Home

John H. Doveton: Kansas Geological Survey

Stephen E. Prensky, U.S. Geological Survey, Denver

[Originally published in 1992, The Log Analyst, v. 33, no. 3, p. 286-303.]

Geological log analysis has emerged as a distinctive sub-discipline of log analysis. It is built on and an extension of the traditional treatment of logging data in terms of reservoir engineering and geophysics. This paper uses summaries of published case studies to review recently introduced techniques and the major developments in geological applications. Although presented as separate sections, many of the topics discussed are interrelated The nature of these interconnections and the diversity of topics reflects the significant growth in geologic log analysis.

For many years, the practical business of petrophysics appeared to be the domain of the engineer, with a tight focus on the economic implications associated with the properties of porosity and water saturation. Even the presentation of log-analysis theory seemed an alien departure from the traditional training of many geologists. This perception contributed to the huge publishing success of the textbook by Asquith and Gibson (1982) in which basic log analysis concepts were presented by geologists for a geological audience. In the same vein, Asquith (1985) went on to publish techniques applicable specifically to carbonates and a practical summary of all the models that have been applied to shaly sandstone evaluations (Asquith, 1990). During the same period, Darwin Ellis accepted a challenge by Stanford to teach a course on logging that would realistically address the needs of earth scientists. The experience led to the textbook by Ellis (1987), which is marked by a particularly readable and authoritative treatment of the physics of the latest generation of tools and how they relate to rock properties. Systematic methods using well logging to measure the physical properties of rocks is also the focus of the text by Hearst and Nelson (1985) who emphasize fundamental concepts of tool theory and rock physics.

Books that elaborate on the broader geological (rather than the narrower reservoir engineering) aspects had a slower start for a variety of reasons. However, their arrival was inevitable, as increasing numbers of geologists worked routinely with logs, and the SPWLA moved inexorably away from a membership dominated by engineers to a majority of geologists. The classic book, Geologic well log analysis by Pirson (1970) had no rival for many years, although its authorship by an engineer is reflected in both its style and topic selection. The prolific and creative work of Oberto Serra and his coworkers finally led to a monumental two-volume treatise (Serra, 1984, 1986). The concept of "electrofacies" was a particularly useful contribution by Serra and Abbott (1980) and has been widely adopted as a bridge to connect logging measurements with the classical facies approach of sedimentary geology. In addition, Serra made extensive use of dipmeter analyses, closely integrated with other logs and profiles of bedding and textural properties. Collectively, these provide a valuable atlas in the interpretation of sedimentary environments from logs (Serra, 1985).

Rider (1986, revised 1991) published a readable and popular book on the geological interpretation of logs, drawn partly from his work with Serra's group, although he pointedly declined to write on dipmeter interpretation. In his book, Doveton (1986) emphasized the role of computer methods in the transformation of logs to profiles of lithology and mineralogy within individual wells and as maps of variation across regional areas. The appearance of the "Bibliography of Well-Log Applications" in The Log Analyst (Prensky, 1987), followed by annual updates, provided a valuable reference source for geological applications. Even a cursory glance through these references (or those listed in this article) shows the relatively limited penetration of mainstream geological journals, at the present time.

The conventions of the SPWLA and its sister societies have always provided forums for geological studies drawn from logs. However, in 1988, a two-day meeting with the theme "geological applications of wireline logs" was convened in London by the Geological Society. In the proceedings volume of this meeting, Hurst et al. (1990) boldly asserted in the introduction that they believed this meeting to be the first of its kind. The enthusiastic response to this meeting and the published proceedings demonstrated the wide interest in the subject and a second meeting on this theme was held in 1991 (GAWL II), also in London. Maybe it was worth the wait, because the developments of the last 10 years have been truly innovative and have the potential to make a major impact on mainstream geology.

AUTOMATED LOG CORRELATION

The principal geological application of logs has always been subsurface stratigraphic correlation. Manual correlations are highly labor-intensive tasks whose end results are subjective and often hotly debated. The potential advantages of automated correlation have been the stimulus for the design of many different computer methods since the pioneering study by Moran et al. (1962). Real progress has been made in the last few years toward the solution of significant problems that have often stymied earlier attempts. Although computer programs can correlate trivial examples, they often fail in more complex situations. Solutions may be either geologically or geometrically absurd or the program method may be unable to distinguish between several reasonable alternatives. The application of artificial-intelligence (AI) methods and the development of expert systems has now provided a means to codify both geometrical constraints and human experience within sets of rules. An expert-system component can guide the data-handling and numerical processing of correlation algorithms, and one has been successfully field-tested on realistic correlation problems (see, for example, Olea and Davis, 1986; Kuo and Startzman, 1987; and Figure 1).

Previous methods of automated correlation could handle simple relative vertical displacement, but often had limited success in accommodating missing sections and differential stretch of correlative intervals between wells. These features result from a normal geological history that includes multiple episodes of erosion and nondeposition, which punctuate periods of continuously varying rates of sedimentation. The occurrence of "gaps" and "stretching" are common to sequence matching problems in other scientific areas, such as linguistic analysis and genetic-string comparison. Dynamic programming methods developed in these fields have been adapted to correlation of wireline logs to select the optimum correlative match from all possibilities. Both missing sections and differential stretch are accommodated automatically. The methodology has proved widely successful, and applications are described by a number of authors, including Wu and Nyland (1986), Lineman et al. (1987), and Griffiths and Bakke (1988).

