Model Inversion for Induced Seismicity

Similar documents
Reservoir Geomechanics and Faults

Simplified In-Situ Stress Properties in Fractured Reservoir Models. Tim Wynn AGR-TRACS

4D stress sensitivity of dry rock frame moduli: constraints from geomechanical integration

An Open Air Museum. Success breeds Success. Depth Imaging; Microseismics; Dip analysis. The King of Giant Fields WESTERN NEWFOUNDLAND:

J.V. Herwanger* (Ikon Science), A. Bottrill (Ikon Science) & P. Popov (Ikon Science)

Microseismic Reservoir Monitoring

Statistical Rock Physics

3D Finite Element Modeling of fault-slip triggering caused by porepressure

Reservoir characterization in an underground gas storage field using joint inversion of flow and geodetic data

Internal microseismic event pattern revealed by waveform cross-correlation analysis

Production-induced stress change in and above a reservoir pierced by two salt domes: A geomechanical model and its applications

Lab 5 - Aquifer Elasticity and Specific Storage. Due October 12 14, 2010

Modeling pressure response into a fractured zone of Precambrian basement to understand deep induced-earthquake hypocenters from shallow injection

Storage: Deep Monitoring and Verification

Tensor character of pore pressure/stress coupling in reservoir depletion and injection

Geomechanics for reservoir and beyond Examples of faults impact on fluid migration. Laurent Langhi Team Leader August 2014

GAUSSIAN PROCESS REGRESSION

Current challenges at CO 2 Sites

Jihoon Kim, George J. Moridis, John Edmiston, Evan S. Um, Ernest Majer. Earth Sciences Division, Lawrence Berkeley National Laboratory 24 Mar.

Coupled geomechanics and InSAR inversion of CO2 injection parameters

Fully Coupled Geomechanics, Multi-Phase, Thermal and. Equation of State Compositional Simulator

Reservoir Modeling for Wabamun Area CO2 Sequestration Project (WASP) Davood Nowroozi

Mechanics of Earthquakes and Faulting

Gaussian Processes. Le Song. Machine Learning II: Advanced Topics CSE 8803ML, Spring 2012

Integrating Lab and Numerical Experiments to Investigate Fractured Rock

WP 4.1. Site selection criteria and ranking methodology. Karen Kirk

Know Before You Go. Steve O Connor. Bryony Youngs. How GeoPrediction helps with Production Optimization and Assessing the Size of the Prize

5 IEAGHG CCS Summer School. Geological storage of carbon dioxide (a simple solution)

Risk Assessment for CO 2 Sequestration- Uncertainty Quantification based on Surrogate Models of Detailed Simulations

Modeling Fault Reactivation, Induced Seismicity, and Leakage during Underground CO2 Injection

Training Venue and Dates Ref # Reservoir Geophysics October, 2019 $ 6,500 London

Dynamic System Identification using HDMR-Bayesian Technique

A review of friction laws and their application for simulation of microseismicity prior to hydraulic fracturing

The Frictional Regime

FRACTURE REORIENTATION IN HORIZONTAL WELL WITH MULTISTAGE HYDRAULIC FRACTURING

Reliability Monitoring Using Log Gaussian Process Regression

Mathematical Modelling of a Fault Slip Induced by Water Injection

CO 2 storage capacity and injectivity analysis through the integrated reservoir modelling

Assessing the Value of Information from Inverse Modelling for Optimising Long-Term Oil Reservoir Performance

Numerical Study of Relationship Between Landslide Geometry and Run-out Distance of Landslide Mass

Synthetic Seismicity Models of Multiple Interacting Faults

Finite element modelling of fault stress triggering due to hydraulic fracturing

Use of Seismic and EM Data for Exploration, Appraisal and Reservoir Characterization

URTeC: Abstract

Implementing and Monitoring a Geomechanical Model for Wellbore Stability in an Area of High Geological Complexity

Nonparametric Bayesian Methods (Gaussian Processes)

Brittle Deformation. Earth Structure (2 nd Edition), 2004 W.W. Norton & Co, New York Slide show by Ben van der Pluijm

Integration of Geophysical and Geomechanical

Lecture: Gaussian Process Regression. STAT 6474 Instructor: Hongxiao Zhu

Workflows for Sweet Spots Identification in Shale Plays Using Seismic Inversion and Well Logs

Developments on Microseismic Monitoring and Risk Assessment of Large-scale CO 2 Storage

Practical Geomechanics

A PRESSURE VESSEL FOR TRUE-TRIAXIAL DEFORMATION & FLUID FLOW DURING FRICTIONAL SHEAR

Gain information on the development of Engineered Geothermal Systems (EGS) and enhance our understanding of long-term reservoir behaviour.

