Neural network interpretation of LWD data (ODP Leg 170) con¢rms ...

Neural network interpretation of LWD data (ODP Leg 170)con¢rms complete sediment subduction at the Costa Rica

convergent margin

Erik Moritz a, Stefan Bornholdt b, Hildegard Westphal c, Martin Meschede a;*a Institut fu«r Geologie und Pala«ontologie, Universita«t Tu«bingen, Tu«bingen, Germany

b Institut fu«r Theoretische Physik, Universita«t Kiel, Kiel, Germanyc RSMAS, Marine Geology and Geophysics, University of Miami, Miami, FL, USA

Received 9 November 1998; accepted 13 October 1999

Abstract

The internal structure of a convergent plate boundary was the focus of ODP Leg 170 in 1996 at the subduction zoneoff Costa Rica. Although the structure of the subduction zone is rather well known from seismic surveys, prior todrilling of ODP Leg 170 it was a matter of discussion whether it is accretionary or non-accretionary. With a neuralnetwork approach, we confirm the evidence gained during drilling of Leg 170, that at least presently no lower-platesediments are transferred to the upper plate by accretion. To supplement lithological information, Logging-While-Drilling geophysical data have been included in this study and were interpreted in terms of lithology using a geneticallytrained artificial neural network. ß 2000 Elsevier Science B.V. All rights reserved.

Keywords: neural networks; data processing; Leg 170; subduction zones

1. Introduction

Costa Rica belongs to the Central Americanvolcanic arc system where the Cocos Plate hasbeen being subducted beneath the CaribbeanPlate for at least 90^100 Ma [1]. The convergentCosta Rica Plate margin has been extensivelystudied by geophysical investigations during thepast 10 years [2^9]. One of the early ¢ndingswas that a major part of the strata approaching

the subduction zone is underthrust beneath thetoe of the margin, and that the thickness of thesediments distinctly decreases within the ¢rst fewkilometers landward of the trench (Fig. 1). This isinterpreted presumably as a result of signi¢cantdewatering [10]. The mechanisms of accretionand underplating, tectonic erosion, deformationand dewatering, however, are still not unequivo-cally established [9].

During ODP Leg 170 the convergent margin o¡the Paci¢c coast of Nicoya, NW Costa Rica, wasdrilled. The general objectives of Leg 170 were todetermine the sediment, chemical, and £uid massbalances within the subduction complex and thelarger subduction system (including the volcanic

0012-821X / 00 / $ ^ see front matter ß 2000 Elsevier Science B.V. All rights reserved.PII: S 0 0 1 2 - 8 2 1 X ( 9 9 ) 0 0 2 7 0 - 8

* Corresponding author. Fax: +49 7071 5059;E-mail: [email protected]

EPSL 5312 20-12-99

Earth and Planetary Science Letters 174 (2000) 301^312

www.elsevier.com/locate/epsl

arc). One of the primary results was that most ofthe lower-plate sediment is bypassing the toe re-gion through underthrusting [11]. In this study wegive more precise data for the compaction processby correlating the reference section Site 1039,which is situated in front of the trench on theCocos Plate, with the underthrust section belowthe decollement at Site 1043 on the CaribbeanPlate (Fig. 1) with the aid of an arti¢cial neuralnetwork. The distance between the two sites isabout 500 m. We interpret Logging-While-Drill-ing (LWD) geophysical data in terms of lithologyusing an arti¢cial neural network. In combinationwith other geophysical parameters this enables usto quantify the rate of £uid expulsion out of theunderthrust section.

LWD data are measured during the drillingprocess at the frontal part of the drill string andare thus of outstanding quality. It is, therefore,possible to avoid measurement errors due to bore-hole breakouts and irregularities produced by theborehole liquids [11]. We demonstrate that theautomated lithological interpretation by arti¢cialneural networks is a highly feasible tool. Based onour correlations we con¢rm that in the outermostpart of the Costa Rican Plate margin virtually nosediment is scraped o¡ from the subducting CocosPlate and accreted to the overriding CaribbeanPlate. Compaction a¡ects predominantly theupper part of the subducted sedimentary pile.

