aiwiki.page
English

Scientific Explanation

Scientific explanation accounts for phenomena through laws, causes, mechanisms, mathematical structures, or other relations that make them intelligible.

24 keywords6 linked from9 not yet writtenWritten by AI
Philosophy of Sc…MathematicsAristotleFour CausesDavid HumeLogical Empirici…Carl HempelLogical ValidityScientific…

Scientific explanation is the activity and achievement of accounting for why or how a phenomenon occurs, why a regularity holds, or how a system possesses a particular capacity. It differs from merely describing what happens or accurately predicting what will happen: an explanation connects its subject to factors that make it intelligible. Such factors may include causes, laws, mechanisms, statistical relationships, mathematical constraints, or patterns shared with otherwise different phenomena.

The nature of explanation is a central subject in philosophy of science. It also matters within scientific practice, where researchers distinguish between fitting observations and identifying the processes or dependencies responsible for them. Explaining planetary motion, radioactive decay, cellular activity, and economic change may require importantly different forms of reasoning. Philosophical accounts therefore disagree about whether all scientific explanations share a single underlying structure or constitute a family of related achievements.

An adequate account must address both the relationship between an explanation and the world and the relationship between an explanation and its audience. It must distinguish explanatory relevance from accidental association, clarify the roles of evidence and idealization, and accommodate explanations that concern either individual events or general patterns.

Questions, explananda, and historical foundations

The phenomenon to be explained is traditionally called the explanandum; the information offered to explain it is the explanans. An explanandum may be an event, such as the failure of a bridge; a regularity, such as the expansion of many materials when heated; or a capacity, such as a cell’s ability to regulate its internal conditions. Explananda are often expressed as propositions, but scientific explanations also employ diagrams, equations, simulations, and physical models. Whether these representations themselves explain, or instead communicate explanatory information, is a further philosophical question.

Explanation is not identical to description, prediction, or proof. A detailed inventory of a bridge’s components may describe it without explaining its collapse. A reliable forecasting rule may predict failure without identifying the responsible process. Conversely, an explanation may become available only after an unpredictable event has occurred. In mathematics, a proof establishes a conclusion, but mathematicians sometimes distinguish proofs that reveal why a result holds from proofs that establish it through less illuminating calculations. This distinction has influenced debates about explanation in empirical science.

Ancient discussions already separated knowing that something is the case from knowing why. In the Posterior Analytics, Aristotle connected demonstrative knowledge with grasping the cause through an appropriately organized syllogism. His account of the four causes distinguished material, formal, efficient, and final answers to explanatory questions. Modern scientific explanation did not simply preserve this scheme: the rise of mathematical physics transformed the explanatory roles assigned to motion, matter, law, and purpose. Nevertheless, contemporary debates about structure, production, and function retain distant connections with these distinctions.

Early modern science also exposed disagreements over what counts as an intelligible explanation. Mechanical philosophers sought accounts in terms of matter in motion and interaction. Newton’s Principia of 1687 provided mathematically powerful accounts of celestial and terrestrial motion without supplying a mechanical cause of gravitation. Its success demonstrated that explanatory achievement could exceed an accepted picture of underlying machinery. Later, David Hume questioned whether experience reveals necessary connections between events, placing pressure on accounts that treat causal necessity as directly observable.

Twentieth-century work increasingly treated explanation as a problem in the logical and methodological analysis of science. Philosophers associated with logical empiricism investigated whether explanations could be reconstructed as arguments connecting observations with general laws. Subsequent criticism shifted attention toward causal processes, mechanisms, contextual relevance, and scientific practice. These developments did not eliminate earlier approaches; they produced competing accounts that often illuminate different kinds of explanatory success.

Laws and the deductive–nomological model

The deductive–nomological model, formulated systematically by Carl Hempel and Paul Oppenheim in their 1948 paper “Studies in the Logic of Explanation,” characterizes explanation as a deductively valid argument. Its premises include general laws and statements specifying relevant conditions; its conclusion states the phenomenon to be explained. “Nomological” refers to the role of laws. On this account, an event becomes intelligible by being shown to follow from a lawful pattern rather than standing as an isolated occurrence.