Apply Rock Mechanics in Reservoir Characterization Msc Julio W. Poquioma

Exploration _Advanced geophysical methods. Research Challenges. Séverine Pannetier-Lescoffit and Ute Mann. SINTEF Petroleum Research

Interpreting lab creep experiments: from single contacts... to interseismic fault models

Dynamic source inversion for physical parameters controlling the 2017 Lesvos earthquake

The Interaction of Reservoir Engineering and Geomechanics (a story)

Uncertainties in rock pore compressibility and effects on time lapse seismic modeling An application to Norne field

Characterizing Seal Bypass Systems at the Rock Springs Uplift, Southwest Wyoming, Using Seismic Attribute Analysis*

Regional-Scale Salt Tectonics Modelling: Bench-Scale Validation and Extension to Field-Scale Predictions

Defmod, an earthquake simulator that adaptively switches between quasi-static and dynamic states

Opportunities in Oil and Gas Fields Questions TABLE OF CONTENTS

Yusuke Mukuhira. Integration of Induced Seismicity and Geomechanics For Better Understanding of Reservoir Physics

Stochastic Spectral Approaches to Bayesian Inference

ON NEAR-FIELD GROUND MOTIONS OF NORMAL AND REVERSE FAULTS FROM VIEWPOINT OF DYNAMIC RUPTURE MODEL

An Investigation on the Effects of Different Stress Regimes on the Magnitude Distribution of Induced Seismic Events

STA414/2104 Statistical Methods for Machine Learning II

Induced seismicity and GeoEnergies: lessons learned from coupled hydro-mechanical modeling

Neutron inverse kinetics via Gaussian Processes

Modeling seismic wave propagation during fluid injection in a fractured network: Effects of pore fluid pressure on time-lapse seismic signatures

Petrophysical Data Acquisition Basics. Coring Operations Basics

Material is perfectly elastic until it undergoes brittle fracture when applied stress reaches σ f

SENSITIVITY ANALYSIS IN NUMERICAL SIMULATION OF MULTIPHASE FLOW FOR CO 2 STORAGE IN SALINE AQUIFERS USING THE PROBABILISTIC COLLOCATION APPROACH

Storage 4 - Modeling for CO 2 Storage. Professor John Kaldi Chief Scientist, CO2CRC Australian School of Petroleum, University of Adelaide, Australia

Fully Bayesian Deep Gaussian Processes for Uncertainty Quantification

MUDLOGGING, CORING, AND CASED HOLE LOGGING BASICS COPYRIGHT. Coring Operations Basics. By the end of this lesson, you will be able to:

Journal of Geophysical Research Letters Supporting Information for

Nonparameteric Regression:

Net-to-gross from Seismic P and S Impedances: Estimation and Uncertainty Analysis using Bayesian Statistics

Checking up on the neighbors: Quantifying uncertainty in relative event location

Recommendations for Injection and Storage Monitoring

Research in Induced Seismicity

Source parameters II. Stress drop determination Energy balance Seismic energy and seismic efficiency The heat flow paradox Apparent stress drop

Multivariate Bayesian Linear Regression MLAI Lecture 11

STA414/2104. Lecture 11: Gaussian Processes. Department of Statistics

ADVANCED MACHINE LEARNING ADVANCED MACHINE LEARNING. Non-linear regression techniques Part - II

Microseismic Geomechanical Modelling of Asymmetric Upper Montney Hydraulic Fractures

Modeling of Fault Reactivation and Induced Seismicity During Hydraulic Fracturing of Shale-Gas Reservoirs

BRIEFING MEMO ON RESERVOIR TRIGGERED SEISMICITY (RTS)

Abstracts ESG Solutions

Probabilistic numerics for deep learning

(Refer Slide Time: 01:15)

