From Monte Carlo to Bayes Theory: The Role of Uncertainty in Petrophysics.

Transcription

1 From Monte Carlo to Bayes Theory: The Role of Uncertainty in Petrophysics. Simon Stromberg Global Technical Head of Petrophysics Senergy

2 Subsurface Characterisation The main goal of the oil-industry geoscience professional is to characterise the complete range of possible configurations of the subsurface for a given set of data and related analogues. This characterisation of the subsurface should lead to a comprehensive description of uncertainty that leads to the complete disclosure of the financial risk of further exploration, appraisal or development of a potential hydrocarbon resource.

3 Subsurface Realisation Tables

4 Petrophysical workflows Calculate Volume of Clay Volume of Clay Calculate Clay Corrected Porosity Effective and Total Porosity Calculate Clay Corrected Saturation Effective and Total Water Saturation Apply cut-offs for net sand, reservoir and pay and averages Average Vcl, Por, Phi And HPVOL

5 Petrophysical Base Case Interpretation

6 Petrophysical Base Case Interpretation

7 Methods of Uncertainty Analysis Parameter sensitivity analysis Partial derivative analysis Monte Carlo Simulation Bayesian Analysis for Diagnostic Reliability

8 Sensitivity Analysis of Input Parameters Single Parameter For example: Volume of clay cut-off for net reservoir If VCL < 0.3 and If PHIE > 0.1 then The interval is flagged as net reservoir

9 Sensitivity Analysis of Input Parameters Single Parameter

10 Sensitivity Analysis of Input Parameters Single Parameters

11 Partial Derivative Analysis In mathematics, a partial derivative of a function of several variables is its derivative with respect to one of those variables, with the others held constant (as opposed to the total derivative, in which all variables are allowed to vary). Partial derivatives are used in vector calculus and differential geometry. The partial derivative of a function f with respect to the variable x is variously denoted by The partial-derivative symbol is. The notation was introduced by Adrien-Marie Legendre and gained general acceptance after its reintroduction by Carl Gustav Jacob Jacobi.

12 Partial Derivative Analysis Example Volume of a cone is: The partial derivative of the volume with respect to the radius is Which describes the rate at which the volume changes with change in radius if the height is kept constant.

13 Partial Derivative Analysis for Waxman- Smits Equation for Saturation ( ) ( ) ( ) ( ) ( ) Rt A A BQv Rt n m n F Qv B n E E F n E F m A E F A E F m E Rt E Rt n n m m A A Rt Rt m ; 1 ln ; ln ; ; ; 1 φ φ φ φ φ δ δ δ δφ φ δ δ δ

14 Partial Derivative Analysis for A Complete Deterministic Petrophysical Analysis ( ) ( ) ( ) ( ) ( ) Rt A A Qv B Rt n m n F Qv B n E E F n E F m A E F A E F m E Rt E Rt n n m m A A Rt Rt m ; 1 ln ; ln ; ; ; 1 φ φ φ φ φ δ δ δ δφ φ δ δ δ ( ) ( ) ( ) ; 1 ; ; fl ma b fl ma b ma fl fl ma fl b ma b b fl fl ma ma fl ma b ma ρ ρ ρ φ ρ ρ ρ ρ ρ φ ρ ρ ρ ρ ρ φ δρ ρ φ δρ ρ φ δρ ρ φ δφ ρ ρ ρ ρ φ + + ( ) ( ) φ φ φ δ δφ φ δ δ φ g n Foil E F m g n Foil g n Foil Foil Foil g n g n Foil Foil g n Foil 1 1 / / / 1 geol sys stat total δ δ δ δ + + samples number of n std dev n stat ; 1 ; σ σ δ

15 Partial Derivative Input Data Zonal Averages Sumavs well zone gross net ng por 1 HSGHL HSGHL HSGHL HSGHL HSGHL HSGHL HSGHL

16 Partial Derivative Input Uncertainty Field: Zone: Porosity Hydrocarbon Saturation parameter value 1 std dev part error % error parameter value 1 std dev part error % error count 6 count 6 w average w average δstat % δstat % rhob % Rt % rhoma % % rhofl % por % A % m % n % B % Qv % δsys % δsys % δgeol % δgeol % average 0.58 δtotal % average δtotal % net/gross Foiln/g*por*Sh count 7 count 6 w average w average δstat % n/g % por % Sh % δgeol % δsys % average δtotal % average δtotal %

17 Partial Derivative Analysis Results Uncertainty (Normal) Distribution Curves porosity Hydrocarbon Saturation distribution porosity avg0.58, std0.010, std3.8% net/gross µ 3σ0.9, µ+3σ0.87 distribution Sh avg0.636, std0.093, std14.7% µ 3σ0.356, µ+3σ0.916 Foil distribution net/gross distribution Foil avg0.449, std0.131, std9.1% µ 3σ0.057, µ+3σ0.840 avg0.074, std0.04, std33.% µ 3σ0.000, µ+3σ0.147