2. Stratigraphy and material

The Cocos Plate carries a succession of hemi-pelagic and pelagic sediments from Early MiddleMiocene (approximately 16.5 Ma) to Holocene[10]. The reference section Site 1039 covers thecomplete and undeformed succession from Mid-dle Miocene to Quaternary sediments and termi-nates in a gabbroic sill. Three sedimentary andone intrusive unit were distinguished at Site1039 [10] (from top to bottom):

b Unit U1 consists of dark olive green diatoma-ceous ooze with ash layers. It is further subdi-vided into Subunit U1A (Holocene to Pleisto-cene; 0^5 mbsf), distinguished by abundant

graded sand layers interpreted as turbidites,and grades downward into Subunit U1B (Pleis-tocene; 5^84 mbsf), where graded sand layersare sparse to absent;

b Unit U2 is distinguished by a sharp decrease inbiogenic sediment. Subunit U2A (Early Pleisto-cene to Late Pliocene; 84^133 mbsf) consists ofdark olive green silty clay, and grades down-ward into Subunit U2B (Late to Early Plio-cene; 134^152 mbsf). Subunit U2B is a darkolive green silty clay interbedded with lightolive green calcareous clay; ash layers are com-mon throughout Unit U2;

b Unit U3 records a sudden increase in biogenicsedimentation from the nearly barren clays ofUnit U2 to calcareous and siliceous oozes. Sub-unit U3A (Early Pliocene to Late Miocene;152^180 mbsf) is composed of ivory coloredsiliceous nannofossil ooze interbedded with cal-careous clay; ash layers are sparse. SubunitU3B (Late to Middle Miocene; 180^280 mbsf)consists of ivory to light green and mottled sili-ceous nannofossil ooze with minor ash layers.Subunit U3B grades downward into SubunitU3C (Middle Miocene; 280^378 mbsf), andconsists of a mottled ivory-colored calcareousooze, interbedded in the lower part with sili-ceous ooze and matrix-supported breccia of cal-careous ooze clasts.

b Unit U4 (378^381 mbsf) consists of ¢ne- tomedium-grained glassy pyroxene gabbro withplagioclase glomerocrysts. Multiple chill zoneswere recovered within the gabbro re£ecting itsintrusive origin.

Drilling at Site 1043 penetrated the toe of thewedge into the underthrust sequence (Fig. 1). Onelithologic unit is de¢ned at this site above thedecollement: Unit T1 (Recent to Early Pliocene;0^150.5 mbsf) consists of thick intervals of clayand silty clay interbedded with relatively thin in-tervals of matrix-supported breccia. Thin layersand small pods of volcanic ash, and interbeds ofsilt and sand, appear throughout Unit T1. Belowthe decollement, Units U1B through U3A of thereference Site 1039 were recognized by the ship-board scienti¢c party. Coring stopped just abovethe contact with the nannofossil ooze. The 5.5 m

EPSL 5312 20-12-99

E. Moritz et al. / Earth and Planetary Science Letters 174 (2000) 301^312302

Fig. 1. (Top) Bathymetric map of the Middle America Trench o¡ the Nicoya Peninsula, Costa Rica, showing the location of theseismic line CR-20 [3] and the ODP Sites 1039^1043. (Bottom) Part of the depth-migrated seismic line CR-20 showing the pene-tration depth of Sites 1039, 1040, and 1043 [10].

EPSL 5312 20-12-99

E. Moritz et al. / Earth and Planetary Science Letters 174 (2000) 301^312 303

thick turbidite-rich Subunit U1A of Site 1039 wasnot observed in the cores at Site 1043 [11].

3. Methods

In situ downhole measurements were collectedusing the LWD technology. The LWD logs usedin this study respond to several petrophysicalproperties. Bulk density (romt) and resistivity(psr) logs re£ect changes in porosity, while gam-ma ray (gr) and photoelectric factor (pef) logsre£ect changes in lithology, e.g., clay mineralabundance [12].

For a detailed correlation of Sites 1039 and1043, a high-resolution stratigraphy was estab-lished to allow for the comparison of the twosites, one of which (1043) was not entirely cored.The logged records have been divided into four`Logging Units' (Table 1) based on obvious

changes in the log response [10]. Since the inten-tion of this study was to establish a correlation asdetailed as possible we used a cluster analysis tode¢ne more di¡erentiated units from the LWDlogs. It was possible to distinguish seven di¡erentcluster units in the logs of Site 1039 (Fig. 2, Table1).