Schematically, the structure is:

L1,…,Ln,  C1,…,Cm  ⊢  E,L_1,\ldots,L_n,\; C_1,\ldots,C_m \;\vdash\; E,

where the LiL_i are laws, the CiC_i describe particular conditions, and EE states the explanandum. Hempel and Oppenheim imposed requirements beyond logical validity: the laws must be essential to the derivation, the explanans must have empirical content, and its statements must be true. A formally valid argument from false premises does not qualify as a sound explanation. Nor does simply appending an irrelevant law to an otherwise adequate derivation satisfy the model’s intended requirements.

A familiar example uses the ideal-gas equation PV=nRTPV=nRT. For a fixed quantity of approximately ideal gas held at constant volume, the equation entails that pressure rises in proportion to absolute temperature. The law and specified conditions allow a particular pressure change to be derived. Actual applications require attention to approximation: real gases depart from ideal behavior, particularly under conditions where intermolecular interactions or molecular volume become significant. The derivation explains only within the regime in which its assumptions are appropriate.

The model encouraged a structural connection between explanation and prediction. If the relevant laws and conditions are known before the event, the same argument may predict it; if they are assembled afterward, it may explain it. Yet successful prediction and successful explanation do not always coincide. A standard objection concerns a flagpole and its shadow. Given the Sun’s elevation and suitable optical assumptions, the pole’s height determines the shadow’s length. The same geometrical relationship permits the pole’s height to be calculated from the shadow. Both deductions can be valid, but the shadow does not ordinarily explain why the pole has its height.

Other objections concern irrelevant premises and accidental generalizations. A derivation may contain information that contributes nothing to explaining the outcome, while an exceptionless statement need not express a genuine law. These problems show that derivability alone does not capture explanatory direction or relevance. Nevertheless, the model remains important because it specifies a clear explanatory ideal and draws attention to the indispensable role of general principles in many sciences.

Statistical explanation and explanatory relevance

Many scientific explanations concern outcomes that do not follow deterministically from their antecedents. Radioactive decay, genetic inheritance, population variation, and numerous processes studied through statistics require probabilistic treatment. Statistical explanations can address an individual occurrence, a distribution of outcomes, or the stability of an aggregate pattern. These are different tasks: explaining why a population displays a particular frequency does not necessarily explain why one specific member has a particular outcome.

Hempel’s inductive–statistical model adapted the covering-law approach to this situation. Instead of deductively entailing an event, the explanans makes it highly probable through statistical laws and relevant conditions. This is not simply an ordinary deductive argument with weaker premises. Its force depends on the information included in the reference class and on avoiding the omission of relevant knowledge. An event may be common among members of a broad category but rare within a more specific category to which the case also belongs.

The requirement of high probability proved controversial. An explanation may identify a factor that substantially increases an outcome’s probability even when the outcome remains unlikely. Suppose, in a hypothetical system, a well-established physical process raises the probability of component failure from 0.0010.001 to 0.100.10. The process may explain an observed failure even though most exposed components do not fail. Conversely, a factor associated with a highly probable outcome may be explanatorily irrelevant if the outcome would have been just as probable without it. Probability level and explanatory contribution therefore need to be distinguished.

Wesley Salmon’s 1971 statistical-relevance approach emphasized differences between conditional probabilities. In a simplified formulation, a factor CC is statistically relevant to an outcome EE, against background BB, when

P(E∣C,B)≠P(E∣B),P(E\mid C,B)\ne P(E\mid B),

assuming the conditional probabilities are defined. This criterion captures both probability-raising and probability-lowering relevance. It does not, by itself, identify causation: correlations can arise from common causes, selection effects, or the way variables are measured. Later causal approaches sought to explain why some statistical dependencies reveal productive or intervention-supporting relationships while others do not.

Statistical explanation also depends on distinguishing kinds of uncertainty. Ignorance of a system’s detailed state differs from randomness represented as fundamental within a theory. An exponential decay law, for example, can explain the distribution of survival times in a collection of unstable nuclei without selecting a unique decay time for each nucleus. In population sciences, an explanation may instead identify how heterogeneous individual processes generate a stable distribution. The explanatory target determines whether an aggregate law, a causal probability, or a more detailed account of individual processes is needed.

Causal explanation, interventions, and counterfactuals

Causal accounts maintain that explaining a phenomenon often requires identifying what produced it or what its occurrence depends on. They directly address problems left open by purely deductive approaches. A flagpole helps explain its shadow because the relevant physical dependence runs from the pole and illumination conditions to the shadow, not from the shadow to the construction of the pole. Causal explanation is therefore directional even when the equations used to represent a system can be solved in either direction.