Estimating fault slip rates, locking distribution, elastic/viscous properites of lithosphere/asthenosphere. Kaj M. Johnson Indiana University

Rock and fluid thermodynamics control the dynamics of induced earthquakes

TexNet and CISR: An Update on Monitoring and Understanding Seismicity in Texas

New Insights into History Matching via Sequential Monte Carlo

Results of the Lacq pilot s monitoring Focus on microseismicity. Joëlle HY-BILLIOT 10 may - Venice

One or two deficiencies of current fault seal analysis methods

Transcription:

Model Inversion for Induced Seismicity David Castiñeira Research Associate Department of Civil and Environmental Engineering In collaboration with Ruben Juanes (MIT) and Birendra Jha (USC) May 30th, 2017

Outline Slide 2 1. Introduction CCS Induced Seismicity 2. Methodology Forward Modeling Inverse Modeling 3. Results for CO 2 sequestration Model Model description Model uncertainties Bayesian solution for inverse problem 4. Conclusions

CO 2 sequestration and Induced Seismicity Slide 3 Carbon Capture and Storage (CCS) challenges: Economic (and political!) Engineering Risks and uncertainties Risks of Induced Seismicity: Uncertainty is always present (e.g., fault stability) Evaluating impact of CO 2 injection on fault stability is more difficult. Need to quantify the risk of induced seismicity due to CO 2 injection. Source: University of Edinburgh Proposed solution to evaluate seismic risk: To develop (forward) model that simulate the fundamental physics. To implement a (inverse) problem workflow to match observations To deploy model for designing operational constraints. Source: BGR

Forward Modeling: Coupled flow-geo simulation Slide 4 Existing computational solutions for coupling flow and deformation. Numerical simulation framework to couple a multiphase flow simulator with a mechanics simulator. Unconditionally stable fixed-stress scheme for sequential solution of two-way coupling. Rigorous formulation of nonlinear multiphase geomechanics to handle strong capillary effects. Mechanical stability of faults modeled from fluid pressures and multiphase flow behavior. Faults represented as surfaces embedded in a 3D medium by zero-thickness interf. elements (to accurately model fault slip under dynamically evolving fluid pressure and fault strength). Reference: Jha, B., and R. Juanes (2014), Coupled multiphase flow and poromechanics: A computational model of pore pressure effects on fault slip and earthquake triggering, Water Resour. Res., 50, doi:10.1002/2013wr015175. Applications: Algorithms can be used for prediction fault slip, fault activation, subsidence, identification of unswept zones in both real and synthetic reservoirs, etc. Research: Improving forward-model Reservoir Simulation Analytical models Data- driven models

Inverse Modeling: Uncertain parameters Slide 5 Large-scale parameters Tectonic boundary conditions Size of the reservoir. Aquifer description: Size of the aquifer and aquifer/reservoir permeability ratio. Fault description: strike, dip, permeability PVT parameters Rock parameters Permeability, Porosity, NTG... Geomechanical properties Poisson ratio (u ), Static coefficient of fraction, dynamic friction coefficient, compressibility intercept (a), compressible wave velocity (vp), shear wave velocity (vs), etc.

Inverse Modeling: Observations Slide 6 Driving index Samaria (P50) 1 Depletion Drive Index (DDI) Connate water and rock expansion Index (CRI) Water Drive Index (WDI) Gas Injection Drive Index (GIDI) 0.9 0.8 Reservoir flow data Well static bottom hole pressure (WBHP data) Well flow rates (production/injection) Mechanical data Surface deformation (measured from InSAR, GPS, etc) Reservoir compaction (measured using radioactive markers). Seismic data Seismic data reflects earthquake dynamic; it should brings valuable information to the inverse problem. Problem with seismic data: it s massive, noisy, difficult to interpret without specialized knowledge. We need an effective approach to incorporate seismic data into inverse problem workflow: earthquake attributes Driving Index 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0 1974 1979 1984 1989 1994 1999 2004 Date SL No Stand Pipe Parasite String TD Ecd Ql Qg Max Cuttings 1 2 3 4 5 6 7 8 13 13 115 115 10 10 10 9 135 12 12 10 10 8 7 8 ECD 5.ppg 4.75 4.81 4.54 4.66 4.35 4.55 4.73 4.85 5.16 5.24 4.97 5.13 4.82 5.09 5.3 4.3 1.0918% 1.1841% 1.2816% 1.4624% 1.6120% 1.9141% 2.1340% 2.0640% 1 2 3 4 5 6 8 8 8 75 7 7 125 11 10 10 9 8 ECD 4.ppg 3.56 3.66 3.75 3.58 3.52 3.67 3.96 4.1 4.23 4.05 4.03 4.25 1.3000% 1.5000% 1.7011% 1.7207% 1.9871% 2.3089% 1 2 3 4 5 55 55 55 55 5 125 10 95 9 9 ECD 3.ppg 2.7 2.86 2.89 2.93 2.71 3.06 3.32 3.37 3.44 3.22 1.2880% 1.8690% 1.9893% 2.1220% 2.1976% Cas Shoe ECD 2009

