well 1

Page 10: MACHINE LEARNING APPLIED TO EARLY FORMATION …...Advanced Gas Interpretation HC Flags Simplified Multimin. Model is required since Well-2 does not have full WL logs, s s, l. WORKFLOW

© F

Ba

talle

r, P

etro

phy

sics

Spec

ialis

ts, R

epso

lE&

P, D

ecem

ber

20

19

.

TRAINNING 1 WELL12 ¼” section

(2838/6161Samples)

TRAINNING

Training dataset

Validatingdataset

Trainning with 46% of the data

Model for PhiT

Model for SwT

Model for Reservoir Flag

Model for Pay Flag

Regression

Regression

Clustering/Regression

Clustering/Regression

Clustering

Model for HC Flag

Well 1

Page 11: MACHINE LEARNING APPLIED TO EARLY FORMATION …...Advanced Gas Interpretation HC Flags Simplified Multimin. Model is required since Well-2 does not have full WL logs, s s, l. WORKFLOW

© F

Ba

talle

r, P

etro

phy

sics

Spec

ialis

ts, R

epso

lE&

P, D

ecem

ber

20

19

.

VALIDATING WELL 112 ¼” section

(3323/6161Samples)

VALIDATING Validating with 54% of the data

Training dataset

Validatingdataset

Model for PhiT

Model for SwT

Model for Reservoir Flag

Model for Pay Flag

Regression

Regression

Model for HC Flag

Clustering/Regression

Clustering/Regression

Clustering

Well 1

Page 12: MACHINE LEARNING APPLIED TO EARLY FORMATION …...Advanced Gas Interpretation HC Flags Simplified Multimin. Model is required since Well-2 does not have full WL logs, s s, l. WORKFLOW

© F

Ba

talle

r, P

etro

phy

sics

Spec

ialis

ts, R

epso

lE&

P, D

ecem

ber

20

19

.

VALIDATING Well 1

Page 13: MACHINE LEARNING APPLIED TO EARLY FORMATION …...Advanced Gas Interpretation HC Flags Simplified Multimin. Model is required since Well-2 does not have full WL logs, s s, l. WORKFLOW

© F

Ba

talle

r, P

etro

phy

sics

Spec

ialis

ts, R

epso

lE&

P, D

ecem

ber

20

19

.

VALIDATING WELL 212 ¼” section

(5961Samples)

VALIDATING

Model for PhiT

Model for SwT

Model for Reservoir Flag

Model for Pay Flag

Regression

Regression

Clustering

Model for HC Flag

Validating with 100% of the data from Well 2

Clustering/Regression

Clustering/Regression

Train interval in Well 2 was not used fortraining, it was used for validating

Well 2

Page 14: MACHINE LEARNING APPLIED TO EARLY FORMATION …...Advanced Gas Interpretation HC Flags Simplified Multimin. Model is required since Well-2 does not have full WL logs, s s, l. WORKFLOW

© F

Ba

talle

r, P

etro

phy

sics

Spec

ialis

ts, R

epso

lE&

P, D

ecem

ber

20

19

.

VALIDATING Well 2

Page 15: MACHINE LEARNING APPLIED TO EARLY FORMATION …...Advanced Gas Interpretation HC Flags Simplified Multimin. Model is required since Well-2 does not have full WL logs, s s, l. WORKFLOW

© F

Ba

talle

r, P

etro

phy

sics

Spec

ialis

ts, R

epso

lE&

P, D

ecem

ber

20

19

.

DATAPOINTS

MM4 MM5 LM4LM2 LM3

Gro

un

dTr

uth

Pre

dic

ted

Porosity (PhiT)

Wel

l-1

Wel

l-2

CC = 0,7 27,3% 24,2% 22,7% 17,7% 17 % 17,8%

LM1

16%

LM5

15%

Eoc

Gro

un

dTr

uth

Pre

dic

ted

Oligo MM4 LM3LM1 LM2MM5 LM4

21,9%** 24,1%* 20,9%* 21,1%** 19,1%** 16,9% 16,1% 17,6%

*Used for Trainning** Partially used for Trainning

HOW GOOD WERE THE MODELS?

