Explanation in Earth Science

In §2 we saw that earth scientists have at least two aims, namely to describe and to explain the (history of) inanimate processes on the earth. Indeed, it seems plausible that all sciences aim at explanation and understanding, over and above mere description, of the phenomena in their domain (see de Regt and Dieks 2005).

4.1 Do historical narratives explain?

The issue of whether historical narratives can explain, and if so, how they do, is one on which opinions greatly diverge. Some authors claim that narratives explain by integrating an event into a bigger picture (Hull 1989). A narrative thus conveys some greater holistic insight, particularly in history and the social sciences. But others sneeringly call historical sciences varieties of “stamp collecting.” Still others see a logical pattern in historical explanation that is fundamentally different from deductive-nomological and inductive-statistical explanation schemes. For human historical narratives it has been argued that the plot, or argumentation structure, of the story (e.g., perspective, ordering in time) conveys an explanation of the events in terms of their necessary (but not sufficient) conditions and their rela­tions (see von Wright 1971).

We will argue that there is more to earth-scientific narratives than merely description. Like evolutionary biology, earth science is partly histori­ography: it reconstructs and tries to explain past events. The purpose of historical sciences is to provide a correct narrative of the sequence of past events and an account of the causal forces and antecedent conditions that led to that sequence.

The historical narrative approach provides two kinds of explanations relevant for earth science: robust-process explanations and actual-sequence explanations (Sterelny 1996, 195). While the latter specifies the (causal) chain of events, the former focuses on underlying (robust) causes of the phe­nomenon. Both kinds of explanation are important in earth science, even for the same event, because they convey distinct breeds of information. Consider two explanations of a catastrophic landslide or mudflow. We may model an observed flow in high detail with physical models. In retrospect, the initiation of the flow that happened at that precise moment in time and at that location can be explained by, for example, heavy rainstorms. This actual-sequence explanation leaves out that the flow was waiting to happen because a certain amount of mud or rubble was on the verge of slope failure. If this specific heavy rainstorm had not initiated the flow, it would very likely have been initiated by another rainstorm, or the spring melting of accumulated snow, or an earthquake. So earth scientists not only model such flows with empirical or physical models but also map areas prone to mudflows as a part of the robust-process explanation for these flows.

In order to be genuinely explanatory, narratives in earth science need not be reduced to physical causal-law explanations. The practice of earth science demonstrates that a domain can be ontologically dependent on another deeper (i.e., physical) domain while at the same time being explanatorily autonomous. In historical narratives, an event is not explained by subsuming it under a generalization.

4.2 Underdetermination problems in earth science

While the historical narratives provided by earth scientists can thus be genuinely explanatory, they face a major problem, which we will discuss now, namely the fact that they are usually underdetermined by the available evidence. Philosophers distinguish between weak (practical) underdeter­mination and strong ( logical) underdetermination. In the former case, there is insufficient available evidence to choose one theory over its rivals, while in the latter case theory choice remains impossible no matter how much evidence is gathered. Contrary to physicists, practicing earth scientists very often face situations in which theories are underdetermined by the available evidence. In fact, it is hard to find papers that do not contain at least a paragraph on the way underdetermination was dealt with in practice (although it is usually not explicitly referred to as “underdetermination”). Typical examples of underdetermination problems are the following (see also Turner 2005).

First, the time scale involved in shaping the earth is orders of mag­nitude larger than the life of human observers or even written history. It is therefore problematic to detect and observe the long-term effects of slow processes that might be extrapolated to the past.

Second, many earth-scientific processes and phenomena cannot (yet) directly or even indirectly be observed. Sometimes a phenomenon eludes direct detection by instruments, for instance deep-mantle convection within the earth and other planets. Also, landforms and sediment deposits (with all the clues to processes and events in the past) often have been obliterated by erosion, mountain building, or flooding. An additional practical problem is that current techniques often disturb the observed processes.

Third, many processes are intrinsically random or chaotic and may be very sensitive to initial conditions. A veritable reconstruction of past events from the geologic record becomes extremely arduous because many events and phenomena are so complex that, in theory, an infinite number of possible laws and initial conditions could be involved. This problem has especially received attention in hydrology where it was named “equifinality.” Noise and chaos lead to a certain uniqueness of geomorphological phe­nomena: duplications of events are seldom found and the probability that the long-term river channel development in a particular river delta would take the same course when repeated is near zero.

While complete causal (deductive-nomological) explanations would be highly underdetermined, a narrative explanation does not require an exhaustively detailed set of observations and initial conditions. In many earth-scientific studies, initial and boundary conditions are implicitly (and without the details) given in the reference to the length and time scale of the phenomenon under study or the time period or study area, and in many other studies it is apparent from the context, such as the name of the scientific journal. For example, the Rhine-Meuse delta in the Netherlands has been studied by Henk Berendsen and Esther Stouthamer (2001) for the Holocene time period, which is the past 11,760 years (in 2010).