The set of four geophysical parameters selectedfrom the LWD logs for this study is interpreted interms of lithologic composition. Ideally, the chos-en parameter set contains full lithological infor-mation; in this case the facies units can bemapped uniquely by the cluster units representedin the data. In our study the choice of parametersand the high quality of the LWD data enable ahigh degree of correlation between cluster andlithologic units (see Table 1). This allows for anaccurate lithological interpretation by an auto-mated algorithm. Here, we use a geneticallytrained arti¢cial neural network to do this task,

Table 1Comparison between lithology, Logging Units [10], and Cluster Units

Unit Subunit Depth (mbsf) Age CorrespondingLogging Unit

CorrespondingCluster Unit

U1 U1A 0^5 Holocene toPleistocene

Logging Unit 1 C1A

U1B 5^84 Pleistocene C1B; ashes de¢nedas C5

U2 U2A 84^133 Early Pleistocene toLate Pliocene

Logging Unit 2 C2

U2B 134^152 Late to EarlyPliocene

U3 U3A 152^180 Early Pliocene toLate Miocene

U3B 180^280 Late to MiddleMiocene

Logging Unit 3 C3; ashes de¢nedas C6

U3C 280^378 in Hole B;363^445 in Hole C

Middle Miocene

U4 378^382 in Hole B;423^445 in Hole C

Logging Unit 4 C4

EPSL 5312 20-12-99


a new approach which allows a highly accurateinterpretation of the data. Its details are describedin Section 4.

4. Neural network approach

Arti¢cial neural networks are data processingdevices useful for modeling large amounts ofdata, in particular when the exact relationshipsbetween di¡erent variables are not known or toocomplicated to derive exact models. Neural net-works are trained in a training phase and, aftertraining is completed, can be used to interpret yetunknown data. Therefore, they are useful formany geological problems of interpreting largeamounts of empirical data. The application of ar-ti¢cial neural networks to geological problems is arelatively recent approach. Baldwin et al. were the¢rst to propose the use of neural networks foridenti¢cation of lithologies [13]. Wiener et al.showed that porosities can be estimated using aneural network [14]. The development of the ap-plication of neural networks on well log interpre-tation is summarized in [15].

Arti¢cial neural networks were originally moti-vated by biological neural networks (for an over-view and a collection of classical papers see [16]).Such devices are capable of parallel data process-ing and are well suited for pattern recognition anddata classi¢cation or clustering tasks. A neuralnetwork is trained with a training data set in or-

der to create an internal representation of clus-tered data. It can perform such tasks even whenthe volume, number of variables or the diversityof the data is very large [17]. In the present casethis ability is used to obtain a representation ofthe cluster units in such a network, on which basisthe facies units will be classi¢ed during the auto-matic interpretation of lithology.

A neural network consists of interconnected

Fig. 3. Interconnected computing cells of a neural network:four input cells receiving the values of four input parame-ters; seven intermediate processing units; seven output cells,representing seven di¡erent cluster units.

Fig. 2. Cluster analysis of selected parts of Hole 1039 D. As an example, the cross plot of gamma ray vs. photoelectric factor isshown.

EPSL 5312 20-12-99


computing cells (sometimes called `neurons' aftertheir biological counterparts). The connectionsbetween the neurons transmit signals with di¡er-ent strengths, given by an associated weight valuefor each connection. The application of an arti¢-cial neural network consists of two phases: (1) atraining phase, during which the network learnsand (2) a utilization phase during which the net-work can be used to interpret new unknown data.During the learning phase, the network learns topredict a desired output from a training data setby iteratively modifying the weights and compar-ing the predicted output with the given desiredone.

