Successfully reported this slideshow.
Your SlideShare is downloading. ×

Paris data ladies #6

Ad
Ad
Ad
Ad
Ad
Ad
Ad
Ad
Ad
Ad
Ad

Check these out next

1 of 99 Ad

More Related Content

Similar to Paris data ladies #6 (20)

Advertisement

Paris data ladies #6

  1. 1. Paris Data Ladies#6 N I N A B E R T R A N D J U S T I N E D E S H A I S A D È L E G U I L L E T M A R I N E K R O L C H A R L O T T E L E D O U X A U R É L I A N È G R E @parisdataladies
  2. 2. Spark Streaming: From 101 to Production MEETUP DATALADY PARIS - APRIL 2018
  3. 3. About me Huikan XIANG ⬢ Data Engineer @ Deezer Core Data ⬢ 7 years experience on data related applications ⬢ Worked in startup environments 3
  4. 4. The Core-Data Team at Deezer 4 Recommendations/Search Analytics Royalties CRM Data WarehouseUsers app interactions Serving all data related needs for the entire company
  5. 5. The Core-Data Team Operates on a Massive Scale 5 1900 Daily jobs 1.5 PB of historical data 2.5 TB/day 100 machines 50+ data engineers & scientists 14M Active Users
  6. 6. The problem... 6
  7. 7. As Soon As Possible 7 I need the data to know what happened on Deezer
  8. 8. 8 REAL TIME DATA
  9. 9. 9 Adapting Recommendations in Real-Time Mozart
  10. 10. 10 Adapting Recommendations in Real-Time Beethoven
  11. 11. 11 Adapting Recommendations in Real-Time Bethoven
  12. 12. Why is real-time data analysis important? 12 Recommendation & search reactive content recommended and precise search ranking BI & product meteometer for product features, bug tracking Fraud detection fake fans detection as fast as possible CRM Fast emailing campaigns
  13. 13. Solutions: 13 DATA
  14. 14. Solutions: 14 Batch Processing DATA
  15. 15. Solutions: 15 Batch Processing DATA Streaming Processing
  16. 16. Batch vs Streaming 16 Batch Processing Streaming Processing Data scope All or most of the data Most recent data managed by rolling time window Data size Larges batches > GB Individual records or micro batches of few records Performance Latencies in minutes to hours Seconds or milliseconds Analytics Complex analytics Simple aggregates and rolling metrics (Not true anymore for Spark Streaming)
  17. 17. Why Spark Streaming 17 Tech Integration Already our batch processing engine, reusable component Multi-data sources integration Kafka, tables, files and third-tier APIs etc. Windowing Functions Rich and adaptive functions Distributed and Scalable Data partitioned
  18. 18. Spark Streaming 101 exercises 18 Goal Get your hands dirty Official Example NetCat and WordCount More Fun ! Trending topic in Twitter
  19. 19. Demo - Twitter Trending Hashtag 19 HTTP POST Processing Details batch interval: 2 secs Query Semantics aggregate window: 10 minutes Hardware local 2 cores, memory 512M Twitter Streaming API Spark Streaming Application Data Visualisation Application dropped Python, Flask framework1% Tweets Developer's trial app
  20. 20. Demo - Twitter Trending Hashtag https://github.com/Huikan/datalady-demo 20
  21. 21. Demo - Sample Code 21 TwitterUtils .createStream(ssc, None, filters) val ssc = new StreamingContext (sparkConf, 2 Seconds) Process details How often? Source
  22. 22. Demo - Sample Code 22 TwitterUtils .createStream(ssc, None, filters) .flatMap{status => status.getText.trim.split(" ").filter(txt => txt.startsWith("#")) .map((_, 1)) .reduceByKeyAndWindow(_ + _, Minutes(10)) .map{case (topic, count) => (count, topic)} .transform(_.sortByKey(false)) val ssc = new StreamingContext (sparkConf, 2 Seconds) Process details How often? Source Query semantics How to group? same as batch