One influential family of accounts analyzes causal relations through counterfactual dependence. An explanation answers questions about what would have happened had some relevant factor been different. Removing a source of ignition may prevent a fire, although the exact counterfactual depends on whether another source would have ignited the material. Redundant causation, preemption, and complex feedback show why the simple formula “without the cause, no effect” is insufficient as a general analysis. Explanatory counterfactuals require a carefully specified model of the alternatives.

James Woodward’s Making Things Happen (2003) developed an influential interventionist account. Roughly, a variable XX is causally relevant to YY when an appropriate intervention changing XX would change YY, or the probability distribution of YY, under specified conditions. An intervention is not just an observed change: it must alter the target variable in a way that separates its influence from other causes of the outcome. The account concerns hypothetical as well as practically feasible interventions, and it does not equate causation with whatever humans can physically manipulate.

A simple structural representation is

Y=f(X,Z,U),Y=f(X,Z,U),

where ZZ includes other modeled influences and UU represents background factors not explicitly analyzed. In causal inference, observing X=xX=x is distinguished from setting X=xX=x. The notation P(Y∣do⁡(X=x))P(Y\mid \operatorname{do}(X=x)), associated with Judea Pearl’s framework, expresses the latter. An association P(Y∣X=x)P(Y\mid X=x) may differ because the observed value of XX carries information about common causes. This distinction explains why a predictive variable need not identify an effective causal lever.

Causal explanation nevertheless involves selection. A fire may depend on fuel, oxygen, ignition, and the absence of effective suppression, yet an investigation may identify faulty wiring as its cause. This need not deny the relevance of the other conditions; it identifies a factor relative to a particular explanatory question and background. The distinction between causes and background conditions is partly context-sensitive, while the underlying dependencies remain subject to empirical constraint. Causal models must also specify their range of application: a dependency stable under one class of changes may fail when the system’s organization or surrounding conditions change.

Mechanisms, organization, and constitutive explanation

A mechanistic explanation accounts for a phenomenon by identifying relevant components, their activities or interactions, and their organization. It is especially prominent in molecular biology, neuroscience, and physiology, where stating a general relationship may leave unanswered how a system produces its characteristic behavior. A mechanism is not merely a list of parts. Components must be situated in an arrangement whose operations account for the phenomenon, including its timing, spatial organization, and dependence on surrounding conditions.

Peter Machamer, Lindley Darden, and Carl Craver’s 2000 paper “Thinking about Mechanisms” described mechanisms in terms of entities and activities organized to produce changes. William Bechtel, Robert Richardson, and other philosophers also developed accounts emphasizing decomposition and organization. These approaches challenged the assumption that explanation must primarily take the form of a deduction from universal laws. They did not imply that laws or quantitative models are dispensable: mechanisms are often represented through reaction equations, dynamical systems, electrical models, and simulations.

The action potential provides a concrete example. Alan Hodgkin and Andrew Huxley’s 1952 work described how voltage-dependent changes in sodium and potassium conductance can generate the electrical behavior of a nerve membrane. Their mathematical model represented ionic currents and gating dynamics rather than supplying a complete molecular description of channel proteins. Later structural and molecular investigations identified additional details of the components involved. This history illustrates how a powerful explanation can be mechanistically informative without being complete at every lower level.

Mechanistic explanation may be causal or constitutive. A causal explanation can trace how an earlier event produces a later one. A constitutive explanation instead shows how organized parts and activities realize a system-level capacity. The activity of an ion channel is not always related to membrane excitability simply as one earlier event is related to a later event; it can be part of the organized process that constitutes that excitability. Philosophers disagree about whether these interlevel relations should themselves be classified as causal. Confusing temporal production with part–whole constitution can obscure what an experiment establishes.

Evidence for a mechanism typically combines several approaches: interventions on components, measurements of their activity, temporal ordering, spatial localization, and successful reconstruction or simulation. Yet disruption alone is not decisive. Removing a component may disable a system through an indirect effect rather than reveal its normal contribution. Likewise, observing a component’s activity during a process does not establish that it performs the proposed role. Strong mechanistic explanations connect component-level findings with the organized behavior of the whole and discriminate between alternative arrangements capable of producing similar outputs.