Inverse Modeling: Bayesian Solution Slide 7 Observations (D) Model parameters (q ) q D) = Posterior Distribution Bayes Rule: q D) = D q ) q ) D) = ò D q ) q ) D q ) q ) dq D q ) q ) D) = Likelihood function = Prior distribution = Evidence (or marginal distribution) Why Bayesian? Challenges: Inference from noise and limited data. Complex structure of the posterior distributions Natural mechanism for regularization Expensive forward model and large dimensionality of Natural framework for uncertainty quantification and the inversion parameters VoI analysis.

Inverse Modeling: Gaussian Processes Slide 8 Gaussian Processes (GP) represents a powerful regression technique for I/O data. GP finds joint distribution of infinite-dimensional multivariate normally distributed random variables. GP is actually a distribution over functions! One needs to define Kernel matrix (i.e., covariance of multivariate normal distribution, e.g., square exponential), which controls the level of smoothness in your non-linear regression model. (It) can be used for multiple input/output data Multivariate Normal Distribution: æ N ç 1 d q ik - q K( q i, q j; a ) = s exp - ç å 2 è 2 k= 1 lk 2 2 jk ö ø

Recap: Modeling Induced Seismicity Slide 9 Source: BGR PHYSICAL MODEL Forward Model Computational Model Predictions Data Science Observations + + - Inverse Problem 1. Updated geological model 2. Risk Analysis and Operational Constraints

Case Study: 3D MODEL FOR CO 2 SEQUESTRATION 2016 Annual Founding Members Meeting 10

CO 2 injection (2D model) Slide 11 CO 2 injection in a deep confined aquifer for the purpose of geologic carbon sequestration. CO 2 is injected at 1500 m Aquifer is bounded on the top and bottom by a low-permeability caprock, and the fault is (in principle) impermeable to flow. Aquifer is hydraulically compartmentalized with a sealing fault that cuts across it. 2D plane-strain model with fault under normal faulting conditions (vertical principal stress due to gravity is the largest among the three principal stresses). Storage capacity of the aquifer is limited by overpressurization and possible slip on the fault. -> Modeling needed (!)

CO 2 injection (3D model) Slide 12 CO 2 injection in a deep confined aquifer for the purpose of geologic carbon sequestration. 3D model with (4 x 4 x 2) km simulation domain. CO 2 is injected at 1500 m in 200m anticlinal aquifer. t Aquifer is hydraulically compartmentalized with a sealing fault that cuts across it. Aquifer is bounded on the top and bottom by a low-permeability caprock. failure stable s Normal faulting conditions (vertical principal stress due to gravity is the largest among the three principal stresses). Storage capacity of the aquifer is limited by overpressurization and possible slip on the fault. -> Modeling needed (!)

CO 2, Uncertain Model Parameters Slide 13 Static coefficient of friction (µ f ) : It s assumed that sliding begins when the ratio of shear to normal stress on contact surfaces reaches the value of static coefficient. Dynamic coefficient of friction(µ d ): after sliding occurs the coefficient of friction decreases and reaches the dynamic friction coefficient. Slip-weakening model (d 0 ): The linear slip-weakening friction model produces shear tractions equal to the cohesive stress plus a contribution proportional to the fault normal traction that decreases from a static value to a dynamic value as slip progresses Rock permeability (k) : Permeability affects how quickly pressure affects the fault stability. 4 permeability multipliers (for 4 reservoir layers) Rock porosity (f ): Rock porosity affects the total volume of fluid in the system. Fault permeability (k fx and k fz ): To model leakage across and along fault.