  23. 23. Demo - Sample Code 23 TwitterUtils .createStream(ssc, None, filters) .flatMap{status => status.getText.trim.split(" ").filter(txt => txt.startsWith("#")) .map((_, 1)) .reduceByKeyAndWindow(_ + _, Seconds(10)) .map{case (topic, count) => (count, topic)} .transform(_.sortByKey(false)) .foreachRDD( rdd => {/** HTTP POST rdd top 10 */}) val ssc = new StreamingContext (sparkConf, 5 Seconds) Process details How often? Source Query semantics How to group? same as batch Sink/Receiver
  24. 24. Inspiring for production? 24 HTTP POST Twitter Streaming API Spark Streaming Application Data Visualisation Application
  25. 25. Production components 25 Rate Control Memory Tuning Failover Kafka Spark Cluster Durable Storage Monitoring Schema Registry
  26. 26. Deezer listen stream in real-time 26 Computing dynamic alloc 3 executors, 2 cores, 4GB RAM each executor Input rate peak to 4K records/s, 6MB/s Latency queryable in less than 1 minute
  27. 27. Next steps More applications 20% of our pipelines in streaming fashion Sync with spark streaming release State, structured API Build an universal message driven system 27
  28. 28. We recruit data ladies (and data boys) deezerjobs.com 28
  29. 29. Huikan Xiang Data Engineer hxiang@deezer.com linkedin.com/in/huikanxiang Thank you
  30. 30. DIFFUSION LIMITÉE – MODÉLISATION DE RETARDS POUR L’EXPLOITATION DES GRANDES GARES Marie de Faverges CEDRIC – CNAM DPF – SNCF Réseau Marie.milliet-de-faverges@reseau.sncf.fr
  31. 31. – ROUTAGE DES TRAINS EN GARE C’EST QUOI UN GOV ?
  32. 32. – ROUTAGE DES TRAINS EN GARE INFRASTRUCTURE COMPLEXE
  33. 33. – ROUTAGE DES TRAINS EN GARE INFRASTRUCTURE COMPLEXE
  34. 34. – ROUTAGE DES TRAINS EN GARE DES GOV FAITS ET VÉRIFIÉS À LA MAIN
  35. 35. – RÉSOLUTION Programmation linéaire en nombres entiers +Données : • Liste de trains qui arrivent ou partent de la gare • Liste d’itinéraires traversant la gare +Variables booléennes affectant un train à un itinéraire +Contraintes d’infrastructure, contraintes commerciales, etc Complexité +En théorie : problème NP-complet +En pratique : exple de Montparnasse : • 800 arrivées et départs par jour • 28 voies en gare • 1000 itinéraires d’arrivée et 1000 itinéraires de départ
  36. 36. – ROBUSTESSE Mais les trains n’arrivent pas toujours à l’heure ! +Le planning n’est plus valable +Les retards se propagent et s’amplifient à travers le réseau !" temps A !# Comment rendre les GOV plus robustes ? +Détecter les scenarios les moins robustes +Les pénaliser en objectif +La solution finale tendra à absorber une partie des retards +MAIS paramétrage très complexe et arbitraire
  37. 37. – LES RETARDS But +Modéliser les retards de trains quelques jours en avance +Utiliser ces résultats pour paramétrer l’outil Mais… +Des retards plus exceptionnels qu’on ne le pense +Des causes trop variées et rares pour prédire précisément But de l’étude +Estimer le risque de retard en prédisant sa loi +Cibler uniquement les petits retards (inférieurs à 20 minutes)
  38. 38. – MÉTHODOLOGIE 38 Données Brutes • Historique des retards • Grands travaux • Météo • Vacances Création du data set •Encodage des données •Troncature des retards •Sélection des variables Modélisation •Modèles linéaires généralisés •Validation Prédiction de probabilités de retards