The neural network used here consists of anumber of interconnected computing cells (Fig.3): four input cells receive the values of fourLWD logging parameters: resistivity, bulk den-sity, photoelectric e¡ect, and gamma log. Thesesignals are transmitted to seven intermediate proc-essing units and subsequently to seven outputcells, representing seven di¡erent cluster units(C1a, C1b, C2, C3, C4, C6, and C7). These layersof computing cells are interconnected with di¡er-ent strengths, given by an associated weight valuefor each connection. The total number of 91weights is determined during a training phase,where the network learns the correlations betweenLWD logs and the corresponding facies, de¢nedby the Cluster Unit. In our case, a training dataset of 176 representative values is taken from Site1039 D, where both LWD logs and lithologicaldata were available. Standard procedures existfor this training task, for example back-propagat-ing the error through the network [18]. Thesestandard training procedures, based on so-calledgradient descent optimization, do not guaranteeobtaining the optimal trained network for a givendataset. This is, however, a crucial requirementfor a high-resolution LWD interpretation as re-quired in the present study. Therefore, we usehere an approach based on a global optimizationtechnique to adjust the weights, called a geneticalgorithm [19]. This method of well log interpre-tation using neural networks trained with geneticalgorithms has been applied ¢rst and described indetail in [20].

The basic principle of a genetic algorithm is an

optimization procedure that is inspired by biolog-ical evolution. In the current problem, 50 neuralnetworks are set up to compete with each other,and by a procedure of selection of the ¢ttest, re-production, mutation and crossover, the qualityof the networks is gradually increased (Fig. 4).For the purpose of this algorithm, each networkis represented as a binary number by encoding theweights as a sequence of binary numbers, eachstring of numbers de¢ning an individual neuralnetwork. An initially random set of networks,i.e. random strings of binary numbers, is eval-uated by comparing the actual output of the net-works with the desired output, that is, in this case,the correct lithologic unit as known from Site1039 D. The neural networks with the lowest per-formance are removed whereas successful net-works are reproduced. Thus the optimizationprocess involves the principle of `survival of the¢ttest'. By means of simulated mutation andcrossover new networks are produced on the basisof the successful networks. This procedure is iter-ated until the networks converge towards the de-sired output, usually for a few thousand itera-tions. The resulting network is then ready tointerpret geophysical LWD data in terms of litho-logical facies as de¢ned by the Cluster Unit.

The neural network used in this study wastrained using a training data set of 176 values.The data set consists of a list of input parametersand the corresponding facies, de¢ned by the Clus-ter Unit. The network was allowed to learn untilno further e¡ective improvement occurred. Thetrained network then was used to interpret thecomplete geophysical data from the Sites 1039and 1043.

5. Results

After successful training, we used the networkto interpret the LWD data from Sites 1039 and1043. To evaluate the prediction capabilities ofthe network when confronted with new, unknowndata, we ¢rst tested the trained network by com-paring its interpretation of the geophysical welllogs from Site 1039 to the known lithologiesfrom this well. Since lithologies are known at

EPSL 5312 20-12-99


Site 1039 from cores, the interpretation of thenetwork can be taken as an evaluation of its per-formance. Fig. 4 compares the output of thetrained network with the Logging Units and lith-ologies de¢ned by the shipboard scienti¢c party ofLeg 170.

Fig. 4 illustrates that the boundaries of theCluster Units of Site 1039 identi¢ed by the net-work correlate closely with the Logging Units ofthe same site. Cluster Unit C1A (`turbiditiclayers') is found to correlate to the lithologicUnit U1A in the uppermost section (from 0 to4 mbsf). The lithologic Units U1B (diatomaceousooze), U2 (silty clay) and U3 (calcareous ooze)correspond to the Cluster Units C1^C3. Litho-logic Unit U4 (pyroxene gabbro) is recognizedas Cluster Unit C4 at a depth of 395 mbsf. Errorsof correlation between the units are in the range

of a few to less than a meter. The ash layers(Cluster Units C5 and C6) interpreted by the neu-ral network correlate exactly with major ashlayers described in the cores of 1039 (i.e. at84 mbsf, 91 mbsf, 351 mbsf; not shown in thelithology column of Fig. 4). The distinction be-tween Cluster Units C2 and C6 is sometimes am-biguous since they are features of very similargeophysical parameters which is demonstrated inthe overlap of the clusters (Fig. 2). These units,however, are restricted to depth intervals whichdo not overlap, thus allowing for an exact corre-lation.