The lack of digital data, where and when you need it, remains a major obstacle to the routine application of computer methods for regional or even fieldwide correlation. This situation continues to slowly improve with the growth of digitized log databases and archives of digital well-log tapes. Recent changes in technology have radically accelerated the costly and time-consuming mass transfer of log curves into digits. Logs are now routinely scanned by devices such as facsimile machines to yield raster (bitmap) images, which can then be stored efficiently on optical discs. Software is available to screen out background grids automatically and to follow log curves for a raster to vector (digital) conversion with minimal user intervention. These recent developments, summarized by Leonard (1990), hold great potential for transforming the vast amount of logging data now frozen in paper records into digits for use in correlation as well as for computer-processed subsurface geologic studies.

ELECTRICAL IMAGING METHODS

Since their introduction in 1986, electrical borehole-imaging tools have made a great impact in a variety of commercial and academic geological applications. Gray-level processing of multiple microresistivity curves can result in images of the borehole wall with impressive resolution (Figure 2). The technology of these tools is described elsewhere in this issue in a companion review paper by Maute (1992) and marks a logical step in the evolution of the dipmeter.

The images often mimic conventional pictures in a striking fashion and so can be compared readily with core. This property has been found particularly useful for boreholes drilled for scientific research, e.g., the Ocean Drilling Program (ODP), where extensive intervals are cored or the borehole is continuously cored. For intervals where core recovery is poor, the images can be used to interpolate missing information, and for intervals where core is recovered, images and core can be matched to establish core orientation.

The images represent conductivity measurements of the borehole wall and register fractures (both open and healed), thin beds, sedimentary structures and other features. The techniques of interpretation are described by Harker et al. (1990); useful reference material for images of clastic rocks is provided by Luthi (1990) and of carbonate rocks by Nurmi et al. (1990). Image processing is now performed routinely on workstations, either as an unaided interactive task or with assistance from an AI procedure (Startzman and Kuo, 1989).

A major application of electrical imaging is in the detection and evaluation of fractures (Casarta et al., 1989; Standen, 1991), a goal that has always been elusive using traditional logs. The recent boom in horizontal drilling, with its primary targets of fractured formations, has stimulated additional interest in this area. Laubach et al. (1988) compared fractures detected by electrical imaging with those obtained from the acoustic borehole televiewer, and their validation by core. Hornby et al. (1990) related fracture aperture widths computed from both electrical scans and reflected Stoneley waves. A number of useful articles are also reprinted in the SPWLA Borehole Imaging volume (Paillet et al., 1990).

RESERVOIR MINERALOGY AND LITHOFACIES ANALYSIS

Multiple "porosity" logs (density, neutron and sonic) have been used for several decades to arrive at true effective porosities independent of matrix mineralogy. Although manual crossplots are still used for this purpose, computer solutions to this problem are increasingly common and generate a compositional profile of minerals as a byproduct. The power of these tools to discriminate mineralogy has also been significantly enhanced through the addition of the photoelectric-absorption curve from the spectral density log. The mathematical process involved is essentially the solution of simultaneous equations that link the unknown volumes of selected minerals and their hypothetical log responses with log records of sedimentary sections. The result is displayed typically as a graphic profile such as shown in Figure 3, and represents the inversion of the original log curves to compositional traces. Consequently, their vertical resolution is controlled by that of the tools (about 0.5 m) and thus represents a moving average of actual variation.

The computer program for solving a determined system (when the number of logs is sufficient for a unique component solution) is simple and can be run quickly on even the smallest computers. The ready availability of fast computers makes them a practical medium for interactive geological processing of wireline logs. Some degree of interaction is necessary for thoughtful analysis due to limiting assumptions concerning linearity, idealized log responses, correct identification of mineral components, borehole environment, and tool errors. By involving the user as a participant, changes in mineral suites at various depths can be recognized and an appropriate selection made from alternatives at all levels. The strategy of running multiple models in parallel is described by Quirein et al. (1986), with the final choice dictated either by probability concepts or user intervention.

In reality, the link between rock compositions and log responses is both underdetermined and nonlinear. Because lithologies are typically dominated by a few components and linearity is generally a reasonable approximation, an experienced analyst can usually generate a satisfactory solution. More expansive approaches to the problem incorporate constraints, accommodate nonlinearities, and take account of tool errors in an iterative process that locates an optimal result with minimal incoherence between the solution and the logs. However, the relative slowness of these more complex programs makes interaction impractical. Furthermore, the difficulties in specifying many of the input parameters means that mathematical optimality does not necessarily mean geological reality. Consequently, interaction with simpler models on workstations is still usually favored by most analysts for routine applications, as discussed by Marett and Kimminau (1990).

Statistical Methods for Lithofacies Analysis

The inversion methods described in the previous section are rooted in a deductive or a "top-down" approach, where solutions of mineral composition are the consequences of models set by the user. The alternative strategy is inductive or "bottom-up," where inferences on lithofacies are drawn directly from patterns observed in log-response associations. Distinctions between characteristic response groupings are then used for classification and subdivision of log sequences. The basic concept was set out by Serra and Abbott (1980), who coined the term "electrofacies" and described a manual procedure based on simple graphical motifs. The pattern-recognition tasks involved were later adapted as a semiautomated procedure in which the complementary skills and limitations of machine and human are put into play. The use of principal-component analysis reduces the dimensionality of the problem from a potentially bewildering multiple-log representation to a lower order space. The axes of this space are composite logs that systematically absorb the information content while screening out statistical noise from various sources. Using cluster analysis, electrofacies are then identified by localized clouds of points. However, the intelligent analyst intervenes at this stage to ensure that the final clusters have interpretable geological meaning based on core observations or geological insight. This automated electrofacies method was originally introduced by Wolff and Pelissier-Combescure (1982), and a useful multiwell case study is described by Widdicombe et al. (1984).