  39. 39. – MODÈLES LINÉAIRES GÉNÉRALISÉS Trois composantes dans le modèle : +La composante aléatoire : ! = ($%, … , $() suivant une loi de la famille exponentielle +Un prédicteur linéaire : *+ = ∑-+.+ +Une fonction de lien / entre la composante aléatoire et le prédicteur linéaire : *+ = / 0+ Soit : 1 ~345678+91 : ; < =>? Les paramètres ? sont estimés par maximum de vraisemblance Exemple +Loi négative binomiale tronquée au-delà de 20 minutes +Deux paramètres 0 et @ ! ~ABC <, D EF/ < = >G?G EF/ D = >H?H
  40. 40. – PREDICTIONS R = 2 – Arrivée à 12h45 de La Rochelle un jeudi de vacances R = 16 – Arrivée à 20h de la Rochelle un mercredi de vacances R = 0 – Arrivée à 20h45 de Nantes un mercredi hors vacances
  41. 41. – PROBLÈME DE VALIDATION Difficulté +Défaut d’homogénéité entre les prédictions et les observations +Pas de méthode de validation adaptée au cas continu Validation des modèles binaires +Calibration : correspondance entre les probabilités prédites et les taux observés +Discrimination : capacité à différencier les succès des échecs en se basant sur les probabilités prédites ON VA ADAPTER LES VALIDATIONS BINAIRES À NOTRE MODÈLE CONTINU EN TESTANT LA CDF EN PLUSIEURS POINTS
  42. 42. – VALIDATION PAR CALIBRATION Calibration plot : +Pour un entier t on calcule pour chaque train ℙ(# ≥ % |') probabilité d’avoir un retard supérieur à t +On trie les trains par probabilités croissantes +On crée g groupes de tailles égales +On compare pour chaque groupe la probabilité prédite et la fréquence observée APPROCHE GRAPHIQUE
  43. 43. – VALIDATION PAR CALIBRATION Test Hosmer-Lemeshow : +Pour un entier t on calcule pour chaque train ℙ(# ≥ % |') probabilité d’avoir un retard supérieur à t +On trie les trains par probabilités croissantes +On crée g groupes de tailles égales +Statistique de Hosmer-Lemeshow : ) = + ,-. / 0., − 2., 3 2., + 05, − 25, 3 25, +Si le modèle est le bon H suit une loi du chi-deux à g-2 degrés de liberté +On veut que la p-valeur soit non significative pour ne pas pouvoir rejeter l’hypothèse nulle APPROCHE STATISTIQUE
  44. 44. – VALIDATION PAR CALIBRATION ! = # $%& ' (&$ − *&$ + *&$ + (-$ − *-$ + *-$ ! = 6,17 2 = 0,63 Groupe (- *- (& *& [0.03,0.29] 420 437 155 138 [0.29,0.34] 384 394 191 181 [0.34,0.38] 374 368 200 206 [0.38,0.42] 349 345 226 230 [0.42,0.46] 337 324 238 251 [0.46,0.50] 298 300 276 274 [0.50,0.54] 282 276 293 299 [0.54,0.59] 249 251 325 324 [0.59,0.65] 231 221 344 354 [0.65,0.86] 173 170 402 405 Régression logistique ponctuel/retard Test set de 5700 trains, g = 10
  45. 45. – VALIDATION PAR DISCRIMINATION Sensitivité et spécificité !"#!$%$&$%' = )* )*+,- !."/$0$/$%' = )- )-+,* La courbe roc représente la !"#!$%$&$%' et 1 − !."/$0$/$%' pour tout cutpoint % ∈ 0,1 Aire sous la courbe ROC Pour un couple 67, 68 avec 67 = 1 et 68 = 0 de probabilités de succès prédites 97 et 98 alors : :;< = ℙ 97 > 98
  46. 46. – RÉSULTATS GLM avec une loi négative binomiale entraîné sur 17 000 arrivées de TGV à Montparnasse Le modèle est évalué sur un set de 6 000 retards
  47. 47. – CONCLUSION Objectif de la thèse +Estimer le risque de retard de trains quelques jours en avance +Utiliser ces résultats pour paramétrer un outil de routage des trains en gare +But à terme : limiter la propagation des retards en gare Approche +Modèles linéaires généralisés pour prédire les lois de probabilité de retard +Validation par calibration et discrimination +Résultats : bonne calibration et discrimination stable
  48. 48. MERCI DE VOTRE ATTENTION ! DES QUESTIONS ?