18 Partial derivative Analysis - Saturation Saturation Uncertainty (d) vs. Saturation for various Porosity Classes 0.50 Uncertainty (d) por0.1 por0.165 por0.5 por0.875 por

19 Monte-Carlo Simulation Multiple repeated calculation of all deterministic equations All input parameters can be sampled from a distribution of expected range in parameters Co-dependency can be honoured All output can be analysed Relative contribution to error can Be analysed Requires a computer Calculate Volume of Clay Calculate Clay Corrected Porosity Calculate Clay Corrected Saturation Apply cut-offs for net sand, reservoir and pay and averages

20 Monte-Carlo Simulation Reference Case Uncertainty Definition Sensitivity MONTE CARLO TORNADO PLOT Test Well 1 Error Analysis for : PhiSoH Reservoir Vcl Cut Res/Pay m exponent RHOB a factor n exponent Gr Clean TNPH Rho Wet Clay Neu Wet Clay LLD Phi Cut Res/Pay Res Clay SGR ND Neu Clean Hc Den ND Neu Clay ND Den Clean ND Neu Clean1 ND Den Clean1 ND Den Clay Res Clay Gr Clay Res Clean Rxo Clay MSFL Rmf Rho mud filt SD Son Clay Cut Res/Pay SD Den Clay SP Clay Rho Dry Clay DTLN SD Den Clean SD Son Clean PhiT Clay Neu Clean SD Den Clean1 Neu Clay SP Clean SD Son Clean1 Rho GD PhiSoH Reservoir Zone : All Shift Low Initial Values Shift High % % % 5% R 0.005R % % % % % % 0% % 0.005R 0.005R 0% % PhiSoH Pay Zone : All Mean : 4.84 Std Dev : PhiH Res Zone : All Mean : 18.5 Std Dev : Phi Cut Res/Pay Mean : Std Dev : Results MONTE CARLO RESULTS HISTOGRAMS Test Well 1 Number of Simulation : PhiSoH Res Zone : All Mean : Std Dev : Av Phi Pay Zone : All Mean : 0.04 Std Dev : PhiH Pay Zone : All Mean : Std Dev : Av Phi Res Zone : All Mean : Std Dev :

21 Models and Equations

22 Results MONTE CARLO RESULTS HISTOGRAMS Test Well 1 Number of Simulation : 408 PhiSoH Pay Zone : All Mean : 4.86 Std Dev : HPVOL (ft)

23 Sensitivity Vcl Cut Res/Pay Cut Res/Pay m exponent RHOB n exponent a factor Gr Clean Rho Wet Clay Neu Wet Clay SGR Hc Den LLD TNPH Res Clay Phi Cut Res/Pay ND Neu Clean ND Neu Clay ND Den Clean ND Neu Clean1 Gr Clay Rmf ND Den Clean1 Res Clay ND Den Clay Rxo Clay MSFL Res Clean PhiT Clay Rho Dry Clay Rho GD Rho mud filt DTLN MONTE CARLO TORNADO PLOT Test Well 1 Error Analysis for : PhiSoH Pay PhiSoH Pay Zone : 3 Shift Low Initial Values Shift High % % R 0.005R 5% 5% 0% % % % % % % % 0.005R 0.005R 0% %

24 Bayesian Analysis For Reliability Bayes allows modified probabilities to be calculated based on 1. The expected rate of occurrence in nature. A diagnostic test that is less than 100% reliable For example There is a 1:10,000 (0.0001%) occurrence of a rare disease in the population There is a single test of the disease that is 99.99% accurate A patient is tested positive for that disease What is the likelihood that the patient tested positive actually has the disease?

25 Bayesian Theory Bayes theory is the statistical method to revise probability based on a assessment from new information. This is Bayesian analysis. To set up the problem: Consider mutually collective and collectively exhaustive outcome (E1, E.En) A is the outcome of an information event, or a symptom related to E. If A is perfect information, Bayes theorem is NOT needed. ) ( ) ( ) ( ) ( ) ( ) ( ) ( 1 A P E A P E P AE P E P AE P A E P i N j j j i i i