The mathematical methods used in this approach are taken from the standard multivariate analysis toolbox and so are readily available for any log analyst who works with logs on a computer. In recent years, practical applications appear to be increasing in number, probably due in large part to the increasing use of microcomputers and workstations by many analysts. Clear signs of this grass-roots change are indicated by the wide readership of the popular "petrophysics" section published in Geobyte, under the direction of Robert Elphick, and the burgeoning membership of the computer-oriented geological societies.

Explanations of multivariate statistical techniques and their application to well logs for pattern recognition and classification tasks can be found in a number of dispersed sources. Doveton (1986) devoted a chapter of his book to mathematical analysis of log trends and patterns, whereas Hayes (1989) provided a useful overview of statistical well-log pattern-recognition methods in his Ph.D. dissertation. Elek (1988) showed how principal-component analysis could be applied to zonation and well-log correlation. Both Busch et al. (1985) and Anderson et al. (1988) described the application of discriminant function analysis to lithological classification from well logs. Other techniques and case studies included cluster analysis (Robinson and Reeves, 1989), fuzzy-set theory (Griffiths, 1989), and Kruskal multidimensional scaling (Matyas, 1990).

Statistical Prediction of Physical Properties

The primary mission of the statistical methods described in the previous section is to serve as automated pattern-recognition devices that link log responses with associations of rock properties. In more traditional applications, statistical line- and curve-fit methods have been used for many years for both calibration and prediction. Because core analysis data are commonly accepted as the reference standard for reservoir evaluation, porosity logs are generally calibrated against core data. The differing vertical resolution of the two measurements requires smoothing of the core data to give common vertical resolution.

The choice of line-fit and estimation procedure by one of several statistical models is still debatable. Only a few papers have been written in this area, but they are generally thoughtful studies that provide useful insights into systematic data analysis for the working log analyst . Etnyre (1982, 1990) wrote a two-part series on weighted least-squares methods applied to formation evaluation, followed by an explanation of the petrophysical uses of the robust-Marquardt statistical procedure (Etnyre, 1990). Rodriguez et al. (1989) described the determination of confidence intervals for petrophysical parameters.

Many log analysts are unfamiliar with the finepoints underlying the concepts and interpretation of statistical methods as applied to petrophysical data. However, the increasing role of digital databases as a component in routine log analysis is a major stimulus for their increased understanding of statistical processing. Basic statistical procedures have been great sources of argument and discussion during the major field unit operating disputes of the 1980s. It is interesting to speculate that the amount of money that hung on the choice of regression method in these equity battles is the greatest in the history of statistics! Basic explanations of statistical methods in a logging context such as the paper by Mitchell and Nelson (1988) are useful for a wide readership. However, a certain degree of caution is appropriate, because of the potential snares of misinterpretation and blind faith in "numbers."

The problems caused by differing sampling volumes associated with core and various logging devices have been recognized for many years. The most common and practical means to bring the measurements to a common vertical resolution is through the statistical smoothing of core data to the coarser scale of the wireline logs. This loss of detail has prompted efforts over the years to reverse the process, by enhancement of the vertical resolution of logging measurements.

Research in this area has been further stimulated by the great interest in thinly bedded reservoirs in recent years (see e.g., Gundeso and Gronvold, 1990; Chaudhary and Vashist, 1991). Applications of log analysis techniques to thin beds were described by Ruhovets (1989), their integration with core data by Sinha et al. (1989), and potential pitfalls in thin-bed enhancement resolution by Minette (1990).

Some progress in finer vertical resolution has been made through the introduction of improved tool designs (Tittman, 1991), particularly with respect to induction logging (Silva and Spooner, 1991) and electrical borehole imaging (discussed earlier). The alternative approach is to create a more finely resolved log through computer processing of data recorded by logging tools with coarser resolution. The actual vertical variation of the logged property can be considered to be averaged or "convolved" by a filter, which is determined by the tool's measurement characteristics. The goal of "deconvolution" is equated with the design of an inverse filter or procedure that essentially reverses the averaging process. The desirability of this is easy to understand, but difficult to implement in a practical and convincing manner. The nonlinear responses of the resistivity tools are mathematically difficult to deconvolve, even for the service companies who are privy to their design characteristics. The measurements of the nuclear tools are stochastic (statistical, rather than deterministic) and so are confounded with counting error. Attempts to amplify the signal in these statistical data must therefore not allow the "noise" to be amplified beyond tolerable limits. Looyestijn (1982) provided a useful and sobering review article that explains these problems. Nevertheless, the great value of even modest improvements continues to stimulate research in this area, such as that reported by Galford et al. (1986), Elkington et al. (1990), and Nelson and Mitchell (1990).