  49. 49. your.name@uwa.edu.auName et al., 201X Visualizating the deep Australian continent with advanced geophysical signal processing Ph.D. Candidate Sophie Monnier Supervisors Prof. David Lumley, A/Prof. Jeffrey Shragge, Dr. Rie Kamei
  50. 50. Paris Data Ladies #6Monnier, 2018 Outline • Geophysics is cool!! • Seismic Data and Model Inputs • Seismic inversion method: Full Waveform Inversion • Results and Impacts
  51. 51. Paris Data Ladies #6Monnier, 2018 Geophysics: a tool to probe the Earth’s Interior Crust Mantle Outer Core Inner Core ~ 6400 km We live on the surface of planet Earth, but we cannot penetrate its depths… Earth’s deepest drilling: • Kola superdeep borehole • 12 km (0.1% Earth’s radius)
  52. 52. Paris Data Ladies #6Monnier, 2018 Geophysics: a tool to probe the Earth’s Interior Crust Mantle Outer Core Inner Core We live on the surface of planet Earth, but we cannot penetrate its depths… We cannot physically penetrate the Earth!!! ~ 6400 km Earth’s deepest drilling: • Kola superdeep borehole • 12 km (0.1% Earth’s radius)
  53. 53. Paris Data Ladies #6Monnier, 2018 So how do we know what we know? The study of seismic waves gives us a picture of the interior of the earth 1. A seismic source propagates in the earth… 2. …is recorded at seismic stations 3. …Gathering information shows seismic discontinuities in the Earth Earthquake seismology: deep earth scale (100 – 6000 km)
  54. 54. Paris Data Ladies #6Monnier, 2018 Exploration seismology: shallow earth scale (0.05 – 6 km) Seismology not only is cool but useful! 2D seismic image of buried geological layersFoetal ecography
  55. 55. Paris Data Ladies #6Monnier, 2018 " " !( !( !( !( !( !( !( !( !( !( !( !( !( !( 114°0'0"E 114°0'0"E 112°0'0"E 112°0'0"E 110°0'0"E 110°0'0"E 20°0'0"S 20°0'0"S 22°0'0"S 22°0'0"S Ü 0 100 km -100-3000-6000 Seafloor elevation (m) Bart 2D line Gascoyne abyssal plain Exm outhPlateau 110 E 112 E 114 E 20 S 22 S Cuvier abyssal plain New developments in Australian OBS technology New Australian OBS fleet: • 18 State-of-the-art seismic sensors • Inaugural survey • Acquisition year: 2014 – 2015 • Investigate deep WA structure
  56. 56. Paris Data Ladies #6Monnier, 2018 PhD Research Objectives receivers (OBS) ØObtain a high-resolution image of Australia’s deep continental structure offshore Western Australia ØInvestigate the performance of state-of-the-art seismic inversion methods on an innovative dataset ØHybrid project: crustal-scale seismology (5–30 km)
  57. 57. Paris Data Ladies #6Monnier, 2018 Outline • Geophysics is cool!! • Seismic Data and Model Inputs • Seismic inversion method: Full Waveform Inversion • Results and Impacts
  58. 58. Paris Data Ladies #6Monnier, 2018 !" !# !$ !% Moho Crust Mantle Source Receivers (OBS) Controlled-source acquisition of seismic data Crust Mantle Outer Core Inner Core ~ 12 800 km Ø Crustal-scale seismology: [10" ,10# ] km ~ 30km
  59. 59. Paris Data Ladies #6Monnier, 2018 !" !# !$ !% Crust Mantle Receivers (OBS) Controlled-source acquisition of seismic data Ø Crustal-scale seismology: [10# ,10" ] km ()*) = ,(!", !#, !$, !%)Source Moho Crust Mantle Outer Core Inner Core ~ 12 800 km ~ 20km
  60. 60. Paris Data Ladies #6Monnier, 2018 Modelling and inversion of seismic data !"#" = %('(, '*, '+, ',) Forward modelling Inversion • Seismic modelling “Mathematically simulate the propagation of seismic waves from a geological model of the subsurface” (Stekl et al., 1998)
  61. 61. Paris Data Ladies #6Monnier, 2018 !"#" = %('(, '*, '+, ',) Forward modelling Inversion • Seismic inversion '( '* '+ ', “Estimate a 2D or 3D Velocity model of the Earth” (Lailly, 1983) Modelling and inversion of seismic data
  62. 62. Paris Data Ladies #6Monnier, 2018 ! " # ! " $ What does seismic data look like?