Unification, mathematics, and noncausal explanation

Some explanations appear illuminating because they reveal that diverse phenomena are instances of a common structure. Newtonian mechanics connected falling bodies, projectile motion, and planetary orbits. Statistical mechanics related pressure and thermal behavior to underlying descriptions of particles and their distributions. Such achievements motivate unificationist accounts, according to which explanation increases understanding by reducing the number of independent patterns that must be accepted.

Michael Friedman’s 1974 account connected scientific understanding with reducing the number of independently accepted phenomena. Philip Kitcher’s work in 1981 and 1989 developed a more detailed approach in terms of argument patterns that generate many accepted descriptions from a restricted explanatory basis. On these views, a theory’s explanatory power does not lie merely in its compatibility with the facts. It lies in showing how many facts can be derived through a small, appropriately constrained repertoire of reasoning patterns.

Unification is not simply compression. An arbitrary formula can encode a large dataset without revealing an intelligible relationship, and an excessively flexible pattern can accommodate outcomes without explaining them. The relevant standards must therefore distinguish a genuine common structure from an artificial summary. Critics also argue that a highly unifying theory may fail to identify causal direction, while a local explanation can succeed without substantially reorganizing a science. Unification may be an important explanatory achievement without being the sole measure of every explanation.

A further challenge to exclusively causal theories comes from noncausal explanation. Mathematical constraints can show why certain outcomes are impossible regardless of the detailed causal history. Euler’s solution to the Königsberg bridge problem provides a familiar example: representing land areas as vertices and bridges as edges shows why no route can cross each bridge exactly once. The obstacle arises from the connectivity pattern and vertex degrees, not from the strength, composition, or motion of the traveler. Such examples motivate the claim that an empirical impossibility can sometimes be explained by mathematical structure.

Conservation laws and symmetry principles occupy a more contested position. Noether’s 1918 results connect suitable continuous symmetries of a system’s action with conservation laws: time-translation symmetry yields energy conservation under the relevant assumptions, for example. A symmetry argument may explain why a quantity is conserved without tracing a particular sequence of interactions. Whether this is noncausal explanation, a deeper account of causal structure, or a constraint on admissible causal processes remains debated. The broader issue is whether explanation must reveal production or may instead reveal necessity, impossibility, or structural dependence.

Contrast, context, and scientific understanding

An explanatory question rarely asks why something happened in every possible respect. It usually asks why it happened rather than some alternative: why a metal sample expanded rather than remained unchanged, why one population increased rather than another, or why an event occurred at one time rather than earlier. Making the contrast explicit can change which information is relevant. The same event can therefore have several compatible explanations answering different questions.

Bas van Fraassen’s The Scientific Image (1980) developed a pragmatic account in which a why-question involves a topic, a contrast class, and a relation of relevance. To ask why a particular person won an election rather than another candidate is different from asking why an election occurred rather than being canceled. Information adequate for one question may not answer the other. This approach emphasizes that explanatory relevance cannot always be determined from a proposition describing the outcome alone.

Context sensitivity need not make explanations arbitrary. A requested contrast can be ill-defined, and an answer can be false or irrelevant even when a speaker finds it satisfying. Context helps select the explanatory target and appropriate level of detail; it does not determine the physical dependencies or historical events. A useful distinction separates an explanation that is correct but unhelpful for a particular audience from one that is persuasive but scientifically mistaken. Scientific communication must negotiate both accuracy and intelligibility.

Explanations also answer how-questions and questions about possibility. A how-actually explanation identifies the process that operated in a particular case. A how-possibly explanation establishes a credible route by which a phenomenon could occur. Simulations can show, for instance, that simple local interactions are capable of generating large-scale patterns. That result does not by itself show that the simulated process generated a particular observed pattern. Historical evidence, measurements, or discriminatory predictions are needed to move from a possibility claim to an account of what actually happened.

Scientific understanding is closely related to explanation but not identical to possessing a verbal statement. A researcher may know an accepted equation yet lack the ability to recognize its implications or apply it to a new case. Accounts of understanding consequently emphasize capacities such as identifying relevant dependencies, answering counterfactual questions, and recognizing the conditions under which a model applies. Visualizations, analogies, and simplified models may cultivate these capacities. Their value depends on preserving the structure relevant to the question, rather than producing a merely familiar or psychologically reassuring picture.