After this proof of the quality of our trainednetwork for lithologic interpretation, the trainedand evaluated network was employed to interpretthe cluster units from the geophysical data of thelower part of Site 1043 from the decollement at

Table 2Pore space volumes of the incoming sediments, derived by summation of LWD neutron porosity log

ClusterUnit

Thickness(m)

Initial porosity(1039)(%)

Initial £uidcontent (1039)(m3/m2)

Compaction(1043)(%)

Derived£uid content(1043 B) (m3/m2)

Di¡erence(£uid expulsion)(m3/m2)

C 1B 114 80 91 71 64 27C 2 71 82 58 80 46 12C 3 206 63 130 0 130 0Sum 391 279 86 240 39

Fig. 4. Principle of genetic algorithms. Individual networks are encoded as strings of binary numbers. An initially random popu-lation is evaluated. Selection, reproduction and simulated crossover and mutation increase the ¢tness of the population (from[20]).

EPSL 5312 20-12-99


147 mbsf down to total depth of 471 mbsf. Thissite was only cored down to a depth of 282 mbsf.Comparison of the output of the network for Site1043 with that of Site 1039 (Fig. 5) indicates thatthe lithologic succession of Site 1043 is very sim-ilar to that of Site 1039. Almost the entire sedi-mentary column of the subducting plate is presentin the underthrust section of Site 1043. Main dif-ferences are the missing Cluster Unit C1A (`turbi-ditic layers') at the top, and the basal Cluster UnitC4 (gabbro). The absence of Cluster Unit C4 isdue to the fact that drilling did not proceed all theway down to the basaltic sill at Site 1043. LWDdrilling stopped at a total depth of 482.3 mbsf, thesill was projected at V490 þ 10 mbsf. Based on adrastic reduction in the rate of penetration and apiece of broken basaltic material at the drilling bitafter pulling out the drill string it is thought thatdrilling stopped more or less directly above thebasal gabbroic sill [10]. Major ash layers identi¢edat Site 1039 are clearly detected at Site 1043 (e.g.,at 205, 210, 446 mbsf).

In the section of Site 1043 directly beneath thedecollement (150^170 mbsf) a displacement ofbulk density, resistivity and gamma ray log upto higher values is observed. This is probablycaused by a signi¢cant dewatering of this sectiondirectly below the decollement [10]. The shift ofthe geophysical data compared to the respectiveinterval at Site 1039 results in an overestimationof the thickness of ash layers in the output of thenetwork.

The detailed correlation allowed an estimationof the compaction in the individual Cluster Unitsof the underthrust section of 1043. Cluster UnitC1B is compacted to 71% and Cluster Unit C2 to80% of the thickness of the corresponding ClusterUnit in the reference hole. Based on the correla-tion using the identi¢ed ash layers, compaction isthought to be virtually absent in Cluster Unit C3(Fig. 5). Reduction in thickness of this unit is onlyobserved in the interval from 447 to 453 mbsf.

Assuming that the reduction in thickness is

solely a result of loss of porosity and that thepore space is completely saturated initially, weused the determination of compaction in the in-dividual Cluster Units for a simple estimation of£uid expulsion: taking an initial £uid content ofabout 279 m3/m2 (derived by summation of theLWD neutron porosity log [10]) and a total com-paction of 86% for the complete incoming strata,we derive a total volume of around 39 m3/m2

£uid, which is expelled during the ¢rst 500 m ofsubduction (Table 2, Fig. 5).

Taking a convergence rate of about 91 mm/year[21] the subduction of the incoming sediments(distance between the decollement tip line andSite 1043 = 500 m) is achieved in approximately5500 years. The resulting £uid expulsion per m2

would then be 0.007 m3/year assuming a linearrelation. The total volume of expelled £uids overthis distance (regarding a slice of 1 m in width)accumulates to 3.5 m3/year. The derived £uid £owis a rough estimation, because de¢nition of actual£uid £ow requires detailed knowledge of path-ways and pressure gradients within the subductedsediments.