Sensitivity analysis on 2D Model Slide 14 Parameter (Units) Earthquake Time Earthquake Location Earthquake Magnitude µ s 1.980 0 1.807 µ d 0 0 0.007 d c (meters) 0 0 0.057 k (mdarcy) 0.062 0 0.025 k ft (mdarcy) 0.050 0 0.067 k fl (mdarcy) 0 0 ~ 0 Fault Plane 0.020 0.0584 0.084

Prior Distributions and True (synthetic) Model Slide 15 PRIOR DISTRIBUTIONS TRUE MODEL Parameter Distribution Mean Var Min Max Parameter True Value µ s TruncNormal 0.6 0.03 0.55 0.65 µ d TruncNormal 0.2 0.1 0.1 0.3 d c (meters) TruncNormal 0.1 0.2 0.05 0.3 k 1 (mdarcy) TruncNormal 300 200 100 500 k 2 (mdarcy) TruncNormal 300 200 100 500 k 3 (mdarcy) TruncNormal 300 200 100 500 k 4 (mdarcy) TruncNormal 300 200 100 500 f TruncNormal 0.1 0.01 0.08 0.12 k f,t (mdarcy) TruncNormal 5.00e- 06 2.00e- 06 1.00E- 06 1.00E- 05 k f,l (mdarcy) TruncNormal 5.00e- 06 5.00e- 06 1.00E- 06 1.00E- 05 µ s 0.6 µ d 0.15 d c (meters) 0.16 k 1 (mdarcy) 200 k 2 (mdarcy) 300 k 3 (mdarcy) 250 k 4 (mdarcy) 250 f 0.09 k f,t (mdarcy) 4.33E- 06 k f,l (mdarcy) 4.87E- 06

Observations/Model responses Slide 16 Earthquake Time: This time is defined as the first time when rupture is observed along the fault since the start of the simulation. Earthquake Magnitude: While inertial effects are ignored in our quasi-static simulation, a finite-element integration of the total amount of slip over the fault, and the energy liberation associated to that slip, can be used as an estimate of the earthquake magnitude. ò M = G dg 0 d G f 2 M w = log10 ( M o) - 6.0 3 where Mo = seismic moment Mw = earthquake magnitude G = shear modulus d = magnitude of slip vector Earthquake Location: Hypocenter is defined as the exact location (measured as distance from origin) where faultrupture is first observed (Fixed for this study!) 16

Gaussian Processes (GP) Total of 1000 simulations run to build GP model (80% training, 20% validation) Slide 17 Hyperparameter selection via maximum likelihood and cross-validation

Solution Inverse problem using MCMC + GP Slide 18 2D Model Solution q z) = z q ) q ) = z) ò z q ) q ) z q ) q ) dq Parameter Distribution Mean Std p µs Normal 0.50 0.03 - µd Normal 0.30 0.05 - dc (meters) Normal 0.30 0.01 - k (mdarcy) Normal 600 200 - kf,t (mdarcy) Normal 7e-6 1e-6 - kf,l (mdarcy) Normal 7e-6 1e-6 - Fault Plane Bernoulli - - 0.5

Solution Inverse problem using MCMC + GP Slide 19 3D Model Solution ò = = q q q q q q q q d p z p p z p z p p z p z p ) ( ) ( ) ( ) ( ) ( ) ( ) ( ) ( µ s µ d d c f k fl k ft

Solution Inverse problem using MCMC + GP Slide 20 k 1 k 3 Converge plots (sampling history and posterior) k 2 k 4 Geweke plots (two-sample test of means)

Solution Inverse problem using MCMC + GP Slide 21 Earthquake Time Earthquake Magnitude

Posterior Predictive Earthquake Risk Assessment Slide 22 Earthquake Magnitude CO 2 injection rate (MSCF/D)

Conclusions Slide 23 1. Uncertainty analysis seems critical for large-scale deployment of CCS 2. Data-fit models (e.g., GP) make inverse modeling (e.g., MCMC) attainable. 3. Earthquake attributes (time, location and magnitude) can be used to infer key geological properties 4. Posterior models can be used to quantify risk envelopes for key operational parameters.

Questions?