PER FORMATION (mean values)

24% 17,6% 17,2% 13,9% 13,1% 14,6%12,6%

17,1% 17,5% 17,9% 15,1% 14,6% 15,5% 14,4%

Raw results

Wel

l-1

Wel

l-2

CC = 0,73

Page 16: MACHINE LEARNING APPLIED TO EARLY FORMATION …...Advanced Gas Interpretation HC Flags Simplified Multimin. Model is required since Well-2 does not have full WL logs, s s, l. WORKFLOW

© F

Ba

talle

r, P

etro

phy

sics

Spec

ialis

ts, R

epso

lE&

P, D

ecem

ber

20

19

.

DATAPOINTS PER FORMATION (mean values)

MM4 MM5 LM4LM2 LM3

Gro

un

dTr

uth

Pre

dic

ted

Water Saturation (SwT)

Wel

l-1

Wel

l-2

86% 80,8% 92,7% 93,3% 91,7 % 84,3%

LM1

93,8%

LM5

99%

Eoc

Gro

un

dTr

uth

Pre

dic

ted

Oligo MM4 LM3LM1 LM2MM5 LM4

94,7% 95,5% 92,9% 96,8% 97,8% 99,5%97,6%

95,3%** 80,9%* 92,7%* 96,6%** 91,2%** 88,1% 90,8% 94,3%

95,1% 93,5% 91,6% 95,7% 95,5% 93,4% 98,3%

*Used for Trainning** Partially used for Trainning

HOW GOOD WERE THE MODELS?

Raw results

Wel

l-1

Wel

l-2

CC = 0,72

CC = 0,85

Page 17: MACHINE LEARNING APPLIED TO EARLY FORMATION …...Advanced Gas Interpretation HC Flags Simplified Multimin. Model is required since Well-2 does not have full WL logs, s s, l. WORKFLOW

© F

Ba

talle

r, P

etro

phy

sics

Spec

ialis

ts, R

epso

lE&

P, D

ecem

ber

20

19

.

DATAPOINTS Well-2

Reservoir, Pay and HC Flags

Well-1

Only considering validation intervals

MDT Samples

HOW GOOD WERE THE MODELS?

Page 18: MACHINE LEARNING APPLIED TO EARLY FORMATION …...Advanced Gas Interpretation HC Flags Simplified Multimin. Model is required since Well-2 does not have full WL logs, s s, l. WORKFLOW

© F

Ba

talle

r, P

etro

phy

sics

Spec

ialis

ts, R

epso

lE&

P, D

ecem

ber

20

19

.

DATAPOINTS Well-2

Reservoir, Pay and HC Flags

Well-1

Only considering validation intervals

MDT Samples

Although Gas from condansate was not possible to distinguish (maybe because not enough trainning data)

HOW GOOD WERE THE MODELS?

Page 19: MACHINE LEARNING APPLIED TO EARLY FORMATION …...Advanced Gas Interpretation HC Flags Simplified Multimin. Model is required since Well-2 does not have full WL logs, s s, l. WORKFLOW

© F

Ba

talle

r, P

etro

phy

sics

Spec

ialis

ts, R

epso

lE&

P, D

ecem

ber

20

19

.

Well 3Will it work?

Page 20: MACHINE LEARNING APPLIED TO EARLY FORMATION …...Advanced Gas Interpretation HC Flags Simplified Multimin. Model is required since Well-2 does not have full WL logs, s s, l. WORKFLOW

© F

Ba

talle

r, P

etro

phy

sics

Spec

ialis

ts, R

epso

lE&

P, D

ecem

ber

20

19

.

Well 3

PredictedReal

Page 21: MACHINE LEARNING APPLIED TO EARLY FORMATION …...Advanced Gas Interpretation HC Flags Simplified Multimin. Model is required since Well-2 does not have full WL logs, s s, l. WORKFLOW

© F

Ba

talle

r, P

etro

phy

sics

Spec

ialis

ts, R

epso

lE&

P, D

ecem

ber

20

19

.

Well 3

PhiT Predicted vs PhiT ground truth (smoothed!!)

Limitation in highporosity sands

PhiT Real

Ph

iTP

red

icte

d

Page 22: MACHINE LEARNING APPLIED TO EARLY FORMATION …...Advanced Gas Interpretation HC Flags Simplified Multimin. Model is required since Well-2 does not have full WL logs, s s, l. WORKFLOW

© F

Ba

talle

r, P

etro

phy

sics

Spec

ialis

ts, R

epso

lE&

P, D

ecem

ber

20

19

.