Some limited progress continues to be made on the prediction of permeability from logs by the use of multiple-regression methods. Log predictions of permeability are still most commonly based on porosity estimates alone. These predictions are often very poor, unless the rock shows little change in pore-size characteristics, because of the failure to take into account the variations in internal surface area. Internal surface area is often related to rock framework textural, mineralogical, and geochemical properties that influence many logs. Although no log measures internal surface area directly, some logs can function as surrogate variables for surface area and be incorporated with porosity in multiple regression models for permeability prediction and error analysis. A very useful review paper of these methods is given by Wendt et al. (1986), who also describe in detail, and critically, the successes and limitations of their own experiences with Prudhoe Bay data.

In studies that draw on older logs, the choices of additional variables are generally restricted to gamma-ray transforms and multiple porosity log indicators of shale content and changes in matrix minerals. However, the geochemical logs (reviewed in detail later) have a great potential for this type of application, as they record a suite of elemental measures. These elements reflect textural properties through their sensitivity to matrix mineralogy, which is, in turn, the product of depositional and diagenetic processes. Herron (1987) introduced this concept of using geochemical logs as surrogate variables for the internal surface area in a multiple regression model based on the classic Kozeny-Carman relationship.

Both simple and multiple regression models have also been applied to the prediction of organic content in the evaluation of source-rock potentials. Logging tools are now available to estimate organic carbon directly from the carbon/oxygen ratio (Herron, S.L., 1986). However, organic content has a distinctive effect on gamma-ray, sonic, resistivity, neutron, and density logs (e.g., Schmoker, 1981; Mendelson and Toksoz, 1986), and these relationships can be used as the basis for statistical prediction (Krystinik and Charpentier, 1987) when calibrated with core data.

ARTIFICIAL-INTELLIGENCE APPLICATIONS

A few years ago, the field of artificial intelligence was dominated by "expert systems," but "neural networks" have recently emerged as a serious competitor. The two approaches have radically different philosophies, but both have found legitimate and interesting applications in log analysis.

Expert Systems

Expert systems attempt to emulate simple reasoning, drawing inferences from data as dictated by a knowledge-base of rules. The rules are a codified mix of objective constraints and more subjective material, which distill the experience and judgment of acknowledged experts in the problem area. Log analysis is widely recognized as a blend of art and science, so that expert systems provide a useful methodology to attempt to capture a lifetime’s experience from older log analysts. Recent examples of some expert system prototypes for general log analysis are described by Peveraro and Lee (1988) and Einstein and Sutherland (1989). Einstein and Edwards (1988) also discuss a comparison between the performance of human experts and expert systems in log analysis and interpretation.

By far the most well-known log analysis expert system has been the Dipmeter Advisor, developed by Schlumberger (see Smith and Baker, 1983). Dipmeter interpretation is an obvious application because it has always been a particularly tricky mixture of systematic analysis, pattern recognition, and experienced judgments. A general theory of the rule-based approach to dipmeter processing is also described by Kerzner (1988). The Dipmeter Advisor is probably more famous in the AI research community than among log analysts because it is considered one of the few expert systems that is used on a routine, daily basis for real-world applications rather than as an interesting curiosity.

Automated log correlation is also an appropriate task for a rule-based approach, because certain solutions can usually be discarded as geometrically impossible or geologically improbable (see Olea and Davis,1986; Lineman et al., 1987). It is now widely recognized that expert systems generally perform well in applications where the problem has clearly defined constraints and goals, but can "fail" in more complex and subtle situations, most particularly where the human "experts" disagree among themselves!

Neural Networks

Neural networks are drawn from models of the brain that see the processing of information as the result of excitation of simple neurons, which are richly interconnected on a massive scale. Although today’s neural networks cannot attempt to approach the complexity of the human brain, some powerful applications can be developed using the basic design features of simple "neuron" units interconnected as a network in a parallel-processing operation. The easiest task for a neural network to attempt is a supervised problem, where the network "learns" the pattern of input responses (log readings) that correspond with a desired output (matrix or fluid characteristics). The current procedures are often timeconsuming since the learning is an iterative process. However, there are several advantages over the classical statistical pattern recognition methods, including a focus on all sample patterns, rather than just summary parameters, and a lack of assumptions concerning linearity or normality. On the debit side, it is often difficult to establish how a successful solution was arrived at based on input patterns. This contrasts with expert systems, where an "audit trail" can quickly establish which rules were invoked in any system decision.

Simple explanatory examples in the determination of lithology from a neural network are outlined by Rogers et al. (1992), using a back-propagation learning algorithm. This material provides an introduction to the neural network applications to log analysis given by Baldwin et al. (1989a, 1989b), who also describe an "unsupervised" application for lithofacies recognition (see Figure 4). Unsupervised pattern recognition is a more difficult problem because it requires the network to teach itself from log data presented to it. Furthermore, these self-taught patterns should have some utility and meaning to the human network handler. Derek et al. (1990) compare the performance of neural networks and statistical pattern-recognition methods in sandstone lithofacies identification.

There are great potential rewards in using this approach as an aid in interpretation of complex lithologies utilizing all measurements made on current and future logging programs. For example, current statistical methods for predicting permeability from logs are limited by their linear structure and estimation of parameters. By contrast, neural networks can handle non-linearities and are nonparametric so that they may be more effective for permeability prediction (Rui-Lin and Chen-Dang, 1991). Runge and Runge (1991) also show how the simulated annealing property of neural network operation can be applied to obtaining blocked logs from curves.

The application of AI to log analysis is still in its infancy and results are often documented in poorly accessible journals. However, progress in the field can now be monitored by attending or reading the proceedings of the annual conference on "Artificial Intelligence in Petroleum Exploration and Production", held at Texas A&M University.

