Scientific Explanation

Scientific Explanation as a Philosophical Problem

Angela Bogdanova
Published by Aisentica Development
Research Dossier · Philosophy of Science · American English
Reserved canonical: https://philosophybybogdanova.com/dossiers/scientific-explanation/
Internal scholarly edition: November 7, 2026 · Not released

Scientific explanation is the disciplined articulation of why a phenomenon occurs, persists, varies, or takes the form it does. Its object can be an individual event, a repeatable regularity, an organized capacity, a distribution, or a mathematical constraint. An explanation makes relevant dependencies intelligible under a specified description of the phenomenon and commits its claims to forms of assessment appropriate to the discipline. Scientific explanations draw on evidence, theories, models, measurements, mechanisms, and reasoning, yet the possession of any single one of these resources is insufficient to settle explanatory adequacy. The relevant philosophical problem concerns what the resources achieve together.

The history of the subject is a history of increasingly explicit disputes about explanatory relevance. Hempel and Oppenheim (1948) gave a landmark logical formulation in their study of the logic of explanation. Salmon (1984) redirected attention toward the causal structure of the world. Woodward (2003) emphasized counterfactual dependence and intervention in Making Things Happen, while Machamer, Darden, and Craver (2000) examined the organized activities constituting mechanisms in Thinking about Mechanisms. These contributions illuminate genuinely different questions: how a conclusion follows, which facts are relevant, how a system produces an effect, and what would happen if relevant conditions changed.

My governing interpretive claim is that an account of scientific explanation must state its explanatory target and the mode of dependence it purports to disclose. This is a methodological conclusion drawn from comparison of existing accounts, rather than a newly proposed theory. The following chapters examine why this requirement is productive and where its application remains contested.

The Research Question, Corpus, and Method

The central question is: what makes an explanation scientific? It is addressed through historical reconstruction, logical analysis, assessment of rival philosophical arguments, and bounded comparisons with examples from the natural, life, social, and mathematical sciences. The analysis distinguishes primary philosophical claims from the author's interpretation. Scientific validity is treated as a question about publicly assessable grounds, rather than an honorific attached to a particular vocabulary or scientific institution.

The corpus begins with Hempel and Oppenheim's 1948 article and Hempel's subsequent formulation in Aspects of Scientific Explanation (1965). It examines Salmon's statistical and causal work; van Fraassen's pragmatics of explanation; Friedman's and Kitcher's unificationist approaches as documented in scholarly literature; Woodward's interventionism; the mechanistic program in biology and neuroscience; Lange's noncausal explanations; and contemporary discussions of idealization, scientific understanding, and interpretable machine learning. The Stanford Encyclopedia's scientific explanation survey provides a map of debates, while original publications and publisher records establish the relevant authorship, dates, and substantive positions.

Access is differentiated. Open scholarly essays and encyclopedia entries are read as substantive texts; restricted books and journal publications are assessed through verified publisher abstracts, bibliographic records, available excerpts, and corroborating academic exposition. The monograph does not attribute detailed page-level arguments to unavailable editions. Illustrative examples are conceptual reconstructions, not new experiments or independent discoveries. This evidential discipline is part of the philosophical argument: a claim that is persuasive on explanatory grounds still requires an accountable documentary basis.

Explanandum, Explanans, and the Choice of Description

An explanandum is what an inquiry seeks to explain. An explanans supplies the considerations advanced as explaining it. Both terms became central in the logical analysis associated with Hempel and Oppenheim (1948). The distinction establishes a useful grammar: an explanation concerns a relation between a specified question and some relevant information, not merely a collection of true statements about the same subject. A temperature record can describe a heat wave; atmospheric dynamics can explain aspects of its occurrence; a statistical model can estimate the altered likelihood of extreme heat under changed background conditions. These activities overlap without having identical explanatory targets.

The description of the target matters. Explaining why a patient developed a condition can involve mechanisms at the cellular level, population-level risk factors, or the causal contribution of an intervention, depending on the contrast being asked. Van Fraassen's (1980) pragmatics of explanation emphasizes the structure of why-questions and the contrast classes relevant to an answer. His account need not be accepted in full to recognize the methodological importance of identifying the alternatives against which an answer is offered.

An explanatory text therefore should specify whether it seeks the occurrence of an event, the difference between cases, the stability of a relation, or the impossibility of an alternative. Ambiguity here generates false disagreements: one author may demand a process that produces an effect while another explains the mathematical restrictions on possible processes. These are different explanatory undertakings. Making the explanandum explicit also makes failure diagnosable: an answer can be true, informative, and systematically misplaced because it responds to a different question.