26 Bayes Theory INPUTS Input Number Calculation Probability of (state of nature) occurring P(A) Input 0.10 Probablity of state of nature not occuring P(nA) 1-P(A) 0.90 Probability of True Positive Test (that B will be detected if A exists) P(B A) Input 0.90 Probability of False Positive Test (That B will be detected if A does not exist) P(B na) Input 0.0 False Negative Test P(nB A) 1-P(B A) 0.10 Probability of a true negative test. P(nB na) 1-P(B na) 0.80 Total probability of detecting A (whether it present or not) P(B) P(B na)*p(na)+(p(b A)*P(A) 0.7 Total probability of not detecting A (whether it present or not) P(nB) 1-P(B) 0.73 Probability that A is present given that it was detected P(A B) P(B A)*P(A)/P(B) 0.33 OUTPUTS Probability that A is present given that it was not detected (probability of a a false negative) P(A nb) P(nB A)*P(A)/P(nB) 0.01 Probability that A is NOT present given it was detected P(nA B) 1-P(A B) 0.67 Probability that A is NOT present given it was NOT detected P(nA nb) 1-P(A nb) 0.99

27 Using AVO to De-risk Exploration and The Impact of Diagnosis Reliability AVO Anomaly Information Our geophysicist has evaluated AVO anomalies and has assessed that: There is a 10% chance of geological success There is a 90% chance of seeing an AVO if there is a discovery There is a 0% chance of seeing a false anomaly if there is no discovery Question If we have a 10% COS based on the geological interpretation and an anomaly is observed, what is the revised COS if an AVO is observed What is the added value of AVO

28 AVO 1 Discovery 10% 90% 1.1 Anomaly 9% 10% 1. No Anomaly 1% Geologocal COS No Discovery 90% 0%.1 Anomaly 18% 80%. No Anomaly 7%

29 AVO-Bayes Theory INPUTS Input Number Calculation Probability of (state of nature) occurring P(A) Input 0.10 Probablity of state of nature not occuring P(nA) 1-P(A) 0.90 Probability of True Positive Test (that B will be detected if A exists) P(B A) Input 0.90 Probability of False Positive Test (That B will be detected if A does not exist) P(B na) Input 0.0 False Negative Test P(nB A) 1-P(B A) 0.10 Probability of a true negative test. P(nB na) 1-P(B na) 0.80 Total probability of detecting A (whether it present or not) P(B) P(B na)*p(na)+(p(b A)*P(A) 0.7 Total probability of not detecting A (whether it present or not) P(nB) 1-P(B) 0.73 Probability that A is present given that it was detected P(A B) P(B A)*P(A)/P(B) 0.33 OUTPUTS Probability that A is present given that it was not detected (probability of a a false negative) P(A nb) P(nB A)*P(A)/P(nB) 0.01 Probability that A is NOT present given it was detected P(nA B) 1-P(A B) 0.67 Probability that A is NOT present given it was NOT detected P(nA nb) 1-P(A nb) 0.99

30 AVO AVO Anomaly Question If we have a 10% COS based on the geological interpretation and an anomaly is observed, what is the revised COS if an AVO is observed What is the added value of AVO Answer based on Bayes is that if there is an AVO anomaly there is a 33% chance it is a discovery. Also we can calculate, if there in NO AVO anomaly then the chance it is a discovery is 1%/

31 Assessing Critical Porosity in Shallow Hole Sections Synopsis of SPE paper in press New technique for deriving porosity from sonic logs in oversize boreholes Sonic porosity used to determine of porosity is at critical porosity Critical porosity 4 to 45 p.u. If lower than critical porosity rock is load bearing Near or at critical porosity the rocks may have insufficient strength to contain the forces of shutting in the well Sonic porosity (using the new technique) is used to determine if rock is below critical porosity and used as part of the justification for NOT running a conductor

32 Case Study: Assessing Critical Porosity in Shallow Hole Sections LWD acoustic compression slowness below the drive pipe LWD sonic device optimised for large bore holes Data showed conclusive evidence of absence of hazards and were immediately accepted as waiver for running the diverter and conductor string of casing DTCO used to asses if rock consolidated (below critical porosity) Gives the resultant DTCO and porosity interpretation Raymer Hunt Gardner model with shale correction No mention of the reliability of the interpretation Processed DTCO accuracy Interpretation model uncertainty

33 Assessing Critical Porosity in Shallow Hole Sections Question What is the potential impact of tool accuracy and model uncertainty on the interpretation Should this be considered as part of the risk management for making the decision on running the conductor What is the reliability of the method and measurement for preventing the unnecessary running of diverter and conductor string of casing

34 Baseline Interpretation

35 Baseline Interpretation DT matrix 55.5 usec/ft DT Fluid 189 usec/ft DT wet clay 160 usec/ft (uncompacted sediment) VCL from GR using linear method Scale : 1 : 500 DB : Test (-1) 1 GR (GAPI) DEPTH (FT) BigSonic DEPTH (999.89FT FT) 09/01/010 15:16 3 DTCO (usec/ft) PHIT (Dec) GR (GAPI) DEPTH (FT) 3 DTCO (usec/ft) PHIT (Dec)