APPLICATIONS OF TOTAL GAMMA-RAY AND SPECTRAL GAMMA-RAY LOGGING

The gamma-ray log has been widely used in geological interpretation for many years as a means to assess both shale content and implied grain-size variation. So, for example, Selley (1974) described the application of gamma-ray profiles in conjunction with glauconite and carbonaceous material as a "cowboy geology" method to aid in recognition of ancient sedimentary environments. However, because of the different sources of radiation, interpretations of this type are often ambiguous and Rider (1990) points out that care must be taken, especially in systematic work.

The introduction of the spectral gamma-ray tool in the 1970s marked a major advance in our ability to determine the specific contributions of the potassium, uranium, and thorium isotope series. Early applications focused on the resolution of reservoir evaluation problems such as the distinction of micas from clays in Jurassic sandstones of the North Sea (see Hodson et al., 1976), and the recognition of fracture systems with uranium mineralization in the Austin Chalk (see Fertl et al., 1980).

Hassan et al. (1976) explored the differentiation of clay minerals and other radioactive species using the ratio of thorium concentration to potassium concentration. Both thorium (by adsorption) and potassium (chemical composition) are associated with clay minerals, so that the ratio expresses relative potassium richness as one indicator of clay-mineral species, as well as being diagnostic of other radioactive minerals. A widely used thorium-potassium crossplot based on broad expectations of ratio fields that are associated with single minerals was published by Quirein et al. (1982). Hurst (1990) cautioned against the "bland generalizations" that underlie such a crossplot, because the chemistry of thorium and potassium associated with clay minerals are determined both by source and diagenetic history. Also, as a nuclear log, the measurement is a stochastic property, where considerations of precision and accuracy are important, particularly at low counts, such as within sandstones.

Because most shales are composed of a mixture of clay minerals, the use of the photoelectric cross section (Pe) in conjunction with the Th/K ratio is an additional help to interpretation. The photoelectric cross section is a direct function of the aggregate atomic number. Ellis (1987) points out that differences in atomic number between quartz and clay minerals can be attributed mostly to iron content. As a result, values in photoelectric absorption are ordered from a low in kaolinite to successively higher values, through smectite, illite, to chlorite, basically as a function of increasing iron content. However, in the final analysis, systematic volumetric estimates are made difficult by the presence of other accessory minerals, as well as the variation in composition of clay minerals, so that these wireline measures provide generalized indications of the compositional aspects of shales.

Based on their analyses of numerous rock samples, Adams and Weaver (1958), in a classic paper, demonstrated the utility of the thorium-to-uranium ratio as an indication of relatively oxidizing or reducing conditions. The two elements are normally associated geochemically. Whereas thorium has only one valency state, which is insoluble, uranium has two valency states, of which the lower is also insoluble, but the higher is soluble and can be removed in solution. The ratio therefore provides a useful indication of relative reduction or oxidation, but whether this can be attributed to depositional or diagenetic mechanisms requires additional information. Adams and Weaver (1958) further suggested that ratios of <2 were highly suggestive of relative uranium enrichment and, by implication reducing conditions, as contrasted with ratios >7, which indicated preferential removal of uranium, possibly by leaching.

Doveton (1991), using ratios from logs run in a Cretaceous/Permian sequence in central Kansas, offers an example of geological interpretation from spectral gamma-ray logs using these concepts. Figure 5 shows striking, and readily interpretable patterns: an abrupt shift in the Th/K ratio that occurs at the Cretaceous-Permian contact and highlights clearly the major basal Cretaceous unconformity. At this depth, the potassium-rich illite-feldspar signature of the Cedar Hills Sandstone changes to a Lower Cretaceous trace, which oscillates between illitic and kaolinitic clay minerals facies, possibly linked with marine and deltaic freshwater environments, respectively. The high-amplitude variations of the Th/K ratio log in the Graneros Shale and Greenhorn Limestone may reflect the occurrence of volcanic ash (bentonites) observed in the drill cuttings, interbedded with normal illitic marine shales.

Based on Th/U ratios and the diagnostic values suggested by Adams and Weaver (1958), an oxidizing environment is indicated for much of the Cedar Hills Sandstone, which would be consistent with its postulated origin as eolian sands. Stacked repetitions of high and medium Th/U ratios characterize the Dakota Formation. These probably reflect high lateral variability in clastic facies and interplay between mostly brackish and freshwater regimes of distributary channels, bays and marginal marine deposits, which would be expected to typify a delta complex. The relatively smooth, long-term cyclic pattern of the Th/U ratio in the marine sequence of the Upper Cretaceous is an excellent indicator of a broad transgression/regression couplet on an open marine shelf. The broad sine-wave feature conforms precisely with the outcrop interpretation of the Greenhorn Cycle as a classic example of a symmetric, third-order tectono-eustatic cycle (Glenister and Kauffman, 1985). The transgressive phase of the cycle started in the uppermost part of the Dakota Formation, continued through the Graneros Shale, and reached maximum development in the Greenhorn Limestone. The regressive hemicyclothem was initiated at the top of the Greenhorn and continued through the Fairport Chalk and Blue Hills Shale, to terminate in the Codell Sandstone. The overlying Fort Hays marks the renewal of a major marine transgression, that is marked by a distinctive drop in the Th/U ratio log value.