6. Discussion

The correlation between Sites 1039 and 1043enabled us to assess changes in thickness of theindividual sedimentary units. Our correlation cor-responds to the correlation based on ash layers bySaito and Goldberg [22] but includes all litholo-gies present in the boreholes. Moreover, our neu-ral network approach avoids the subjectivity in-herent in visual approaches. The correlationshowed reduced thicknesses in Cluster Units C1and C2 of the underthrust section of Site 1043(Fig. 5). The overriding Caribbean Plate isthought to have rapidly compacted these sedimen-tary units during the subduction process. A sim-ilar response to the subduction was not found inthe lower Cluster Unit C3.

Fig. 5. Correlation between reference section at Site 1039 and the section below the decollement at Site 1043 using the output ofa genetically trained neural network. Also shown is the lithology and Logging Units (taken from [10]), and the calculated com-paction. Note: only Units C1 and C2 show compaction e¡ects. See text for further explanation.

C

EPSL 5312 20-12-99


EPSL 5312 20-12-99


Factors explaining the reduced compactibilityof Cluster Unit C3 could be: (1) its greater age,(2) its lower porosity and, therefore, its lower per-meability, and (3) a longer upward path of £uiddrainage. Additionally, the diatoms of lithologicUnit U3 are able to withstand the pressure of theoverlying masses and, therefore, largely preservetheir initial porosity. Also the low heat £ow at theCosta Rica convergent margin [23], resulting inless dissolution of diatoms compared to otherareas, may play an important role. This canwork both ways as dissolution and reprecipitationof a small amount of silica as cement couldstrengthen the sediments considerably (B. Clen-nell, personal communication).

Between 447 and 453 mbsf of Unit C3 at Site1043 a reduction in thickness is observed. Basedon core descriptions this interval is composed of amatrix-supported breccia [10] which was not ob-served at Leg 170 Site 1040 (Fig. 1) where coresare available down to the basaltic sill layer. Wetherefore suggest that this interval is primarilymissing at Site 1043 and that the reduced thick-ness is not necessarily a result of compaction.

Similarly, the top 5 m (turbiditic layers) of thesedimentary column are missing in the under-thrust section of Site 1043 (Fig. 5). A seismic pro-¢le run between these sites shows, however, thatthis uppermost turbiditic layer pinches out beforereaching the base of the slope. Silver et al. inter-preted the absence of the turbiditic layers as aresult of bottom currents [24]. Our analysis ofthe LWD data con¢rms that the complete succes-sion of sediments approaching the subductionzone (cored at reference Site 1039) is found inthe underthrust section of Site 1043. This meansthat little if any of the incoming sediment is ac-creted frontally to the upper plate. This interpre-tation is strengthened by the composition of sedi-ments above the decollement at Site 1043 (UnitT1). If the deformed sedimentary wedge at Site1043 were composed of material scraped o¡ thesubducting Cocos Plate, then the composition ofthe sediments should be similar to those drilled atSite 1039. These sediments are rich in carbonatesin the lower part (lithologic Unit U3, see Table 1)and change to diatomaceous rich hemipelagicclays in the upper part (lithologic units U1B^

U2B, see Table 1). The sediments above the de-collement are of very di¡erent composition andshow hardly any resemblance to the referencesite [10]. They consist predominantly of pure clayswith rare fossils. It seems likely that these sedi-ments represent material of the sedimentary apronthat has been shed downslope to the toe of theslope [25]. A lithologic unit similar to the greendiatomaceous ooze with ash layers of Unit U1Bor the ivory colored siliceous nannofossil ooze ofUnit 3B has not been observed in sections abovethe decollement at Sites 1040 and 1043. If therewas frontal o¡scraping, at least some of thesesediments should occur in the sedimentary wedgeabove the decollement. Additional evidence forthe complete subduction of the incoming stratacomes from Site 1040 [10] where again only thetop few meters of the turbidites present at Site1039 are missing. Although this `point study' isbased on only two drilling sites it supports thenotion of presently non-accretionary conditionsalong the Middle America Trench o¡ Costa Rica.

7. Conclusions

High-resolution correlation of ODP Sites 1039and 1043 (Leg 170) o¡ Costa Rica based on litho-logical interpretations of geophysical LWD datawas performed using a genetically trained neuralnetwork. The correlation of the two sites con¢rmsthat (1) signi¢cant £uid expulsion occurs veryearly in subduction history, (2) compaction a¡ectsthe underthrust section depending on lithologicalbehavior, and (3) the model of sediment accretionat the Costa Rican convergent margin that wasaccepted prior to the drilling of ODP Leg 170fails to explain the completeness of the incomingsedimentary section below the decollement.