36 DB : Test (1) Rock above CP Scale : 1 : GR (GAPI) DEPTH (FT) 1000 BigSonic DEPTH (999.89FT FT) 10/09/010 10: DTCO (usec/ft) PHIT (Dec) ResFlag () Phi Cut Res Cutoff SensitivityData Wels: BigSonic P90 P50 P Net Reservoir Phi Cut Res Cutoff Net Reservoir - Al Zones 1 GR (GAPI) DEPTH (FT) 3 4 DTCO (usec/ft) PHIT (Dec) ResFlag ()

37 Uncertainty Analysis (Monte Carlo) Distribution Default + - DTCO (usec/ft) Gaussian log 5 5 GR clean Gaussian GR Clay Gaussian DT wet clay (usec/ft) Gaussian DT Matrix (usec/ft) Gaussian DT Water (usec/ft) Gaussian Cutoff for critical Porosity (v/v) Square

38 Uncertainty Analysis (Monte Carlo) Footage where porosity exceeds critical porosity Gross Section P10 0ft P50 3 ft P ft MONTE CARLO RESULTS HISTOGRAMS BigSonic Number of Simulation : 000 Net Res Zone : All Mean : 9.01 Std Dev : 5.93 MONTE CARLO TORNADO PLOT BigSonic Error Analysis for : Net Reservoir Shift Low Initial Values Shift High 00 Gr Clean Phi Cut Res/Pay Gr Clay Sonic water Sonic matrix DTCO 5 5 Sonic Wet Clay Net Reservoir Zone :

39 Results of Monte-Carlo Probability Net Rock > CP

40 Tornado Plot MONTE CARLO TORNADO PLOT BigSonic Error Analysis for : Net Reservoir Shift Low Initial Values Shift High Gr Clean Phi Cut Res/Pay Gr Clay Sonic water Sonic matrix DTCO 5 5 Sonic Wet Clay Net Reservoir Zone : 1

41 Conclusions from Monte Carlo The model processed DTCO and model uncertainty leads to significant doubt that the rock is below critical porosity The most important considerations are: The clay volume and clay correction The actual porosity value for critical porosity (0.41 to 0.45) Tool Accuracy is not a concern (+/- 5 usec/ft) A good question to ask is what is the tool accuracy given the new processing technique and challenges of running sonic in big bore-holes?

42 Bayesian Analysis of Reliability Lets assume that the regional data shows that the 90% of all top hole sections will be below CP and stable Based on Monte-Carlo it is judged that the interpretation will be 80% reliable 1 Below CP 90% Sonic Log Shows Below C P Sonic Log Shows AboveC P 80% % % % 0.00 Well Bore Stability 0.00 Above CP 10% Sonic Log Shows Above CP Sonic Log Shows Below C P 80% % % 0.00 % 0.00

43 Bayesian Inversion of Tree 1 Log Shows "BCP" 74% 97% 1.1 BC P (0.7/0.74) 7% 3% 1. AC P (0.0/0.74) % C ritical Porosity LOG Shows "ACP" 6% 69%.1 BC P (0.18/0.6) 18% 31%. AC P (0.08/0.6) 8% % Wells below CP regionally 0.9 Reliability of the log 0.8 "BCP" "ACP" BCP ACP If Log shows BCP 0.97 If Log Shows ACP 0.31

44 Bayesian Analysis of Reliability Joint Probability Table If the log shows that the section is below critical porosity then we can be 97% certain that it is below critical porosity % Wells below CP regionally 0.9 Reliability of the log 0.8 "BCP" "ACP" BCP ACP If Log shows BCP 0.97 If Log Shows ACP 0.31

45 Bayesian Analysis What if only 60% of the wells in the area are below critical porosity. What is the value of the sonic? % Wells below CP regionally 0.6 Reliability of the log 0.8 "BCP" "ACP" BCP ACP If Log shows BCP 0.86 If Log Shows ACP 0.73 If the sonic log shows below CP there is an 86% chance it is really below CP

46 Reliability of Diagnosis Chart 1.0 Probability that Interval is Below CP Log Diagnostic Reliability Regional Data % of well above Below Cp

47 Conclusions From Reliability Analysis If the Sonic is 100% reliable as a diagnosis of the well being above or below CP then the log data then we can be sure the interval is above/below CP If the Sonic log is not 100% reliable then we need to take into account Regional data trends Reliability of the Sonic log If we believe that the sonic log is less than 100% diagnostic then there is always a risk that the well will be above CP, even if the log data shows otherwise.

48 Conclusions The most important task of the geoscientist is to make statements about Range of possible subsurface outcomes based on: Uncertainty Diagnostic reliability of data There are several ways to analyse the range of outcomes based on uncertain input parameters Single parameter sensitivity Partial derivative analysis Monte-Carlo simulation Bayes analysis for diagnostic reliability The results of uncertainty and reliability analysis can be counter intuitive