The development of portable spectrometers now makes it possible to verify the Th/U log ratios with the actual geology at the outcrop. So, for example, Zelt (1985) concluded that spectral data recorded at outcropping Cretaceous marine shales and chalks in Colorado, New Mexico, and Utah could be used to deduce the proximity of paleoshorelines and directions of sediment transport. Based on field measurements of the Lower Jurassic Cleveland ironstone in England, Myers (1989) showed relatively high Th/K ratios in the oolitic ironstones, and attributed the thorium-rich character to a lateritic weathering origin. Myers (1990) has also reported the integration of spectral measurements made from outcrop, core, and boreholes as important measures of clastic reservoir properties, with refinement of mineral and grain-size interpretations and recognition of shale permeability barriers and intervals of enhanced permeability. Slatt et al. (1991) provide insights on the reliability of interwell correlations in the subsurface based on experiences with detailed outcrop logging of lenticular and continuous turbidite sandstones by both standard logging truck and handheld devices.

APPLICATIONS OF GEOCHEMICAL LOGGING

The major change in geological applications of logs occurred with the introduction of nuclear logging tools to supplement the older electric logs. Geological conclusions based exclusively on the spontaneous potential and resistivity logs are restricted to simple characterizations of shale and pore volume and ambiguous assertions concerning grain-size variations or lithofacies types. Sadly, this approach will continue to be used in the more mature basins where these are the majority of logs available. Because the stratigraphic framework is so well known in these same basins, they are the natural locations to research evolving models of sequence stratigraphy. As a result, antique techniques of log-shape interpretation are mingled with modern sedimentological concepts drawn from computer modeling and seismic stratigraphy.

By contrast, nuclear measurements are rich in information concerning mineralogy and geochemistry. However, they also require more systematic analytic strategies than the often intuitive interpretation style used with older electric logs. Ellis (1990) gives an excellent review of the developments in nuclear logging that led up to the introduction of the geochemical (elemental analysis) logging (Chapman et al., 1987).

The Schlumberger Geochemical Logging Tool (GLT) incorporates measurements from three nuclear logging devices, combined on a single tool string, to estimate concentrations of 10 elements: potassium, thorium, uranium (from the natural gamma-ray spectrum); aluminum (by delayed neutron-activation analysis); and silicon, calcium, iron, sulfur, titanium and gadolinium (from the prompt-capture gamma-ray spectrum measured after a 14-Mev neutron burst). More extensive technical details are provided by Hertzog et al. (1987). The GLT string has been used both by industry and in scientific research logging (ODP, Cajon Pass well, the German KTB project). Commercial (and onshore) applications have focused on sedimentary sections (e.g., Wendlandt and Bhuyan, 1990), while much of the scientific research work, summarized by Anderson et al. (1990), has been directed to the analysis of igneous and metamorphic sections.

Figure 6 (also cover illustration) presents a spectacular example of a geochemical logging run during ODP Leg 115, in the Indian Ocean, and shown in the ODP Wireline Logging Manual (Borehole Research Group, 1990). This log provides a clear example of the geological evolution of a volcanic island that subsided and the subsequent development of a reef in Lower Eocene times. Cores from the volcanic sequence are mostly vesicular olivine-basalt flows with weathered zones, succeeded by plagioclase basalt. The basaltic composition is shown by the relatively high amounts of iron, aluminum, and silicon contents. The high aluminum spikes are coincident with weathered "soil" horizons between the flows. A thin calcarenite zone (interpreted from core as beach deposit) is succeeded by a distinctive titanium-rich basalt which marks the termination of volcanic activity. Core recovery in the overlying reef was only 5%, but this was sufficient to show an upward transition from grainstones to packstones and faunal changes that collectively mark a progressive deepening of water. The reef limestone is contrasted starkly with the volcanic basalt, by low iron, aluminum, and silicon contents, but high calcium content. The sulfur curve is of particular interest as it shows zones, possibly cyclic, of high sulfur concentration within the reef. The sulfur has been interpreted to reflect sulfate content associated with evaporite zones. Although no evaporites have been observed in the limited core available (5% recovery), the log may be a depth record of eustatic changes in sea level, with low stands marked by sulfur anomalies. Amplitude spectra from the sulfur trace show distinctive peaks at wavelengths of 25 and 50 ft, suggesting a cyclic pattern that may be related to the Eocene low stands of 36, 40, 42, 49, and 54 Ma of the Vail eustatic curve. As this example shows, great insights can be made into the geological history of a sequence from the raw elements recorded by the geochemical logs.

The major thrust of research connected with these logs has been aimed at the production of realistic mineral transforms. "Normative" minerals calculated from oxide analyses have been widely used in igneous petrology since the CIPW (Cross Iddings Pirsson Washington) norm was introduced by Cross et al. (1902). These normative minerals are contrasted with modal compositions, which are those mineral phases actually observed in the rock. The normative concept can be extended to sedimentary sequences in attempts to compute mineral assemblages, based on the 10 elements currently measured on the geochemical logging tool string (Herron, M.M., 1986). In calculating classical igneous norms, oxides are assigned to minerals in an allocation scheme that attempts to conform with their crystallization history. By contrast, elements from geochemical logs are transformed to normative or "chem" minerals by the inversion procedures discussed earlier. Herron (1988) studied terrigenous sands and shales in terms both of core and geochemical-log data and suggests that new methods of classification may be necessary. Strictly speaking, there will almost always be more minerals than elements to solve for them, so that the problem is always underdetermined. However, as Herron et al. (in press) notes, the overwhelming majority of sedimentary minerals can be numbered as 10: quartz, 4 clays, 3 feldspars, and 2 carbonates. In practice, reasonable compositional solutions can be generated using relatively small mineral subsets, provided that they have been identified correctly and that the compositions used are both fairly accurate and constant. In common with all new technologies, the approach is both exciting and controversial, but even modest successes should be of enormous benefit to a variety of geological studies.