  63. 63. Paris Data Ladies #6Monnier, 2018 What does a velocity model look like? Critical acquisition parameters: Ø Source Frequency content Ø Source-Receiver distance (offset) Ø Source/Receiver sampling Ø Choice of Objective Function ! 2 = $ 2% (Kamei et al. 2013) Full Waveform Inversion (FWI) Depth(km) Distance (km) km/s FWI Resolution:
  64. 64. Paris Data Ladies #6Monnier, 2018 Outline • Geophysics is cool!! • Seismic Data and Model Inputs • Seismic inversion method: Full Waveform Inversion • Results and Impacts
  65. 65. Paris Data Ladies #6Monnier, 2018 Full Waveform inversion: a local optimization approach Starting Model !" ≤ $ ? No: Update model !&'( = !& + +! Synthetic DataPreprocessed Field Data Data Misfit Yes: Convergence Raw Field Data
  66. 66. Paris Data Ladies #6Monnier, 2018 Full Waveform Inversion Pipeline 1. Build initial velocity model 2. Prepare and model data 3. Inversion parameter tuning 4. Interpret results
  67. 67. Paris Data Ladies #6Monnier, 2018 AuSREM, crustal component (Salmon et al., 2013) AusMoho (Kennett et al., 2011) 1D velocity profiles (Goncharov et al., 2016) Maurice Ewing cruise (Geoscience Australia, 2001) Synthetic Model " " !( !( !( !( !( !( !( !( !( !( !( !( !( !( 114°0'0"E 114°0'0"E 112°0'0"E 112°0'0"E 110°0'0"E 110°0'0"E 20°0'0"S 20°0'0"S 22°0'0"S 22°0'0"S Ü 0 100 km Bart 2D line Velocity Model Building 1. Build initial velocity model 2. Prepare and Model Data 3. Inversion parameter tuning 4. Interpret results
  68. 68. Paris Data Ladies #6Monnier, 2018 AuSREM Crustal model below Bart 2D line Distance from first shot point (km) Distance from first shot point (km) Depth(km) Bart 2D line Ewing (2001) Moho contour from Maurice Ewing velocity model Moho depth from AusMoho model Moho depth estimated from Goncharov et al. 2016 Velocity Model Building 1. Build initial velocity model 3. Inversion parameter tuning 4. Interpret results 2. Prepare and Model Data
  69. 69. Paris Data Ladies #6Monnier, 2018 Field Data " " !( !( !( !( !( !( !( !( !( !( !( !( !( !( 114°0'0"E 114°0'0"E 112°0'0"E 112°0'0"E 110°0'0"E 110°0'0"E 20°0'0"S 20°0'0"S 22°0'0"S 22°0'0"S Ü 0 100 km ? Bart 2 Field Receiver Gather, near-offset traces muted & AGC applied (window length: 30 s) Bart 2 Ø Processed field data 1. Build initial velocity model 3. Inversion parameter tuning 4. Interpret results 2. Prepare and Model Data
  70. 70. Paris Data Ladies #6Monnier, 2018 Bart 2 Synthetic Receiver Gather, with AGC applied (window length: 30s) Ø Bart 2 Synthetic " " !( !( !( !( !( !( !( !( !( !( !( !( !( !( 114°0'0"E 114°0'0"E 112°0'0"E 112°0'0"E 110°0'0"E 110°0'0"E 20°0'0"S 20°0'0"S 22°0'0"S 22°0'0"S Ü 0 100 km Bart 2 2D Synthetic Modelling 1. Build initial velocity model 3. Inversion parameter tuning 4. Interpret results 2. Prepare and Model Data
  71. 71. Paris Data Ladies #6Monnier, 2018 Ø Frequency and Offset for Subsurface Illumination Inversion Parameter Tuning Low frequency High frequency 1. Build initial velocity model 3. Inversion parameter tuning 4. Interpret results 2. Prepare and Model Data S R
  72. 72. Paris Data Ladies #6Monnier, 2018 1. Build initial velocity model 3. Inversion parameter tuning 4. Interpret results 2. Prepare and Model Data Ø Offset and Source-Receiver Sampling for Resolution 1 km 22 km Sampling Illumination Inversion Parameter Tuning