Logical Empiricism and the Search for Explanatory Form

The mid-twentieth-century logical approach reconstructed explanation by examining relations among statements, laws, and evidence. Hempel and Oppenheim's (1948) deductive-nomological proposal sought to show how a description of the event to be explained could follow from general laws together with relevant antecedent conditions. The achievement was conceptual: an informal appeal to laws became a structure capable of logical assessment. On this approach, explanatory arguments should have genuine empirical content, their premises should be well-supported, and at least one general law should be essential to the derivation.

This proposal captured an important kind of scientific reasoning. A well-supported law and specified initial conditions can show why an observed result was to be expected under a theoretical description. At the same time, the derivational form places considerable weight on how laws are identified and on whether expectedness amounts to relevance. The Stanford Encyclopedia's account of Hempel traces how the program also addressed statistical laws and the explanation of laws by theories. Treating every subsequent dispute as a simple rejection of Hempel obscures the precision the approach introduced.

The durable lesson is that logical structure is one dimension of explanatory quality. Deductions can reveal what a theory commits an investigator to; consistency and explicit assumptions make explanatory proposals contestable. Yet a correct inference may also preserve a dependence that is explanatorily accidental. We must therefore distinguish the validity of an argument from the explanatory significance of its premises. A valid calculation can answer a prediction question while leaving the mechanism, cause, or contrastive relevance unsettled.

The Deductive-Nomological Model

Consider an idealized metal rod subject to a specified temperature change. If its thermal expansion is accurately approximated by a law in the relevant regime, the initial length, material coefficient, and temperature difference allow a derivation of the expected length change. Under the model of Hempel and Oppenheim (1948), this is the paradigm of an explanation by subsumption under a law. The logical argument makes clear which assumptions are needed and how the conclusion depends on them. It also supports calculations that can be checked against observation.

The example illustrates the strengths and limitations of formal reconstruction. Physical laws used in practice are often restricted by domains, background assumptions, approximations, and error tolerances. The approximate linear law describes a restricted regime and is not an exceptionless universal description of all rods under all temperatures. A derivation may still be scientifically illuminating because the approximation tracks an explanatory dependency in that regime. A rigid identification of explanation with exact universal-law derivation would misrepresent how physicists use domain-limited models. Potochnik's (2017) study of scientific idealization shows why simplifying causal complexity can be integral to scientific inquiry.

The model's deeper difficulty lies in the relation between formal symmetry and explanatory direction. The same mathematical relation can sometimes permit prediction in either direction even when an accepted causal account identifies a privileged explanatory orientation. Such problems suggest that laws and deductive entailment do not exhaust what an account must represent. Their genuine contribution lies in disciplined articulation of theoretical consequences, an achievement preserved within broader causal, mechanistic, and model-based approaches.

Asymmetry, Relevance, and the Classic Objections

If the dimensions and angle of the sun are known, geometric optics can be used to calculate the length of a flagpole's shadow under suitable idealizations. Given the shadow, the sun's angle, and additional information, geometry can also calculate the flagpole's height. Both inferences can be mathematically valid. The flagpole's height helps explain the shadow's length in the intended causal setup; the shadow's length is not ordinarily treated as explaining the flagpole's height. The textbook example isolates a mismatch between derivational reversibility and explanatory direction, as analyzed in the scientific explanation literature.

The objection requires careful handling. The reverse calculation is epistemically valuable: from a measured shadow one can infer an unknown height. That inferential success supplies evidence about height but does not itself reverse the process that generated the shadow. Distinguishing explanation from measurement-based inference resolves the apparent contradiction without depreciating the reverse calculation. A second family of objections concerns relevance. Statistical regularities or universal generalizations may be true while containing irrelevant factors. An empirically successful inference may include a feature that is merely correlated with the cause of an effect.

These objections matter because they reveal what philosophical analysis should preserve: scientists need valid inferences, credible laws, and causal or other dependency information. The mistake would be to demand that one of these dimensions stand in for all the others. A satisfactory evaluation asks what feature of the world makes the explanandum depend on the explanans and whether the representation actually isolates it.

Prediction and Explanation

Prediction concerns the anticipation or estimation of outcomes; explanation concerns what makes an outcome, pattern, or constraint intelligible. Many scientific achievements perform both functions. Newtonian dynamics enables explanations and predictions of motion within its domain, just as physiological models may describe processes that also permit targeted forecasts. The two accomplishments are connected through theory, but their standards can diverge. A rule that repeatedly predicts a variable can exploit a stable proxy without identifying the process that produces it.

Hempel's logical program placed explanatory and predictive arguments in close structural proximity, a relation examined in his account of scientific explanations. Later causal approaches identify why successful forecasts alone leave open questions of manipulation and counterfactual dependence. Woodward's (2003) interventionist analysis asks which relationships remain relevant when particular variables are altered under specified conditions. That question exceeds the information supplied by predictive accuracy on an unchanged data distribution.