Mineral solutions may be calculated by two alternative strategies. In the first, the average chemical compositions of minerals drawn from a large database are used as end-member responses and resolved by standard matrix inversion procedures. This result is normative and generic in the sense that it is based on a sample drawn from a universal mineral reference set and applied to a specific sequence where local mineral compositions may deviate from the global average. The result is hypothetical but has the particular advantage that comparisons can be made among a variety of locations and do not require expensive ancillary core measurements. In a second approach, the solution is calibrated to core data, where laboratory determinations of mineralogy and elemental geochemistry are analyzed by multiple regression techniques to determine local mineral compositions. This result is linked to petrography and so is philosophically closer to an estimated modal solution, rather than the more hypothetical normative model.

Several detailed studies have been made to assess the strengths and limitations of geochemical logging through exhaustive comparisons of borehole data and core elemental and mineralogical analyses. Examples include comparisons in the Conoco Research well, Ponca City, Oklahoma (Hertzog et al., 1987); discussion of the results from an Exxon research well that penetrated Upper Cretaceous siliciclastic rocks in Utah (Wendlandt and Bhuyan, 1990); and an assessment of data from three Shell wells in the Netherlands, Oman, and the U.S. (van den Oord, 1990). Figure 7 shows a typical comparative example of mineralogy for core and geochemical-log estimates (from van den Oord, 1990).

In general, the prognosis for this infant technology is quite good, particularly with regard to the relatively good match between in-situ borehole measurements and laboratory measurements from core. Teething problems are related primarily to the determination of the appropriate mineral-transform strategies to obtain useful results. Most authors working on the problem agree that local core calibration is a necessary step, rather than resorting to a generic normative solution. At the same time, it is recognized that the precise resolution of sedimentary mineral assemblages is inherently a complex problem and that in some respects, the technology is ahead of our understanding of the distribution of elements in sedimentary sequences. So, for example, Wendlandt and Bhuyan (1990) point out that some knowledge concerning the controls on distribution patterns of gadolinium and titanium would prove to be a very useful aid in future work.

In addition to the immediate display of lithofacies types, there are numerous potential applications of successful mineral transforms of geochemical logging data including: quantitative estimates of grain-size, cation-exchange capacity, and permeability, and using the minerals as surrogates for other petrophysical properties (Chapman et al., 1987). Although there may be some differences of opinion on how far these goals have been met, they certainly set forth a worthwhile agenda of research targets. Accurate clay-mineral typing and geochemical clues to diagenesis have immediate applications to improve reservoir engineering practice. Selley (1991) considers that the "third age of log analysis" has arrived with the advent of geochemical logging and that they are useful discriminators of a variety of diagenetic effects of cementation and solution, especially when used in conjunction with other logs. Cheshire (1991) advocated the use of computer-based diagenetic modeling as the appropriate methodology to deduce diagenesis from wireline logs, as the overall effects of diagenetic features are fairly subtle. Denham and Tieh (1989) suggest that measurements of thorium and uranium have potential in exploration studies as a means to delineate migration paths for fluids associated with the generation of hydrocarbons.

As mentioned earlier, an important application of the ODP geochemical logging work has been in igneous and metamorphic rocks. The models represent a significant advance on the interpretations associated with older logs run in geothermal wells where most techniques represented simple pattern recognitions based on older wireline logs (e.g., Keys, 1979). Normative solutions of mineral compositions have been computed in a variety of lithologies, using the inversion methods discussed earlier. Modal analyses from representative thin

sections, coupled with the use of best-fit and self-consistency criteria within the computer algorithm, provide important constraints to guide the solution to a feasible result, as described by Anderson et al. (1988). The strengths and drawbacks of a variety of mineral transform models for this purpose are discussed by Harvey et al. (1990).

STRATIGRAPHIC CYCLES AND MODELS OF CLIMATE CHANGE

Over the years, a considerable amount of research has been applied to the concept that many stratigraphic successions are composed of cyclic repetitions of simple lithological sequences. Duff et al. (1967) summarized the extensive literature on stratigraphic cycles, most of which was based on outcrop studies. Controversy centered on whether cyclic characteristics were real or illusory, and, if real, whether the cycles were caused by tectonic or climatic changes.

In the last decade, mounting evidence suggests that oscillations in global climate are responses to orbital perturbations of the earth. These are cyclic in time, and distinctive periods can be assigned to the orbital parameters of eccentricity (95,000 years), obliquity (41,000 years), and precession (19,000 -23,000 years). The periods are collectively known as "Milankovitch cycles" and provide a key to the recognition of climate events in the remote geological past, such as glaciation and sea-level changes and their effects on the stratigraphic record. The model can also be linked with the Vail coastal-onlap curves developed from seismic stratigraphy and offers an exciting potential to assign specific dates to stratigraphic units. The recent sedimentary successions of the deep-sea floor have proved to be a useful testing ground for verification of whether these cyclic climatic changes can be detected by changes in lithological properties. These effects should be shown by changes in clay mineralogy and clay content, grain-size, types and abundances of planktonic fossils, caused by cycles of temperature and aridity/humidity driven by changes in solar energy.