Results per wellWell-1 Well-2 Well-3

Ph

iTSw

TP

red

icte

d

Real

Pre

dic

ted

Real

Pre

dic

ted

Real

Pre

dic

ted

RealP

red

icte

dReal

Page 23: MACHINE LEARNING APPLIED TO EARLY FORMATION …...Advanced Gas Interpretation HC Flags Simplified Multimin. Model is required since Well-2 does not have full WL logs, s s, l. WORKFLOW

© F

Ba

talle

r, P

etro

phy

sics

Spec

ialis

ts, R

epso

lE&

P, D

ecem

ber

20

19

.

PhiT (All Wells)

PhiT Real - Phi Predicted =

PhiT predicted vs PhiT ground truth

PhiT Real Smoothed

Ph

iTP

red

icte

dSm

oo

thed

Is this enough??

Page 24: MACHINE LEARNING APPLIED TO EARLY FORMATION …...Advanced Gas Interpretation HC Flags Simplified Multimin. Model is required since Well-2 does not have full WL logs, s s, l. WORKFLOW

© F

Ba

talle

r, P

etro

phy

sics

Spec

ialis

ts, R

epso

lE&

P, D

ecem

ber

20

19

.

Fluid Type Flags ResultsWell-1 Well-2 Well-3

Page 25: MACHINE LEARNING APPLIED TO EARLY FORMATION …...Advanced Gas Interpretation HC Flags Simplified Multimin. Model is required since Well-2 does not have full WL logs, s s, l. WORKFLOW

© F

Ba

talle

r, P

etro

phy

sics

Spec

ialis

ts, R

epso

lE&

P, D

ecem

ber

20

19

.

CONCLUSIONS

- Although this methodology will never replace a conventional logging suit and petrophysicalevaluation, results are promissing. (+70% accuracy, +85% considering average and smoothed values). Predictive value is clear but isit enough? Can we improve it?

- Limitations and potential improvement areasidentified- Higher sample rate? Bigger data base? FLAIR

and Resistivity?

Page 26: MACHINE LEARNING APPLIED TO EARLY FORMATION …...Advanced Gas Interpretation HC Flags Simplified Multimin. Model is required since Well-2 does not have full WL logs, s s, l. WORKFLOW

© F

Ba

talle

r, P

etro

phy

sics

Spec

ialis

ts, R

epso

lE&

P, D

ecem

ber

20

19

.

CONCLUSIONS

- Models using imbedded algorithms in Geolog havebeen created and validated with acceptableenough results for a preliminary quickinterpretation and fluid characterization

- For now, these models are considered “very local”, but if a big enough DB is built….

- If this can be done with Geolog, which is not anspecific Data Analytics software, it is very likelythat it can be improved using adequate software and data scientists with petrophysical background

Page 27: MACHINE LEARNING APPLIED TO EARLY FORMATION …...Advanced Gas Interpretation HC Flags Simplified Multimin. Model is required since Well-2 does not have full WL logs, s s, l. WORKFLOW

© F

Ba

talle

r, P

etro

phy

sics

Spec

ialis

ts, R

epso

lE&

P, D

ecem

ber

20

19

.

POTENTIAL APPLICATIONSGeological Operations- Logging program and tool configuration optimization- Fluid Sampling optimization- Pore Pressure prediction- Difficult logging conditions- Coring point decisión

Exploration and Appraisal- NNVV (How much info can you get from a Masterlog or Mudlog?)- Offset well analysis- Logging optimization in appraisal phase- Improvement on fluid hetereogeneity distribution with reduced

sampling (?)

Development- Additional information from wells not logged. Higher density of data for

static modeling- Identifying anomalous wells for potential (contingent) logging- Improvement of seismic resolution (Quantico Energy Solutions)

Page 28: MACHINE LEARNING APPLIED TO EARLY FORMATION …...Advanced Gas Interpretation HC Flags Simplified Multimin. Model is required since Well-2 does not have full WL logs, s s, l. WORKFLOW

© F

Ba

talle

r, P

etro

phy

sics

Spec

ialis

ts, R

epso

lE&

P, D

ecem

ber

20

19

.

ACKNOWLEDGMENTS

- Ridvan Akkurt from Schlumberger for his wisdom in the first stage of this Project

- Matheu Llurba and Eusebio Oña from Repsol Geological Operations Team.

- Roberto Varade (exRepsol), Ricardo Ferreti (exSLB), Giulio Beda (exRepsol) for their creativity and “willingto do things” attitude when starting this Project.

- To Repsol and its Partners (Total and OMV) for theirsupport and permission to use this dataset