A reliable prediction is evidence that an organized pattern has been captured. Its explanatory interpretation depends on how the predictive variables are related to the phenomenon and how alternative generative processes have been assessed. Conversely, an explanation may identify a mechanism while supporting only probabilistic predictions because of stochastic dynamics, incomplete initial conditions, or intrinsic measurement limits. The epistemic tasks should be recorded distinctly in studies, reviews, and public accounts. This prevents the rhetorical promotion of predictive success into a causal claim and preserves the value of prediction as a separate scientific accomplishment.

Statistical Laws and Explanatory Probability

Many events that science explains are not expected with near certainty. In epidemiology a risk factor may increase the probability of a disease without ensuring any individual's outcome. In statistical physics and certain quantum processes, distributions and transition probabilities enter directly into scientific description. A criterion requiring explanations to make the explanandum highly probable would marginalize important scientific phenomena, especially unusual events governed by well-understood stochastic structures.

Salmon, Jeffrey, and Greeno's (1971) Statistical Explanation and Statistical Relevance contains Salmon's influential argument that statistically relevant factors should inform explanation even when particular outcomes remain improbable. This addresses a limitation of attempts to equate explanatory force with high conditional probability. If a rare event occurs in a characterized process, identifying the process and relevant distribution can illuminate the event without turning rarity into likelihood. The explanation concerns why the event falls within a structured space of possibilities and what changes its relative chance.

The difficulty is specifying relevance. A statistical relationship may be genuine but explained by a common cause, selection process, or measurement practice. Controlling for every observed variable can also eliminate information about mediating processes or introduce bias when conditioning on a collider. The philosophical point is that conditional probability alone does not identify the entire dependence structure. Statistical relevance is an indispensable clue, to be interpreted in relation to research design, causal hypotheses, and the specific explanatory question.

Explaining an Improbable Event

Imagine that a radioactive nucleus decays within a short observation interval even though the model assigns a low probability to decay during that interval. The event is unlikely, and yet the model can specify a decay process characterized by a hazard rate and probabilistic law. The scientist can explain the event's membership in a well-defined stochastic process without claiming to identify a deterministic trigger for that particular moment. Salmon's (1971) statistical relevance analysis was motivated in part by the philosophical significance of objectively improbable events.

It is useful to separate three questions. Why can decay occur at all in this physical system? Why does the model assign this probability over a specified interval? Why did this nucleus decay at precisely this instant? A scientifically satisfactory answer to the first two may coexist with a principled restriction on answers to the third, depending on the physical interpretation and state information available. Treating all three as one question creates an artificial appearance of incompleteness. The claim that a probability distribution is explanatory must itself be supported by the model's evidential and theoretical credentials.

This example also distinguishes explanation from narrative completion. A desire to assign every realized event a hidden deterministic reason is a further metaphysical commitment, not a logical consequence of asking why. Scientific explanations can specify structured possibility, distributional dependence, and constraints. Their adequacy depends on the explanatory target rather than on the promise to abolish uncertainty.

Salmon's Shift Toward Causal Structure

In 1984 Salmon advanced an influential causal and mechanical conception in Scientific Explanation and the Causal Structure of the World. He distinguished epistemic, modal, and ontic ways of conceiving explanation and argued that explanation should address processes and interactions in the world, rather than reduce explanatory success to the properties of an argument. This shift supplied a philosophical response to the fact that logical and statistical inferences may be formally acceptable while failing to discriminate causal from coincidental connections.

Salmon's position is historically specific. Its technical treatment of causal processes and physical randomness should not be compressed into the assertion that all explanations simply mention causes. The book also addresses the relationship between explanation, realism, and the statistical character of physical science. Its publisher description identifies a causal/mechanical account as the central defended position; detailed technical claims about chapters inaccessible in this research are therefore avoided. The Stanford Encyclopedia's causal explanation survey places these debates in their broader philosophical setting.

The ontic orientation offers an enduring insight: representational success should remain answerable to how the phenomenon is actually generated, organized, or constrained. It also encounters hard questions. How should omissions, equilibrium cases, mathematical necessities, or explanations by higher-level regularities be assessed when a causal process is not the entire story? The historical development of the field shows why causal information has become central without closing the discussion of noncausal scientific explanation.

Causal Explanation and Difference-Making

A causal explanation identifies dependencies that bear on the occurrence or variation of a phenomenon. This formulation encompasses several traditions and does not itself select one metaphysics of causation. In a laboratory context, researchers ask whether an intervention on a variable changes an outcome under controlled conditions; in an observational context, they use a combination of background knowledge, measurement, statistical adjustment, natural experiments, and sensitivity analysis to evaluate the same type of dependence. A scientific causal claim must specify both the target of change and the assumptions under which a difference-making interpretation is warranted.