This approach is not only practical but often also defensible (Schumm and Lichty 1965). First of all, there is a correlation between the relevant length scale and the relevant time scale of most phenomena (see Figure 9.2), so it is not necessary to study the evolution of a large object on a very short time scale or a small object on a very large time scale. For instance, current ripples, which have lengths of about 0.2 m, are formed and destroyed in minutes to hours, so to study ripples over decades is probably useless. Mountain ranges of thousands of kilometers long, on the other hand, are formed in millions of years. In other words, it takes a much longer time to build or break down large things than to build or break down small things.

The reason is that the range of energy available is between two close limits (compared to the extreme energy levels familiar to astronomers): a lower limit necessary to exceed thresholds such as friction or entraining boulders by stream flow, and an upper limit given by the maximum energy available on earth. Consequently, no significant change occurs to the whole mountain range in a period of hours, even though one plot on one slope may have changed significantly by a flood or by mass wasting. Therefore, it is commonly not useful to study large phenomena over very short periods or very small phenomena over large time periods. Notable exceptions are relatively unique events such as the impacts of large meteorites or earthquakes.

4.3 Underdetermination and explanation

How precisely do underdetermination problems lead to the impossibility of finding complete causal explanations and to the use of narrative explanations in earth science? Consider a typical earth-scientific research project that aims to construct a spatiotemporal description and an explan­ation of the course of the river Rhine in the past 10,000 years (see Berendsen and Stouthamer 2001).

The predominant methodology of historical sciences is that of formu­lating various competing explanations (causes) of present phenomena (effects) and discriminating between them by searching for “smoking guns.” Consider a candidate causal explanation for the presently available evidence, for example, regarding the courses of the river Rhine through time. This would take the form of a description of the initial conditions, say, 10,000 years ago, plus a set of causal laws that govern the dynamics of the system. The laws are, as we argued above, derived from physics and chemistry (let us suppose they are given and deterministic). What really matters is the choice of the initial conditions; these are actually referred to as “the cause.” But these conditions are not, and indeed can never be, specified completely - they merely give an incomplete description of a part of the universe (and this part is never a closed system). In practice, therefore, explan­ations in earth science are a combination of abductive narratives and causal explanations. The narratives carry most of the explanatory power, for example of how the shifting of the Rhine branches over the delta plane depended on the sea-level rise and other factors.

The causal explanation parts are often applications of computer models to certain aspects of the delta development. Computer modeling based on laws of physics or chemistry is becoming an activity as central in earth science as experimenting and observing. However, the laws in the causal explanations are not simply given and deterministic as assumed above. For instance, for many phenomena it is not prima facie clear which physical laws apply. Moreover, the relatively simple laws of physics can seldom be applied directly to the initial conditions to check whether they (deductively) explain the observations under scrutiny. For instance, conservation of momentum and mass are simple physical laws that apply to fluid flow. These simple laws are the basic components of the Navier-Stokes equations that govern fluid flow, which cannot be solved analytically. Therefore these equations are implemented in so-called “mathematical” or “physical com­puter models.” The equations have to be simplified and discrete time steps and grid cells must be used to model the flow in space and time. These simplifications bring a host of necessary numerical techniques to ensure conservation laws and to minimize numerical (computer-intrinsic) error propagation. When initial or boundary conditions are specified for this model, certain laboratory or field conditions can be simulated and compared to the observations. But one can never be certain that a mismatch between model results and observations is not due to the simplifications and numerical techniques. So this does not solve the problems of underdeter­mination, for as Oreskes, Shrader-Frechette, and Belitz explain:

Verification and validation of numerical models of natural systems is impos­sible. This is because natural systems are never closed and because model results are always non-unique.... The primary value of models is heuristic. (1994, 641)

Systems that are “never closed” exhibit the underdetermination of initial and boundary conditions and the “non-uniqueness” refers to the equifinality problem. So models cannot be used to hindcast or forecast the systems studied in earth science, unless they are calibrated on data from the past. But in the latter case, one can hardly speak of a true “physical” model because the data carry a part of the explanatory weight.

But there are still great benefits to be gained by modeling. First, the human mind cannot comprehend the results of complex sets of equations in space and time under certain initial conditions, but a model pro­vides comprehensive results. Second, a scientist can manipulate a model in ways that are impossible in nature or even in the laboratory. Various scenarios can be studied and “what-if questions” surveyed under the assumptions of the model, even if the modeled scenario is not what really happened according to the observations. Thus, these scenarios play an important role in robust-process explanations, both to extend the explanation and as counterfactuals in view of the observations. Third, a set of model runs may indicate (though not always prove) whether certain hypotheses are possible at all and whether they conflict with physical laws and mathematical constraints. Modeling in this sense is used as a test for the narrative robust-process explanations. Fourth, the comparison of various different models for the same phenomenon may indicate the robustness of the models: if the modeled phenomenon is similar in the different models, then it is to some extent independent of the model schematizations.