  73. 73. Paris Data Ladies #6Monnier, 2018 1. Build initial velocity model 3. Inversion parameter tuning 4. Interpret results 2. Prepare and Model Data Ø Offset and Source-Receiver Sampling for Resolution 1 km 6 km 18 km 22 km 18 km 9 km Sampling Illumination Inversion Parameter Tuning
  74. 74. Paris Data Ladies #6Monnier, 2018 1 km 6 km 18 km 1. Build initial velocity model 3. Inversion parameter tuning 4. Interpret results 2. Prepare and Model Data 22 km 18 km 9 km Sampling Illumination Inversion Parameter Tuning Ø Offset and Source-Receiver Sampling for Resolution
  75. 75. Paris Data Ladies #6Monnier, 2018 Outline • Geophysics is cool!! • Seismic Data and Model Inputs • Seismic inversion method: Full Waveform Inversion • Results and Impacts
  76. 76. Paris Data Ladies #6Monnier, 2018 Starting Model Distance from first shot point (km) Distance from first shot point (km) Depth(km) Bart 2D line Ewing (2001) Moho contour from Maurice Ewing velocity model Moho depth from AusMoho model Moho depth estimated from Goncharov et al. 2016 Initial Model 1. Build initial velocity model 3. Inversion parameter tuning 4. Interpret results 2. Prepare and Model Data
  77. 77. Paris Data Ladies #6Monnier, 2018 Final Model 1. Build initial velocity model 3. Inversion parameter tuning 4. Interpret results 2. Prepare and Model Data Bart 2D line Ewing (2001) Moho contour from Maurice Ewing velocity model Moho depth from AusMoho model Moho depth estimated from Goncharov et al. 2016 Distance from first shot point (km) Final Model Depth(km) Distance from first shot point (km)
  78. 78. Paris Data Ladies #6Monnier, 2018 1. Build initial velocity model 3. Inversion parameter tuning 4. Interpret results 2. Prepare and Model Data Seismic Inversion Results Before After • Crustal visibility: < 10 km • Moho not imaged • Tomographic resolution: 5-10 km • Crustal visibility: down to 25 km • Moho imaged at 20-25 km • FWI resolution: 1-2 km • Sensitivity analysis: • Subsurface Illumination • Resolution • Acquisition Design • Sensitivity analysis used for second OBS deployment over Lord Howe Rise Plateau
  79. 79. Paris Data Ladies #6Monnier, 2018 1. Build initial velocity model 3. Inversion parameter tuning 4. Interpret results 2. Prepare and Model Data Seismic Inversion Results Before After • Crustal visibility: < 10 km • Moho not imaged • Tomographic resolution: 5-10 km • Crustal visibility: down to 25 km • Moho imaged at 20-25 km • FWI resolution: 1-2 km • Sensitivity analysis: • Subsurface Illumination • Resolution • Acquisition Design • Sensitivity analysis used for second OBS deployment over Lord Howe Rise Plateau
  80. 80. Paris Data Ladies #6Monnier, 2018 Conclusion Ø The deep structure of the WA continent was imaged in high resolution thanks to high-quality seismic data, advanced geophysical processing and state-of-the-art seismic inversion methods Ø My analysis enabled better planning of seismic acquisition surveys offshore Australia Ø Geophysical Data Analysis and Machine Learning are meeting halfway! o Automatized geological interpretation o CNNs for induced seismicity prediction o Automatic detection of ore minerals
  81. 81. Paris Data Ladies #6Monnier, 2018 Conclusion Ø The deep structure of the WA continent was imaged in high resolution thanks to high-quality seismic data, advanced geophysical processing and state-of-the-art seismic inversion methods Ø My analysis enabled better planning of seismic acquisition surveys offshore Australia Ø Geophysical Data Analysis and Machine Learning are meeting halfway! o Automatized geological interpretation o CNNs for induced seismicity prediction o Automatic detection of ore minerals