Woodward's (2003) Making Things Happen develops interventionist and counterfactual approaches that relate causal explanation to the potential for manipulation and control. The relevant intervention is an idealized theoretical notion, not a requirement that researchers physically manipulate every cause. This is crucial in astronomy, historical geology, and climate science, where controlled laboratory intervention on the focal phenomenon is generally impossible. Counterfactual comparison can remain meaningful through constrained modeling, historical evidence, and independent tests.

Difference-making highlights explanatory asymmetry, but it depends on causal structure already being responsibly specified. Merely observing that changing a model input alters its output is insufficient to establish that the modeled intervention corresponds to a feasible or well-defined causal intervention in the world. Causal explanation links inference with substantive claims about what alterations would matter; its strength emerges from the scrutiny of these claims, their assumptions, and alternative structures.

Intervention, Invariance, and Counterfactuals

The ability to assess how an outcome changes when relevant conditions change is a central mark of explanatory understanding in interventionist accounts. Consider a simple model in which a parameter governs a chemical reaction rate within a calibrated experimental regime. An adequate causal analysis predicts a pattern of variation when the temperature is altered, subject to assumptions about the reaction pathway and other controlled conditions. The counterfactual question asks which dependency would persist under the relevant intervention, and the invariance question asks over what range it persists.

Woodward's interventionist account of explanation clarifies the difference between an ordinary association and a relation suitable for causal intervention. A manipulable dependency may be more explanatorily valuable than an impressive correlation that breaks down under policy change, experimental control, or environmental shift. Yet invariance comes in degrees. A causal model can be reliable within a stated range without being universal, while an intervention can be poorly defined if it changes several causal variables at once.

A philosophical assessment should therefore ask for the intervention variable, outcome variable, background conditions, admissible comparison, and stability domain. These questions expose gaps often concealed by the phrase 'X explains Y.' They also prevent the overstatement of counterfactual modeling as direct experiment. A computational sensitivity exercise is informative about a model; a causal claim about nature requires further evidence that the model's intervention semantics and empirical premises are adequate.

Mechanisms, Entities, Activities, and Organization

Mechanistic explanation investigates how organized entities and activities produce, maintain, or modify a phenomenon. Machamer, Darden, and Craver (2000) analyzed mechanisms in terms of organized entities and activities, an influential perspective in the philosophy of biology and neuroscience. On this view, the components matter through their activities and their organization: the same inventory of parts can have different consequences when connections, timing, or spatial arrangements differ. A catalogue of molecular components therefore cannot by itself explain a functional process.

A mechanism can also be investigated at several scales. Researchers may identify the relevant cell populations, their interactions, and biochemical pathways, then relate these to changes in behavior or physiological function. The account is explanatory when it shows how the specified organization bears on the phenomenon and how relevant perturbations would affect it. This is more exacting than a diagrammatic metaphor in which arrows connect familiar nouns. The SEP entry on mechanisms in science explores how contemporary mechanistic approaches relate to causal relevance, discovery, and scientific representation.

Mechanistic accounts have genuine limits. The mechanisms underlying a phenomenon may be unknown, multiply realizable, or too complex for exhaustive reconstruction. Higher-level statistical patterns can remain informative even where a complete component-level explanation is unattainable. The philosophical benefit of mechanistic analysis lies in the organization of testable questions and in its demand for productive connections among parts, activities, and the explanandum. A satisfactory explanation must identify what level and granularity of mechanism the question requires.

The Neuroscientific Case: The Action Potential

The action potential provides a bounded example of mechanistic reconstruction. A textbook summary may say that a neuron fires when stimulation exceeds a threshold. A more explanatory account identifies how membrane conductances, electrochemical gradients, voltage-dependent ion channels, and temporally organized changes in permeability contribute to the characteristic electrical event. The investigation moves from observed waveform to dependencies within a biological process. Craver's (2007) Explaining the Brain uses the historical Hodgkin–Huxley action-potential model in discussing multilevel standards of neuroscientific explanation.

The lesson is methodological. Experimental records of voltage and current constrain proposed mechanisms; models represent the dynamics; interventions and comparisons help distinguish which activities are constitutive and causally relevant in the particular explanation. At the same time, the familiar Hodgkin–Huxley model has a defined level of abstraction and does not provide an exhaustive molecular account of every neuron. Its explanatory power depends partly on the question and scale: describing the generation of a membrane potential waveform is distinct from explaining a behavioral choice or a conscious experience.

Mechanistic inquiry should not be mistaken for an unlimited promise of reduction. To explain a complex outcome, the investigator must identify which components and conditions are explanatorily salient. More microscopic detail can sharpen an account, but it can also distract from relevant system organization. The best level of analysis is the level that resolves the actual explanatory question with reliable, testable dependencies.

