Seismic facies classification away from well control - The role of augmented training data using basin modeling to improve machine learning methods in exploration. Per Avseth (Dig Science) and Tapan Mukerji (Stanford University)
Dig Deeper Our vision: Digital transformation of explorational risking Conventional risking (sub-domain silos): Full integration with ML/Bayesian networks (models and data) P(trap) Trap Low sat. (Biogenic) Stiff Tight/ Low Perm P(res) AVO uplift P (Discovery) Reservoir Oil/Gas Source Leaked High Por. (brine) Seep P(source) Geoscience + ML/AI = Faster and better decisions! AVO
Augmenting training data using integrated models from expert domains The sparse data PROBLEM: Noise Data space (Well logs and calibrated seismic) Calibration/ validation Pseudowells Model space (what ifs) Extended seismic calibration Our next prospect We don t have well logs at every seismic trace, and seismic is acquired in a pre-defined sub-set of prospective area. New prospects may be located outside areas sampled by available well log data (and even outside seismic coverage). How do we train a ML algorithm to predict new prospects away from well control? P(x) Prior (obs) Ground truth (knowns & unknowns) Posterior (prediction) Likelihood (model) The rich model SOLUTION: We need domain knowledge and integrated models to augment machine learning Domain knowledge + Machine Learning (e.g. Bayes Ntw) = better and faster predictions Better predictions = More likely correct decisions x
Well-Logs Geology QSI with augmented training data Scenario testing based on geological expertise Augmented Training data Probability Density Functions (PDFs) Statistical Rock Physics Inverted Elastic Seismic Data Properties Seismic Inversion Bayesian Machine Learning Lithofacies Maps and Uncertainty 4
Case Example 1: AVO classification constrained by burial history in Loppa High Area, Barents Sea
AVO classification constrained by rock physics depth trends Once upon a time (The Leading Edge, 2003) We need to dig deeper! Extend technology by adding more G&G domain input/constraints: 1) Include diagenesis 2) Include tectonics (burial, uplift, erosion) 3) Honor sequence stratigraphic principles
Rock physics and AVO modeling constrained by burial history 1. Burial history 2. Diagenetic modeling (Walderhaug) 250 Deposition 0 Geologic Age (M.Yr) Cement volume Porosity Stø Fm Onset cement (Present day burial) Max. burial Burial depth (m) Temperature (degrees C) Depth/Temperature 3. Rock physics modeling (Dvorkin-Nur) 4. AVO modeling (Zoeppritz) Acoustic Imp. Oil Brine Oil Vp/Vs Brine + shale compaction and RP Gradient Deposition Brine Oil - + + Intercept Slide 7 Depth/Temperature - Shale (Background) Max. burial
Burial constrained AVO modeling to create syntethic training data Unconsolidated sand example: Oil-filled sand = AVO class III Burial curve Frying pan Shale Brine sand 70 C Oil Brine Shale Oil Compaction trend Brine Fluid trend Oil Brine AVO constrained by burial
Burial constrained AVO modeling to create syntethic training data Cemented sandstone example: Oil-filled sst = AVO class IIp Burial curve Frying pan Shale Brine sand 70 C Oil Brine Shale Oil Compaction trend Brine sst Fluid trend Oil Brine sst AVO constrained by burial
DIG DEEPER AVO and Burial History in Skalle/Juksa area, Loppa High, Barents Sea (Refs: Avseth and Lehocki, 2016; N. Johansen, 2017) Near Far Skalle Juksa Skalle Juksa Mech. comp. Skalle. Juksa 2km 1km 70C Cementation Uplift Juksa Skalle Skalle mod. Juksa mod. Uplift map (Johansen, 2017) 0 Juksa sst is slightly more cemented than Skalle sst!