The above considerations make it clear that both historical narratives and causal explanations are needed in earth science. The two kinds of explan­ation are in fact complementary: causal explanations can be employed on a small spatiotemporal scale and may give further hints for narra­tive explanations on a much larger scale, such as the scenario modeling approach outlined above, whereas such broadly construed narrative explan­ations may in turn provide more insight into specific causal processes. The border line between these two types of explanation is often unclear because they gradually merge into one another.

4.4 Inference to the best explanation

Until now we have outlined the problems that earth scientists face. We do not want to suggest that these problems cannot be solved, however. In practice earth scientists do come up with explanations, which they regard as more or less corroborated. But how do the scientists arrive at these explanations? An important occupation of earth scientists is formulating hypotheses about possible causes for the phenomena observed. It appears that most earth-scientific explanations are the result of abductive inference, and that earth scientists typically rely on “inference to the best explan­ation” (Lipton 2004), supported by deductive causal explanation (for instance based on computer modeling). Furthermore, abduction in earth science is extended by a method already described by Thomas Chamberlin (1890) as “the method of multiple working hypotheses.” A number of hypotheses are developed which potentially explain the observations. By contrasting and testing a number of (incompatible) hypotheses, a biased attempt at confirmation is prevented. The hypotheses can be processes that are known from observations, or “outrageous hypotheses” of processes that supposedly occurred in past times but are no longer active or that seem to be in conflict with the laws of physics or chemistry. Wegener's hypothesis of “continental drift” is the most famous example of such an outrageous hypothesis: initially rejected as absurd, it is now accepted and sup­ported by plate tectonics. The hypotheses are used to predict testable consequences (deduction), preferably for a wide range of different loca­tions, with the use of different kinds of instruments and often with computer models. When these various data and model scenarios all point to the same (underlying) explanation or common cause, earth scientists accept this explanation as (tentatively) true.

For instance, the occurrence of about 23 glacial-interglacial couplets in the past 2.4 million years (“ice ages”) is best explained by a combination of a unique setting of continents and astronomical influences on the global climate. There are cyclic variations in the inclination and direction of the earth's axis and the obliquity of the orbit, of which the periods correspond with those of the glacial-interglacial cycles. These variations cause solar irradiation variations at higher latitudes, which leads to cooling. This, in turn, causes an increase of snow cover which leads to a larger reflection of sunlight and hence causes more cooling. But this orbital forcing has been the case for much of the earth's late history during which glaciation was not observed, so something more is needed. About 2.4 million years ago, Antarctica was detached from the South American continent. Cir­cumpolar currents developed instead of currents circulating between the equator and the pole, which led to thermal isolation and, given the polar position, cooling of Antarctica. The result was increasing ice coverage and hence larger albedo, and consequently the whole planet cooled a few degrees. This cooling, added to the astronomically induced fluctuations, was enough to trigger the glaciation of higher latitudes.

While this set of hypotheses is not without problems, an alternative set of hypotheses that explains these phenomena equally well is not easily conceived. For example, the fluctuation of solar activity due to its stellar dynamics (“solar forcing” on the earth's climate) is a competitive explanation for global temperature oscillations, of which effects have been demonstrated in the geological record. But there is no stellar theory that predicts the 23 rhythmic cycles, whereas the orbital forcing (systematic variations in the earth's orbit) does.

Most earth scientists believe that the former set of (triangulated) hypotheses is the best explanation for the observations, while the latter set of hypotheses is also true but only modifies the effects of orbital forcing. This is an example of inference to the best explanation: the observed ice ages are surprising, but there are hypotheses explaining part of the evidence that would make the observations not surprising. So the set of hypotheses is tentatively accepted as the explanation, and further confronted with new evidence.

Inference to the best explanation is a form of abduction, and, like in­duction, it has its limitations (Lipton 2004). Foremost, abduction is a method to select hypotheses rather than conclusive explanations. We probably will never know whether the best explanation is also the one and only true explan­ation. In fact, the uniqueness of events and phenomena in the geologic past often requires hypotheses that are at first sight outrageous in view of our present experiences (Davis 1926, Baker 1996). For example, the hypotheses that certain landforms are caused by ice-age glaciers, and that whole con­tinents are drifting and colliding, were once considered outrageous. Over time, however, they were supported with further evidence and robust-process explanations were formulated by extending the hypotheses and by adding causal explanation sketches. The latter are typically developed by using mathematico-physical models, but, as was argued in §4.3, these models are also troubled by underdetermination problems. If the narratives survive tests against an increasing body of evidence, the hypotheses are more generally accepted by consensus as the best explanations.

5.