Mechanistic Evidence and the Problem of Completeness

Descriptions of mechanisms often differ in completeness. A diagram can identify a possible pathway, an experimental perturbation can implicate one component, and a convergent research program may eventually establish many stages of a process. These are different evidence states. Craver's (2007) standards of multilevel explanation illuminate why plausible component stories must be evaluated against intervention evidence, organizational relations, and relevant phenomena.

Completeness is relative to an explanandum. Explaining how an enzyme catalyzes a reaction at a certain scale does not require tracing every quantum interaction, yet it may require enough structural detail to distinguish catalytic pathways from competing accounts. Explaining the emergence of a system-level pattern may require both component interactions and boundary conditions. Scientific practice uses partial mechanisms as research guides, provided investigators label what is hypothesized, directly observed, inferred, and idealized.

A serious danger is the transformation of a mechanism sketch into a settled explanation because its graphical presentation is persuasive. A coherent arrow diagram might depict only hypothesized relationships, and a successful simulation might be compatible with several possible pathways. A better appraisal makes the epistemic status of each link explicit. Mechanism science develops by progressive testing, not by the cumulative accumulation of labeled boxes. This conclusion is supported by the methodological implications of the mechanisms literature and does not require a newly named doctrine.

Unification and the Economy of Scientific Understanding

Unificationist theories place weight on the capacity of an explanatory pattern to integrate phenomena that previously appeared disconnected. Friedman (1974), in Explanation and Scientific Understanding, and Kitcher's later work are central landmarks in the development of this approach. The motivating thought is that scientific understanding increases when a small number of explanatory patterns organizes a broad range of facts. The attraction is clear: a principle that explains multiple classes of observations seems intellectually deeper than an unrelated collection of local accounts.

Unification has a productive historical role. Scientific research often progresses by connecting different empirical regularities under a shared theoretical treatment, thereby supporting novel comparisons and revealing new dependencies. The SEP discussion of scientific unity distinguishes structural, methodological, and explanatory forms of unity. These distinctions matter because a single standardized notation can unify descriptions without uncovering a common causal mechanism, while one theory may reveal a shared dependence across very different phenomena.

A unifying account still requires constraints on acceptable explanatory patterns. An artificially flexible language can compress observations while concealing their differences. The appearance of economy may be purchased through assumptions that make the account untestable or empty. Genuine explanatory unification must therefore preserve empirical and conceptual relevance, distinguish robust integrations from mere redescriptions, and remain responsive to anomalous cases. The philosophical value of unification is substantial; its status as an exclusive criterion for explanation remains contested.

The Limits of Unification

The promise of unity can become a distortion when an explanation compresses distinct processes into the same verbal formula while hiding relevant differences. Two diseases may produce similar observable outcomes through separate molecular pathways. A broad epidemiological generalization can be useful for classification and prediction, yet individual clinical explanation may require distinct mechanisms. In this case, greater descriptive economy is not automatically greater explanatory depth. The SEP treatment of scientific unity documents philosophical disputes concerning how explanatory unity interacts with reduction and disciplinary difference.

Conversely, explanations that appear local may be unified at the level of an underlying principle without being reducible to a single descriptive mechanism. Mathematical symmetry constraints can govern different physical systems even though the realized causal processes differ. Recognizing this permits a more discriminating appraisal: one should ask whether unity results from a shared law, structural relation, mechanism, idealization, or classification. These alternatives have different evidential implications.

I draw the interpretive conclusion that unification is a significant explanatory virtue when it compresses dependencies while preserving distinctions that matter to the question. This conclusion can be defended from the contrast between genuine integration and overgeneralization. It remains a judgment within established debates. It does not establish a novel unificationist theory. The demand for explicit relevant distinctions also clarifies why explanatory depth cannot be read off a simple count of how many phenomena one formula appears to cover.

Pragmatic and Contrastive Why-Questions

Why did this particular experimental sample respond to treatment rather than remain unchanged? Why did one patient recover while another did not? Why does one type of organism occupy a habitat while a similar species does not? These questions request differences among alternatives. Their answers may depend on the contrast class, background conditions, and dimensions of interest. Van Fraassen (1980) addressed these features in The Pragmatics of Explanation, a chapter of The Scientific Image.

Pragmatic relevance helps explain why the same scientific material can support different valid explanations. A physiologist may focus on ionic transport, a clinician on treatment response, and a public health researcher on population exposure. Their questions may concern the same sequence of events, yet differ in which variation must be accounted for. An explanation that resolves the clinical contrast can be incomplete for the physiological question without being defective within its stated scope.

This flexibility does not reduce explanation to what an audience happens to find satisfying. Answers remain answerable to evidence and must correctly discriminate the alternatives invoked. The contrast can change what counts as relevant without changing whether a causal connection is true. The epistemic discipline lies in coupling transparent question selection to defensible knowledge. An explanation succeeds when it gives a supported answer to a specified question, and a philosophical critique must evaluate both the answer and the appropriateness of the question's assumptions.