  82. 82. Paris Data Ladies #6Monnier, 2018 Conclusion Ø The deep structure of the WA continent was imaged in high resolution thanks to high-quality seismic data, advanced geophysical processing and state-of-the-art seismic inversion methods Ø My analysis enabled better planning of seismic acquisition surveys offshore Australia Ø Geophysical Data Analysis and Machine Learning are meeting halfway! o Automatized geological interpretation o CNNs for induced seismicity prediction o Automatic detection of ore minerals
  83. 83. Paris Data Ladies #6Monnier, 2018 Conclusion Ø The deep structure of the WA continent was imaged in high resolution thanks to high-quality seismic data, advanced geophysical processing and state-of-the-art seismic inversion methods Ø My analysis enabled better planning of seismic acquisition surveys offshore Australia Ø Geophysical Data Analysis and Machine Learning are meeting halfway! o Automatized geological interpretation o CNNs for induced seismicity prediction o Automatic detection of ore minerals
  84. 84. Paris Data Ladies #6Monnier, 2018 Conclusion Ø The deep structure of the WA continent was imaged in high resolution thanks to high-quality seismic data, advanced geophysical processing and state-of-the-art seismic inversion methods Ø My analysis enabled better planning of seismic acquisition surveys offshore Australia Ø Geophysical Data Analysis and Machine Learning are meeting halfway! o Automatized geological interpretation o CNNs for induced seismicity prediction o Automatic detection of ore minerals
  85. 85. Paris Data Ladies #6Monnier, 2018 References • Alcock, M. B., Stagg, H. M. J., Colwell, J. B., Borissova, I., Symonds, P. A. and Bernardel, G., 2006. “Seismic Transects of Australia’s Frontier Continental Margins”, Geoscience Australia Record • Goncharov, A., Cooper, A. Chia,P. and O’Neil, P., 2016. “A New Dawn for Australian Ocean-Bottom Seismography.” The Leading Edge 35 (1): 99–104. • Kamei, R., Pratt, R. G. and Tsuji, T., 2013. “On Acoustic Waveform Tomography of Wide-Angle OBS Data-- Strategies for Pre-Conditioning and Inversion.” Geophysical Journal International 194 (2): 1250–80. • Kennett, B.L.N., Salmon, M., Saygin, E. & AusMoho Working Group,2011. “AusMoho: the variation of Moho depth in Australia.” Geophysical Journal International 187: 946-958. • Lailly, P. 1983. “The Seismic Inverse Problem as a Sequence of Before Stack Migrations.” In Conference on Inverse Scattering, Theory and Applications, Society for Industrial and Applied Mathematics, 206–20. • Salmon, M., B. L. N. Kennett, and E. Saygin. 2012. “Australian Seismological Reference Model (AuSREM): Crustal Component.” Geophysical Journal International 192 (1): 190–206. • Stagg, H.M.J., Alcock, M.B., Bernardel, G., Moore, A.M.G., Symonds, P.A. & Exon, N.F., 2004. “Geological framework of the outer Exmouth Plateau and adjacent ocean basins.” Geoscience Australia Record • Štekl, I., and Pratt, R. G., 1998. “Accurate visco-elastic modelling by frequency-domain finiti differences using rotated operators”, Geophysics, 63, 796-809 • Tortopoglu, B., 2015. “The Structural Evolution of the Northern Carnarvon Basin, Northwest Australia”, PhD Thesis • Zelt, C. A., Smith, R. B., 1992. “Seismic Traveltime Inversion for 2-D Crustal Velocity Structure.” Geophysical Journal International 108 (1): 16–34.