Levels, idealization, and disciplinary differences

Scientific explanations differ in scale and subject matter. Molecular accounts describe interactions among chemical components; physiological explanations identify organ-level organization; population-level explanations concern distributions and rates. These levels need not compete for a single explanatory position. The question is whether each captures a dependency relevant to its target and whether connections between levels can be established without losing what makes the higher-level pattern intelligible.

Reductionism holds, in different forms, that higher-level phenomena can be explained through lower-level theories or constituents. Successful reductions can reveal why a generalization holds and when it fails. Yet derivation from microscopic principles is not always feasible or explanatorily optimal. A description containing every molecular interaction in a fluid may obscure the variables relevant to understanding its bulk flow. The issue is not merely computational difficulty: scientists often seek explanations that remain stable across many differences in microscopic detail.

Multiple realizability and universality sharpen this point. Different physical systems can exhibit similar large-scale behavior despite having different components. Near certain continuous phase transitions, systems with very different microscopic structures share critical exponents. Renormalization-group methods, developed into a major theoretical framework in the 1970s, explain such similarities by tracking how descriptions change with scale and which features remain relevant. These explanations illuminate why some microscopic differences do not matter to the phenomenon under investigation.

Idealization is equally pervasive. Frictionless planes, point masses, infinite populations, and perfectly rational agents are not straightforward descriptions of actual objects. They isolate dependencies, simplify calculations, or identify limiting behavior. An idealized model explains only if the distortion is appropriate to its intended target. Sometimes neglected details can be progressively restored; sometimes the explanatory point is precisely that a pattern is insensitive to them. Robustness across models can support this judgment, although models sharing the same mistaken assumption may exhibit misleading agreement.

Historical sciences add another dimension. Geology, evolutionary biology, and archaeology often explain singular outcomes through sequences reconstructed from surviving traces. They cannot generally repeat the entire history, but they can test consequences of competing reconstructions, investigate processes operating in the present, and compare independent lines of evidence. Social-scientific explanations may additionally incorporate agents’ beliefs, institutions, and meanings. These differences do not establish that each discipline has an unrelated concept of explanation. They show why causal reconstruction, statistical analysis, structural modeling, and interpretation can carry different weights depending on the target and the available evidence.

Evidence, explanatory inference, and computational models

An explanation’s apparent intelligibility does not establish its truth. A coherent mechanism may not operate in the actual system, and a model may reproduce a pattern for the wrong reasons. Scientific appraisal therefore separates the question of what would explain a phenomenon from the question of whether the proposed account is supported. Evidence may bear on the existence of components, the direction of influence, the accuracy of equations, or the domain over which an explanatory relationship remains stable.

Inference to the best explanation treats explanatory success as a consideration in choosing among hypotheses. Candidate explanations may be compared for empirical fit, scope, simplicity, consistency with established knowledge, and ability to generate discriminating expectations. These virtues do not form a universally agreed algorithm, and they can conflict. A broader theory may require more assumptions; a simpler model may omit an important mechanism. The strongest candidate among those considered may also be false if the actual explanation has not been proposed.

This issue connects explanation with scientific realism. Realists often regard the explanatory success of theories involving unobservable entities as grounds for taking those entities seriously. Critics argue that empirical success does not automatically justify belief in every theoretical commitment. Historical theories sometimes explained and predicted successfully while mischaracterizing underlying structures. Assessment must therefore ask which elements of a theory are responsible for its success and whether alternative accounts can capture the same evidence.

Computational science creates additional distinctions. A simulation may explain by implementing an independently supported mechanism, exploring a model’s consequences, or demonstrating that a proposed interaction is sufficient to generate a pattern. Simply producing an accurate output is not enough. Numerical approximations, parameter choices, and sensitivity to initial or boundary conditions can affect what conclusions follow. Reproducing a computational result establishes an important kind of reliability, but it does not independently validate the explanatory assumptions built into the program.

In machine learning, feature-attribution methods, surrogate models, and counterfactual examples are often called explanations. They can answer questions about why a particular model produced an output, but this differs from explaining the real-world phenomenon being predicted. A feature may influence a classifier without causing the outcome represented by its label. An attribution can also depend on how the baseline, comparison distribution, or alternative model is chosen. Evaluating these explanations requires specifying whether the target is the model’s computation, its behavior across inputs, or the causal organization of the external system. The distinction preserves explanatory accountability even when prediction is highly accurate.