We conclude that:

1. Compaction a¡ects only the upper 140 m ofthe subducted sedimentary pile, whereas thelower 180 m appear to lack signi¢cant compac-tion related to subduction. We interpret this asthe result of the di¡erence in lithologic compo-sition and longer £uid pathways. Overlyingmasses prior to subduction, the higher compac-

EPSL 5312 20-12-99


tion resistivity of diatoms, and the anoma-lously low heat £ow o¡ Costa Rica couldalso play an important role in withstandingcompaction e¡ects.

2. The stratigraphic succession observed in theunderthrust section (Site 1043) displays thecomplete succession found in the undisturbedsection in front of the subduction zone on theCocos Plate (Site 1039). This con¢rms that cur-rently no frontal o¡scraping, which would addsediment to an accretionary wedge, takes placein the area of Leg 170. The entire sedimentarysuccession is thus subducted beneath the Car-ibbean Plate.

Further research will be required for a betterunderstanding of the fast compaction of theunderlying sediment pile induced by the subduc-tion process. These studies should include a de-tailed examination of £uid £ow, especially consid-ering pressure gradients and £uid pathways withinthe subducted strata. Also, the question on whatit depends whether sediment is accreted or notremains to be solved.

Acknowledgements

We gratefully acknowledge the GEOMAR Re-search Center for Marine Geosciences, Kiel, forproviding the computing capacities. Financialsupport was given by the German Science Foun-dation (Projects Bo 1242/2 and Me 915/13). PeterZweigel helped to calculate the £uid mass balan-ces. With the detailed comments and advice of thejournal reviewers Roland von Huene and ThomasH. Shipley, as well as an anonymous journal re-viewer we were able to considerably improve themanuscript. Thanks to Ben Clenell and Al Boltonfor information on the compaction behavior ofthe sediments ![CL]

References

[1] M. Meschede, W. Frisch, A plate tectonic model for theMesozoic and Early Cenozoic history of the CaribbeanPlate, Tectonophysics 296 (1998) 269^291.

[2] P.L. Sto¡a, M.K. Sen, Nonlinear multiparameter optimi-zation using genetic algorithms: Inversion of plane-waveseismograms, Geophysics 56 (1991) 1794^1810.

[3] T.H. Shipley, K.D. Mcintosh, E.A. Silver, P.L. Sto¡a,Three-dimensional seismic imaging of the Costa Rica ac-cretionary prism: structural diversity in a small volume ofthe lower slope, J. Geophys. Res. 97 (1992) 4439^4459.

[4] T.H. Shipley, P.L. Sto¡a, D.F. Dean, Underthrust sedi-ments, £uid migration paths, and mud volcanoes associ-ated with the accretionary wedge o¡ Costa Rica: MiddleAmerica Trench, J. Geophys. Res. 95 (1990) 8743^8752.

[5] K.D. Mcintosh, E.A. Silver, T.H. Shipley, Evidence andmechanics for forearc extension at the accretionary CostaRica convergent margin, Tectonics 12 (1993) 1380^1392.

[6] R. von Huene, E. Flueh, B. Cropp, T. Csernok, E. Fabel,J. Ho¡mann, K. Emeis, P. Holler, G. Jeschke, C. Lean-dro, I. Perez-Fernandez, J. Chavarria, A. Flores, D. Es-cobedo, R. Leon, O. Barrios, Morphotectonics of the Pa-ci¢c convergent margin of Costa Rica, in: P. Mann (Ed.),Geologic and Tectonic Development of the CaribbeanPlate Boundary in Southern Central America, GeologicalSociety of America Special Paper 295, 1995, pp. 291^307.

[7] K. Hinz, R. von Huene, C.R. Ranero, Tectonic structureof the convergent Paci¢c margin o¡shore Costa Rica frommultichannel seismic re£ection data, Tectonics 15 (1996)54^66.