Explanation and Scientific Understanding

Scientific understanding is often associated with the ability to connect facts, use a representation, answer counterfactual questions, or anticipate the consequences of changing assumptions. An explanation may contribute to these abilities. Their exact relation remains disputed. De Regt (2017), in Understanding Scientific Understanding, examines understanding as a philosophical problem in its own right, including the role of intelligible theories and scientific practice. Khalifa's (2017) Understanding, Explanation, and Scientific Knowledge also treats the relation as a substantive question rather than an automatic equivalence.

A researcher can understand how to manipulate a model without knowing whether the model's assumed structure is correct. A polished visual explanation can produce strong subjective conviction without generating justified belief. Conversely, a correct and detailed technical account may initially be opaque to its intended reader, yet still contain the dependencies required for scientific explanation. These examples require distinguishing objective explanatory merits, competent use, and experienced intelligibility.

The present dossier addresses the philosophical nature of explanation. The adjacent dossier Scientific Explanation and Understanding develops the wider epistemology of understanding as a distinct subject. The boundary reflects a genuine difference between the adequacy of explanatory relations and the cultivation or possession of understanding. It also guards against turning ease of communication into the criterion by which scientific explanations are judged.

Mathematical and Noncausal Explanation

Some scientific answers appear explanatory because they establish why an alternative is impossible under a structural or mathematical constraint, rather than because they identify a causal sequence generating the actual outcome. Lange (2016) examines such cases in Because Without Cause, including examples from mathematics and the sciences. The Königsberg bridge problem is instructive: given an abstract graph with the relevant pattern of odd-degree vertices, there is no Eulerian traversal of every edge exactly once. The impossibility follows from graph structure rather than a story about why a particular walker failed on a particular day.

The distinction depends on the explanandum. If the question is why one person failed to finish a route, fatigue or misinformation could be relevant. If the question is why no permitted route has the required combinatorial property, the parity argument answers it. The same setting therefore admits both causal and mathematical explanations, each directed to a differently formulated phenomenon. The SEP entry on mathematical explanation surveys debates about such cases and the role of mathematics in empirical science.

A noncausal explanation should still make contact with a well-specified target and valid supporting assumptions. Graph-theoretic necessity applies to a graph with given features; whether that graph faithfully represents a physical bridge system is an additional modeling question. Formal necessity can be explanatory without eliminating empirical responsibility for the correspondence between model and situation. The existence of this class of examples challenges the assumption that every explanation must be an account of causal production.

Idealization and Abstraction

Science often explains complex systems by neglecting real factors, imposing simplifying boundaries, or working with limiting cases. Idealization can isolate a dependence that would be hidden in a fully detailed description. Potochnik (2017) defends the importance of idealization in scientific inquiry, connecting simplified models to the representation of causal patterns. Cartwright (1983), in How the Laws of Physics Lie, challenged simplistic understandings of the universal truth of laws as deployed in scientific models.

The philosophical question is when simplification preserves explanatory relevance. A model treating an object as a point mass can accurately explain a trajectory for some purposes while failing to explain rotational effects. A population model assuming homogeneous mixing may illuminate a general transmission mechanism while misstating clustered dynamics. A successful idealization makes its domain, ignored factors, and sensitivity to those omissions assessable. Its explanatory virtues cannot be judged from literal realism alone.

Idealization creates two obligations. First, the scientist should identify which features are intentionally suppressed and why. Second, the account should show that the claimed dependency remains relevant under appropriate relaxations or empirical comparison. Some explanations are especially powerful because their central result is robust to the omitted detail. Others fail precisely when a neglected factor dominates. The distinction between epistemically productive idealization and misleading simplification must be established case by case rather than by praise or condemnation of ideal models in general.

Models as Instruments of Explanation

Scientific models are selective representations that can organize data, express dynamical laws, simulate interventions, and establish mathematical constraints. Some directly target an empirical process; others describe idealized systems for comparison. The SEP survey of models in science surveys the diversity of model-based representation and the philosophical problems of idealization and explanatory understanding. The explanatory role of a model is therefore an achievement requiring assessment, not an automatic consequence of calling a construct a model.

A mathematical model can support explanation by making dependencies explicit. A simulation can show how local interactions generate an aggregate pattern under specified rules. A diagram can help articulate the organization of a mechanism. In each case, the movement from the representation to the world requires independent justification. A simulated pattern cannot by itself demonstrate that the hypothesized process generated an observed event. Empirical calibration, parameter constraints, external validation, robustness checks, and domain knowledge all affect that inference.

Models also produce different forms of explanatory success. A coarse-grained model may capture the difference-making variable relevant to a policy decision, while a detailed microscopic simulation may be valuable for another question. The explanatory adequacy of either depends on the phenomenon specified and the dependencies preserved. This explains why the neighboring Research Dossier Scientific Models deserves its own treatment: representation and explanation intersect in practice, but the philosophy of a representation is not exhausted by its possible explanatory use.

