Predicting Vertical Resistivity By Machine Learning - Presentation - PDF document

See discussions, stats, and author profiles for this publication at: https://www.researchgate.net/publication/333812114 Predicting Vertical Resistivity By Machine Learning - Presentation Presentation · June 2019 CITATIONS READS 0 36 3 authors , including: Alexander Vereshagin Torolf Wedberg M Vest Energy AS smartFeatures.ai 46 PUBLICATIONS 156 CITATIONS 24 PUBLICATIONS 227 CITATIONS SEE PROFILE SEE PROFILE Some of the authors of this publication are also working on these related projects: Continuous follow up of the performance of the CSEM technology as a de-risking from the exploration toolbox. View project Machine Learning in Geoscience View project All content following this page was uploaded by Alexander Vereshagin on 17 June 2019. The user has requested enhancement of the downloaded file.

Predicting Vertical Resistivity By Machine Learning 1

Th_R11_05 (Predicting Vertical Resistivity By Machine Learning) Alexander Vereshagin*, Torolf Wedberg, Aristofanis Stefatos M Vest Energy * alexandre@mvestenergy.no 2

Background 3

Motivation Why measure anisotropy ( Rv ): • More precise estimation of S HC • Important for EM/CSEM: • (Anomaly in Rv, and/or anomaly in Rv/Rh ) ≈ DHI emgs.com • Dipping anisotropy challenge: high anisotropy → strong effect • It helps to cross-check with wells! 4

Motivation • Rh : normally available in wells (deep resistivity) Rv : scarce ( triaxial logging tools ). • 2017: analyzed all publicly available triaxial wells on Norwegian continental shelf (NCS) – 18 wells by that time. ( AAPG/SEG 2017, Wedberg et al. ) • No machine learning, just data • Results: subsurface is mostly anisotropic • Median( Rv/Rh ) ≈2.5 for available data (upscaled to formation level) 5

Limited availability of triaxial log (NCS: 25 wells vs. s. 6000+ in in DIS ISKOS !) DISKOS database, NCS 6

Machine Learning • Goal: Predict anisotropy for wells where triaxial data is not available • What to predict? • Anisotropy ( ani = Rv/Rh ) vs. Rv : ani has less value spread and less correlated to Rh • Type of model – key factors: • Input: basic composite logs • Not many wells, with missing intervals and bad points • → Recurrent, convolutional: heavy pre-processing • Most of the advanced algorithms work • Performance differ • “Sweet” functionalities (error bars, feature importance, …) 7

Features: correlations • Rv has strong correlation to RDEP, contrary to anisotropy • Do not want dominating features, so lets predict anisotropy! 8

Feature selection • Composite log: GR, RDEP, AC, ACS, NEU, DEN, PE, … To predict • More features → better scores But: missing data → larger error, worse scores • What else? depth = z • Vertical depth (Compaction!) • Water depth • Combined features • Geographic coordinates (coverage) • Geological formation (availability → area of application?) • … 9

Model preparation • Resampling • Some cleaning To predict • Filtering: good for scores, BUT: more filters → less feature values diversity → less prediction stability! depth = z z Feature 1 Feature 2 Anisotropy … Gradient boosting/ XGBoost, others can be plugged in 10

Boosted trees (e (e.g., ., XGBo Boost) Tree depth = 1 a True x 1 = p x 1 < p False b x 0 Tree depth = 2 a x 2 < q x 1 = p x 2 = q True b x 1 < p c x 0 False x 2 < q d 11 http://arogozhnikov.github.io/2016/06/24/gradient_boosting_explained.html

Training: cross-validation K-fold cross-validation Estimate how model will perform once fully trained scikit-learn.org Multiple (grid) runs : vary algorithm’s (hyper)parameters, feature combinations, train- test splitting Compute fit scores on test sets for each train-test Area (“class”) data split (try multiple metrics !) 25 wells = 25 “groups” ≈10^5 samples Select the best (hyper)parameters, FIX THEM Re-train on entire database with these fixed parameters. Fit should not change dramatically Save the model (and models for error bars) 12

4 parameters with 7, 5, 3, 7 values, respectively 7*5*3*7 = 735 parameter combinations (“sets”) Nr. of parameter set 735 1 13

Run for every set and plot the fit score… 14

… reorder the parameter sets 15

… reorder the parameter sets 16

… another score 17

… and yet another score 18

… and run for all train -test splits. 19

…which sets look good? 20

…which sets look good? 21

Error bar (quantile regression) Rh, RDEP Rv Anisotropy 22

Feature importance (“sweet” functionality) Dictated by Physics and data availability NEU-DEN depth RDEP DEN depth below water depth separation mudline “Combined” feature “Combined” feature Scikit-learn Gradient Boosting feature importance metrics 23