[8] S. Ye, J. Bialas, E.R. Flueh, A. Stavenhagen, R. vonHuene, G. Leandro, K. Hinz, Crustal structure of theMiddle America Trench o¡ Costa Rica from wideangleseismic data, Tectonics 15 (1996) 1006^1021.

[9] R. von Huene, E. Flueh, A review of marine geophysicalstudies along the Middle America Trench o¡ Costa Ricaand the problematic seaward terminus of continentalcrust, in: H. Seyfried, W. Hellmann (Eds.), Geology ofan Evolving Island Arc, Pro¢l 7, Stuttgart, 1994, pp. 143^159.

[10] G. Kimura, E. Silver, P. Blum, ODP Leg 170 shipboardscienti¢c party, Proc. Ocean Drilling Program, Initial Re-ports, vol. 170, College Station, TX (Ocean Drilling Pro-gram), 1997.

[11] G. Kimura, E. Silver, P. Blum, P.J. Fox, J.F. Allan,T.J.G. Francis, Ocean Drilling Program, Leg 170, Prelimi-nary Report, College Station, TX, 1997, 86 pp.

[12] Schlumberger (1991) Log Interpretation Principles/Appli-cations, Schlumberger Educational Services, Houston,TX.

[13] J.L Baldwin, D.N Otte, C.L Wheatley, Computer emula-tion of human mental process: application of neural net-work simulation to problems in well logging interpreta-tion, Soc. Petrol. Eng., Pap. 19619 (1989) 481^493.

[14] J.M. Wiener, J.A. Rogers, J.R. Rogers, R.F. Moll, Pre-dicting carbonate permeabilities from wireline logs using aback-propagation neural network, Soc. Petrol. Eng. Ex-panded Abstr. Biogr. Tech. Program 1 (1991) 285^289.

[15] J.H. Doveton, Geologic log analysis using computermethods, AAPG Comput. Appl. Geol. 2 (1994) 169 pp.

[16] J.A. Anderson, E. Rosenfeld (Eds.), Neurocomputing:

EPSL 5312 20-12-99


Foundations of Research, MIT Press, Cambridge, MA,1988.

[17] J. Hertz, A. Krogh, R.G. Palmer, Introduction to theTheory of Neural Computation, Addison-Wesley, Red-wood City, CA, 1991.

[18] D.E. Rumelhart, G.E. Hinton, R.J. Williams, Learningrepresentations by back-propagating errors, Nature 323(1986) 533^536, Reprinted in Anderson and Rosenfeld,1988.

[19] S. Bornholdt, Genetic algorithms, in: T. Gramss et al.(Eds.), Non-Standard Computation, Wiley-VCH Verlag,Weinheim, 1998.

[20] H. Westphal, S. Bornholdt, Lithofacies prediction fromwireline logs with genetic algorithms and neural networks,Z. Dtsch. Geol. Ges. 147 (1996) 465^474.

[21] C. De Mets, R.G. Gordon, D.F. Argus, S. Stein, Currentplate motions, Geophys. J. Int. 101 (1990) 425^478.

[22] S. Saito, D. Goldberg, Physical and hydrologic propertiesin the Costa Rica and Barbados Subduction Zones esti-mates from ODP downhole logging, EOS Trans. Am.Geophys. Union 79 (1998) F395.

[23] M.G. Langseth, E.A. Silver, The Nicoya convergent mar-gin ^ a region of exceptionally low heat £ow, Geophys.Res. Lett. 23 (1996) 891^894.

[24] E. Silver, ODP Leg 170 shipboard scienti¢c party, Com-plete sediment subduction and implications of £uid £owin the Middle America Trench o¡ Costa Rica, Joides J. 23(1997) 1^3.

[25] M. Baltuck, E. Taylor, K. McDougall, Mass movementalong the inner wall of the Middle America trench, CostaRica, in: R. von Huene, J. Auboin et al. (Eds.), InitialReports, DSDP, 84, US Government Printing O¤ce,Washington, DC, 1982, pp. 551^570.

EPSL 5312 20-12-99


Neural network interpretation of LWD data (ODP Leg 170) con¢rms ...

Documents

Transcript of Neural network interpretation of LWD data (ODP Leg 170) con¢rms ...