  86. 86. Paris Data Ladies #6Monnier, 2018 Thank you for your attention!
  87. 87. Paris Data Ladies #6Monnier, 2018 Thank you for your attention !! Questions?
  88. 88. Paris Data Ladies #6Monnier, 2018
  89. 89. Paris Data Ladies #6Monnier, 2018 1. Build initial velocity model 3. Inversion parameter tuning 4. Interpret results 2. Prepare and Model Data Inversion Parameter Tuning Ø Receiver Sampling and Objective Function for Resolution
  90. 90. Paris Data Ladies #6Monnier, 2018 1. Build initial velocity model 3. Inversion parameter tuning 4. Interpret results 2. Prepare and Model Data Inversion Parameter Tuning Ø Receiver Sampling and Objective Function for Imaging
  91. 91. Paris Data Ladies #6Monnier, 2018 1. Build initial velocity model 3. Inversion parameter tuning 4. Interpret results 2. Prepare and Model Data Inversion Parameter Tuning Ø Receiver Sampling and Objective Function for Imaging
  92. 92. Paris Data Ladies #6Monnier, 2018 1. Build initial velocity model 3. Inversion parameter tuning 4. Interpret results 2. Prepare and Model Data Inversion Parameter Tuning Ø Receiver Sampling and Objective Function for Imaging
  93. 93. Paris Data Ladies #6Monnier, 2018 1. Build initial velocity model 3. Inversion parameter tuning 4. Interpret results 2. Prepare and Model Data Inversion Parameter Tuning Ø Receiver Sampling and Objective Function for Imaging
  94. 94. Paris Data Ladies #6Monnier, 2018 1. Build initial velocity model 3. Inversion parameter tuning 4. Interpret results 2. Prepare and Model Data Inversion Parameter Tuning Ø Receiver Sampling and Objective Function for Resolution
  95. 95. Paris Data Ladies #6Monnier, 2018 1. Build initial velocity model 3. Inversion parameter tuning 4. Interpret results 2. Prepare and Model Data Inversion Parameter Tuning Ø Receiver Sampling and Objective Function for Resolution
  96. 96. Paris Data Ladies #6Monnier, 2018 1. Build initial velocity model 3. Inversion parameter tuning 4. Interpret results 2. Prepare and Model Data Inversion Parameter Tuning Ø Receiver Sampling and Objective Function for Imaging
  97. 97. L’association créée en 1987 Objectif principal favoriser la présence des femmes en mathématiques. Actions • Journées « filles et maths : une équation lumineuse », • Forum des jeunes mathématiciennes, • Interventions dans les classes, • Statistiques, …. Elle est également un lieu de rencontre et de promotion de la contribution des femmes à la recherche et à l’enseignement des mathématiques. femmes et mathématiques
  98. 98. femmes et mathématiques Depuis 2009, en partenariat avec Animath, Journées organisées autour de 4 temps forts: • Promenade mathématique , • Atelier sur les stéréotypes liés aux maths et découverte des métiers des maths, • Speed-meeting, • Pièce de théâtre-forum Dérivée Les 2 associations ont déjà organisé 76 journées En 2017: Ø 10 villes Ø 15 journées Ø 1250 lycéennes Ø 420 collégiennes En 2018, nous visons: Ø 15 villes Ø 20 journées Ø 1700 lycéennes Ø 550 collégiennes
  99. 99. femmes et mathématiques Nous avons besoin de vous ! Vous pouvez nous aider. 1) en nous faisant un don 2) en participant à des speed-meetings* 3) via du bénévolat de compétences 4) en organisant une journée dans vos locaux…. Vous pouvez nous joindre à fetm@ihp.fr Prochaines dates: 18 mai à Pau, 25 mai au Salon de la culture et des jeux mathématiques Place Saint Sulpice, 30 mai au Lycée la Mare Carré à Moissy- Cramayel

×