Computer Simulation and Explanatory Evidence

Computational simulation extends the range of conditions scientists can explore. Researchers can construct models of systems whose interactions are difficult to manipulate directly, examine possible trajectories, and compare interventions under assumed dynamics. Such work is particularly important where the systems are spatially extended, historically unique, or computationally intractable by ordinary analytical means. Yet a simulation's numerical output is conditional on its governing equations, parameterization, initial state, boundary conditions, and implementation.

The SEP discussion of models provides philosophical context for understanding simulation as a representational practice. A simulation may explain a pattern within a model by showing how specified rules generate it; establishing that this is an explanation of nature requires separate validation. The comparison resembles a deductive argument: valid consequences can follow from assumptions that do not accurately describe the target system. Numerical precision cannot repair an unsupported premise.

An appropriate explanatory report should distinguish internal model dynamics from externally supported physical or social dependencies. Parameter uncertainty, structural uncertainty, choice of resolution, and alternative modeling assumptions should be visible to the argument. The resulting scientific explanation can be strong even when uncertainty remains, provided its conclusions match the evidence and declared scope. What matters is traceable inferential and empirical support, rather than the visual complexity or computational expense of the simulation.

Explanation in Historical Sciences

Geology, evolutionary biology, paleontology, and astronomy investigate events whose crucial conditions cannot usually be recreated in the laboratory. Their explanations often reconstruct sequences from surviving traces, use general processes to delimit possibilities, and compare rival historical scenarios. This practice shows why physical intervention on the original event is not a universal prerequisite for scientific causal explanation. The philosophical question is whether the inference from present evidence to past processes is sufficiently constrained and discriminates credible alternatives.

The SEP entry on scientific unity considers the difficulties of imposing a strict covering-law model on historical sciences. Causal approaches, by contrast, can make sense of singular explanations grounded in mechanisms, configurations, and contingent circumstances. A geological episode may be explained through stratigraphic evidence, chemical signatures, known physical processes, and temporal ordering without possessing an exceptionless law from which the episode could be deductively derived.

Such cases also expose the risk of narrative excess. A plausible chronology is not automatically a scientific explanation. A credible reconstruction explains why particular traces would be expected under the proposed history and less expected under serious rivals. Independent lines of evidence can constrain the same inference. Historical scientific explanation thus integrates context-sensitive narrative with rigorous counterfactual and evidential discipline. It requires neither the fiction of repeating the past nor an abandonment of causal assessment.

Explanation in Population Sciences

Population sciences confront the difference between individual outcomes and patterns of differential risk. A distribution of cases may be explained by a change in exposure, susceptibility, or environmental context even though the precise sequence in every case remains unavailable. Statistical explanation is therefore more than a substitute for an incomplete story about individual causes. It can be the appropriate description of an explanandum defined at population level. Salmon's (1971) statistical relevance work is a central philosophical reference for separating probabilistic explanatory information from mere inductive confirmation.

Consider an idealized epidemiological comparison. A factor may alter incidence rates while most exposed individuals remain unaffected. The investigator must ask whether the association reflects a genuine causal contribution, a shared antecedent, selection effects, measurement differences, or other explanations. A risk ratio is a summary of observed relationships under specified assumptions; it is not automatically a mechanistic account of every individual case. Population explanations can nevertheless be precise about how the risk distribution changes and under which conditions.

The philosophical assessment should track the level at which the question is asked. When policy concerns the reduction of cases in a population, a stable and credible causal population relation can provide useful explanation. When a patient asks why a particular event occurred at a particular time, further individual-level information may be needed and sometimes may remain inaccessible. The scientific status of the first explanation does not depend on its solving the second. This example reveals how question specification guards against false standards of completeness.

Explanatory Relevance and Selection of Variables

Every empirical system contains more facts than a scientific explanation can include. A successful explanation selects variables, background conditions, and dependencies because they matter to the specified question. The selection is not merely editorial compression: it embodies a judgment about which features support the relevant counterfactual, mechanistic, statistical, or mathematical claim. Woodward's (2003) counterfactual account of causal explanation clarifies why some distinctions among variables contribute explanatory information and others remain incidental.

The task creates a vulnerability. If variables are chosen solely for predictive contribution, the resulting model may rely on a proxy that lacks the intended causal significance. If a mechanism is constructed solely from familiar biological components, important environmental or organizational factors may disappear. Conversely, including every correlated variable can undermine intelligibility without improving validity. Explanatory relevance has to be defended by a relation to the explanandum and to alternative hypotheses.