Testing: before including in training set Saturation? Rh, RDEP Rv Anisotropy 24

Testing: aft fter including in training set Saturation? Rh, RDEP Rv Anisotropy 25

Testing: before including in training set Rh, RDEP Rv Anisotropy 26

Testing: aft fter including in training set Rh, RDEP Rv Anisotropy 27

Testing: final scores R2 = 0.57 Scores for log 10 ( Ani ) Effectively, Metrics Testing Final anisotropy ratio MedAE ≈ 0.08 ≈ 0.06 error around 10 % RMSE ≈ 0.14 ≈ 0.12 ≈ 20m R 2 ≈ 0.5 ≈ 0.6 Why low R2? Anisotropy fit: (measured – predicted) 𝞽 ≅ 1 28

Upscaling to formation level R 2 (outliers removed) ≈ 0.8 - 0.95 Physics R V upscaled / R H upscaled R 2 (outliers removed) ≈ 0.8 - 0.95 Outliers: saturation, ehnaced by up-scaling 29

Expanding predictions : Training wells : M Vest recent CSEM inversions Atlantis Apollo 7325/1-1 1. Predicting ani (Rv) for entire interval of the 7324/7-2 7324/2-1 training wells (beyond the training data Wisting 7324/8-1 area 7220/4-1 interval) 7324/8-2 7220/5-2 7324/9-1 7220/5-1 Johan Castberg area CSEM resolves Rv → check with CSEM! 2. 7220/8-1 7220/8-U-1 7324/7-1 Adding 15 wells without triax but with CSEM. 7122/2-1 Hammerfest basin Ivory Stordal Unfortunately, currently we do not have access 6707/10-3 S 6705/7-1 6608/8-1 to modern CSEM cubes for any of our triax wells Phoenix 6608/11-7 S Vema 6706/11-1 (such cubes exist) 6608/11-4 Gymir 6706/11-2 6608/11-5 6608/11-8 Åsgard 6506/12-P-1 AH 6608/11-2 Maria 6608/11-6 6407/1-5 S 6608/11-3 Ormen Lange 6305/8-2 6608/10-6 6608/11-1 Knarr 34/3-3 S Kvitebjørn 34/11-A-16 16/1-21 A 16/1-21 S Ivar Aasen 15/3-A-5 Gudrun 16/1-22 S 15/3-A-12 T2 16/2-13 A Johan Svedrup 16/2-16 AT2 Ekofisk 2/4-K-4 A 30

Unconstrained ( ρ V ) ρ V Comparing to CSEM inversion (Example) Triaxial logs not available CSEM inversion has low vertical resolution, but resolve R Transverse and upscaled Rv and Rh Guided ( ρ V ) Rv Rv ρ V Rh, RDEP Rh, RDEP Missing features Missing features Constrained top ( ρ V ) ρ V Target interval 31

Transverse resistivity over 1.5 km depth interval Inversions of different age Lacking features in shallow section Lacking CSEM sensitivity in deeper section Rh, RDEP Missing features 32

Statistics for anisotropy @ different scales upscaling Sampled to geological formation scale Sampled to measurement scale Median: ≈2.2 Median: ≈ 2.65 Mean: ≈2.6 Mean: ≈ 3.6 33

Wedberg Updated triax Conservative Handling et.al. 2017 database prediction missing features Expanded anisotropy statistics (NCS) Nr. Fm intervals 87 159 365 450 Nr. unique Fms 32 43 80 80 Nr. wells 18 25 40 40 No missing input features Handling missing features Prediction median: ≈ 2.5 Triax database median: ≈ 2.65 34

Anisotropy vs HC Not all HC-bearing formations show strong anisotropy. There are anisotropic formations which are not HC- bearing. But among the formations with high anisotropy, there are more HC-bearing ones 35

Potential applications QC the log, add to database, re-train 3ax log present? • Create Rv log • Look for missed pay zones • Improve resource estimates • Make better background for future CSEM There is a well? • Analyze surrounding wells • Get Rv for the CSEM background • Raise alarm if chance for false-positive QC inversion, re-invert if needed CSEM inversion exist? Feasibility and sensitivity analysis Approach will work for other measurements 36

Conclusions • Predicting ani/Rv from basic composite logs with minimal feature selection • Optimization is important • Predicting very well when upscaled • Match with CSEM results • Which scale we can comfortably predict at? • Make a new score? • Useful for CSEM analysis, missed pays, log QC, etc. • Interesting anisotropy statistics over NCS, numbers stay stable • Future: • Mask before training? • to add latest triax wells • more CSEM results 37

References 38

Acknowledgements / Thank You / Questions Alexander Goncearenco (The National Institutes of Health, US) Lars Lorenz (Geonautika) Daniel Shantsev (EMGS, Norway) Alexander Vereshagin*, Torolf Wedberg, Aristofanis Stefatos M Vest Energy * alexandre@mvestenergy.no 39 View publication stats View publication stats