Generating AVO training data for Skalle well (7120/2-3S) Brine properties: Vp Vs Rho mean 3.2 1.73 2.31 std 10% 10% 5% 1 0.8 0.6 Cov. 0.8 1 0.8 0.6 0.8 1 Vp Vs Rho 3.3 1.6 2.42 5% 5% 5% 1 0.8 0.6 0.8 1 0.8 0.6 0.8 1 Vp Vs Rho 3.0 1.5 2.5 5% 5% 5% 1 0.95 0.8 0.95 1 0.8 0.8 0.8 1
AVO classification constrained by burial history at Skalle well Skalle Juksa Shale Heterol. Brine Oil Gas -log(γ) Most likely brine saturated sandstones predicted at Juksa well
Simulation of AVO training data from burial trends at Juksa well (7120/6-3S) mean Vp Vs Rho 3.4 2.0 2.3 Vp Vs Rho 3.5 1.9 2.45 Vp Vs Rho 3.0 1.45 2.54 std 10% 10% 5% 5% 5% 5% 5% 5% 5% 1 0.8 0.6 1 0.8 0.6 1 0.95 0.8 Cov. 0.8 1 0.8 0.8 1 0.8 0.95 1 0.8 0.6 0.8 1 0.6 0.8 1 0.8 0.8 1
AVO facies/fluid classification constrained by burial history at Juksa well Skalle Juksa Shale Heterol. Brine Oil Gas -log(γ) Most likely oil saturated sandstones predicted at Juksa well
Inputs for nonstationary PDF-s
Stationary pdfs
Non-stationary pdfs Increasing maximum burial
Stationary
Nonstationary
Case Example 2: Integrating statistical rock physics and pressure and thermal history modeling to map reservoir lithofacies in the deepwater Gulf of Mexico (Wisam, Mukerji, Sheirer and Graham, Geophysics, July-Aug. 2018)
Basin Modeling (BPSM in one slide) Honoring the geology and solving for the physics in geologic time Modeling pressure and thermal history and rock properties Basin and Petroleum System Modeling - BPSM Structure Geologic Inputs Stratigraphy Rock Properties + + Simulation Coupled PDEs in time and space Calibration Model Outputs Predicted Rock Properties
0 10 km Comparative Study of QSI Scenarios Value of extrapolating pseudo logs at Well A 2 other wells (C and D) held out for validation Well B Well A Middle Miocene deep water sand reservoirs Thunder Horse North Field NW Well B Well A SE Thunder Horse Field salt Reservoir salt Reference Actual Well A Data Actual Well B Data Basin Modeling-Rock Physics Extrapolation at Well A Scenario 1 Scenario 2 22
Basin Modeling Outputs 2D basin model across Thunder Horse structure Spatial trends in effective stress and temperature conditions A B 23
Spatial Trend of PDFs Scenario 1: PDFs from well B alone Scenario 2: series of interpolated PDFs; Predicted spatial variations of Vp, Vs and density from basin modeling and rock physics Reference Case Scenario 1 Scenario 2 Vp (m/s) sandstone shale Distance (km) 0 2 km Sandstone Shale Bayesian classification Determination of most likely lithofacies and probabilities of lithofacies
Results: Improved sandstone thickness and volume predict Scenario 1: underestimates net volume by 23% Scenario 2: net volume difference of 0.5% only Scenario 1 Reference Ave. thickness error ~ 200 m Scenario 2 Reference Ave. thickness error ~ 25 m Validation well Well C Reference Workflow 1 1 well alone Workflow 2 1 well + BPSM & RP Depth (m) GR 0 1 0 1 0 1 Pr(Sand) Pr(Sand) Pr(Sand) Error (m) 25 Error (m) 25
G&G integrated with ML/AI (summary) Domain knowledge (Sedimentology, Basin Modeling, Rock Physics/QI) augments Machine Learning! Many sources of uncertainty: - geological scenario - geological heterogeneity - imperfect and incomplete data, - approximate rock physics models, Need for multiple possible Earth models (scenarios) Need for Uncertainty Quantification. Remember we are often looking for rare events!
Key take aways Machine learning not a black box We need G&G domain experts! Phase transition in massive computations and machine learning is an opportunity! How do we take advantage of this transition in our research and business?
Geosciences & Machine Learning If we can meet the challenges, If we can avoid the pitfalls, We can benefit from the opportunities Just dig it!
Acknowledgements Thanks to Dig Science colleagues (Kristin Dale, Tore Nordtømme Hansen, Kristian Angard, Carine Zeier, Reidar Muller). Thanks to Ivan Lehocki for key contributions Thanks to Lundin-Norway for collaboration/input to Skalle and Juksa discoveries (Article in The Leading Edge, 2016). Thanks to TGS Nopec for seismic data used in this study.