Wireline logs have proved to be excellent records of many of these phenomena. They have the advantage over core of being both continuous and complete. A major disadvantage (at this time) is that they are primarily depth rather than time records of petrophysical changes. Consequently, conversions to a chronology scale must be made using fossil-dating from cores. Also, the vertical resolution of some logging tools may be too coarse to detect higher frequency cycles, particularly at sites of slow sedimentation rates. Worthington (1990) examined this problem in detail by modeling cyclic sequences with differing sedimentation rates and convolving these with the response functions of common tools. In some cases, the poor vertical resolution of tools, such as the induction resistivity, may cause them to be ineffective in resolving the cycles attributed to precession and obliquity. However, the trend to running tools with higher resolutions has improved this situation. Maltezou and Anderson (1991) described an example of the recognition of Milankovitch cycles from resistivity logs.

Detection of potential cycles is made by Fourier transformation of the logs to amplitude spectra and conversion of depth to time scales deduced from fossil evidence. Cyclic characters in logs appear to be strongly linked with changes in clay content and porosity. These were registered on resistivity and sonic logging tools in deep-sea sediments of Baffin Bay and the Labrador Sea on ODP Leg 105 as described by Jarrard and Arthur (1989). The fluctuations in clay and porosity appear to reflect changes in strength of ocean bottom currents. At times of weaker currents, high-porosity clay-rich sediments accumulated; stronger currents may have resulted in transport and deposition of greater volumes of quartz and other coarse-grained minerals. Ultimately, the waxing and waning of the bottom currents is controlled by warming and cooling cycles at the ocean surface, which in turn reflect migration of upwelling zones.

The use of the geochemical-logging tool string has resulted both in an improvement of vertical resolution and the recording of elements that can be related more directly to geological properties, particularly mineralogy. The power spectrum of the calcium/silicon ratio at the Labrador Sea site (see Figure 8) shows peak developments that can be attributed to all three Milankovitch cycle types (Jarrard and Arthur, 1989). In the south Atlantic (ODP Leg 114), Mwenifumbo and Blangy (1991) derive amplitude spectra of geochemical logs using a moving depth window. They were able to identify cycles in the calcium log that were clearly out of phase with silicon and hydrogen variation. These characteristics are easily attributed to fluctuations of diatom-rich porous sediments as changes in climatic temperature-moved upwelling zones north and south.

Estimates of sedimentation rate are currently keyed to biostratigraphic control and these will obviously vary with time. Molinie and Ogg (1990) considered that the time periods of the Milankovitch cycles is sufficiently well established that they could be used to deduce sedimentation rates as a continuous function from sliding-window spectral analysis. They applied this technique to a gamma-ray log of Jurassic-Cretaceous radiolarian mudstones from the equatorial Pacific and were able to derive reasonable sedimentation rates and detect a major discontinuity.

The emergence of the concepts of Milankovitch cycles and sequence stratigraphy has also encouraged spectral analyses and cyclic interpretations of logs at sites onshore. Examples of recent studies along these lines include the analysis of North Carolinian Upper Triassic lacustrine beds using gamma-ray logs (Hu et al., 1990), Argentinian Lower Cretaceous highstand deposits using a combination of resistivity, sonic and gamma-ray logs (Spalletti et al., 1990), English Upper Jurassic sequences using filtered sonic and gamma-ray logs (Melnyk, 1990), and the Texas Permian, using gamma-ray logs (Borer and Harris, 1991). The sequence stratigraphic link between well logs, outcrops and cores is also described in the popular and profusely illustrated book by Wagoner et al. (1990). Vail and Wornardt (1991) discussed the integration of well logs with seismic stratigraphy in exploration and development.

Much fine-tuning remains to be done both in tracking the changes in periodicities back in geologic time, and the choice of appropriate spectral techniques and logging measurements. However, links with the Vail coastal-onlap curve offer great potential for studies of basin history and a more refined chronology of stratigraphic events. In a parallel development, it appears that practical magnetostratigraphy logs may soon be available even from sedimentary sequences with low magnetization. Experimental logging on Leg 102 of the ODP demonstrated that both the intensity and direction of paleomagnetic polarization could be measured reliably by downhole tools in strongly magnetized volcanic rocks (Leg 102 Scientific Party, 1985). Further work by Tabbagh et al. (1990) in the Couy (France) boreholes produced magnetic-logging results from a sedimentary sequence that showed a good match with core measurements. If natural remnant magnetization polarity can be logged in a practical and reliable manner, there will be important and exciting applications. A fundamental limitation of correlations based on conventional wireline logs is that they are lithostratigraphic whereas magnetostratigraphy logs would be keyed directly to time. Where most correlation ties cross time-lines, because sedimentary layers are diachronous products of transgression and regression, absolute time sequences established from logs of magnetostratigraphy could give revolutionary insights into development of sedimentary basins and characterizations of reservoir structure.

Summary

We have attempted to summarize the recent major trends and developments in the field of geological applications of wireline logs. The field is expanding rapidly and comprehensive coverage is difficult to achieve. This growth is an excellent sign that exciting future developments can be expected and that wireline-log analysis will make increasingly significant contributions to a wide variety of geological studies.

Home