One practical test is to ask what would be lost if the proposed factor were removed from the explanatory account. Would the account cease to distinguish the alternatives, fail an intervention comparison, omit an essential productive activity, or lose the necessary constraint? These questions have different answers depending on the explanation type. Their value lies in exposing the work each selected factor performs. Relevance becomes assessable when the writer states the relation by which a fact earns its place in the account.

Inference to the Best Explanation

Inference to the best explanation concerns the use of explanatory virtues in the evaluation of hypotheses. Lipton's revised second edition of Inference to the Best Explanation (2004) examines how explanatory considerations inform scientific reasoning and discusses contrastive explanation and Bayesianism. The idea is influential because scientists often compare hypotheses that fit some shared observations but differ in coherence, integration, auxiliary commitments, and capacity to account for unexpected details.

The epistemic challenge is that explanatory attractiveness is not equivalent to truth. A simple narrative may seem compelling because it suppresses contrary evidence. A hypothesis may unify existing observations because it was tailored after those observations were known. Claims about elegance, mechanism, and broad scope acquire evidential force only when their relation to independent data and serious alternatives is made explicit. The SEP survey of scientific realism places inference to the best explanation within debates about the status of successful scientific theories.

The scientific merit of such inference depends on the quality of the hypothesis space and the safeguards against circular evaluation. An explanation should be assessed for empirical consequences, stability under new evidence, and robustness against reasonable rivals. The phrase 'best explanation' should designate the result of comparison under transparent standards, not serve as a substitute for that comparison. Explanation can guide inference while remaining subject to independent checks.

Scientific Realism, Empiricism, and Explanatory Commitment

Scientific explanations often invoke entities or processes that are not directly observed. Electrons, genes, causal structures, and theoretical fields acquire roles within explanations through their contribution to well-supported scientific practices. Philosophical realists and empiricists disagree about what explanatory success licenses one to believe regarding the unobservable structure of the world. The SEP discussion of scientific realism sets out competing arguments about truth, reference, and the epistemic implications of successful science.

Van Fraassen (1980) develops constructive empiricism in The Scientific Image, distinguishing empirical adequacy from the stronger commitment that a theory's claims about unobservables are true. His pragmatic analysis also maintains that explanatory strength is not simply additional empirical evidence for a theory. A realist can reply that certain explanatory successes are difficult to understand without commitment to the described structures. The argument is substantive because the same body of successful scientific practice can be interpreted through different epistemic standards.

The present research does not settle this dispute through stipulation. It identifies a useful difference between assessing whether an account answers an explanatory question and assessing which ontological commitments its success warrants. A model may be explanatory in a restricted empirical domain even while its unobservable posits remain debated. Conversely, claims of ontological discovery need evidential support beyond the aesthetic attractiveness of an explanation. Scientific explanation and scientific realism are related inquiries with distinct burdens of argument.

Evidence, Observation, and Theory Dependence

An explanation must engage evidence, and evidence itself is produced through instruments, classification, models, and practices of interpretation. The SEP treatment of theory and observation traces debates about whether observation can be understood independently of theoretical assumptions. Theory dependence matters because an explanatory model may shape which observations are considered relevant, which variables are measured, and how anomalous data are classified.

This interaction does not imply that observation is arbitrary. Instruments can be calibrated, alternative measures compared, assumptions tested, and independent research practices made to converge or diverge. A persuasive explanation should expose how its key observations were obtained and whether competing descriptions remain possible. Multiple methods are valuable when they place different constraints on the same explanatory hypothesis, especially when a single model might accommodate many patterns through flexible parameter choices.

There is a further issue of evidential asymmetry. A datum can support a hypothesis because it is expected under it, yet the hypothesized relation need not explain the datum in the philosophically relevant sense. Inferential support, causal generation, and mathematical entailment need not coincide. Scientific authors should identify whether a cited study establishes a phenomenon, estimates an association, tests a mechanism, or adjudicates an explanatory alternative. That distinction improves both conceptual rigor and the integrity of source-based scientific writing.

Levels, Reduction, and Explanatory Autonomy

Researchers explain phenomena at molecular, cellular, organismic, population, ecological, and other scales. Whether a higher-level explanation can be reduced to a lower-level account depends on what reduction means and on the available theories. Craver's (2007) multilevel mechanistic program investigates how component activities and system organization figure in explanations, while the SEP entry on scientific unity examines distinctions among explanatory, theoretical, and ontological unity.

The existence of lower-level constituents does not settle which variables best explain a higher-level pattern. In some cases lower-level details reveal a mechanism that changes the assessment of a phenomenon. In others a coarse-grained model explains stable system properties precisely because many microscopic configurations make the same macro-level difference. Distinct disciplines can discover regularities with independent explanatory usefulness without implying that their subject matter is metaphysically disconnected.

The philosophical task is to make explanatory scale and cross-level relations precise. One should ask whether a proposed lower-level account genuinely derives the higher-level pattern, whether it identifies a productive mechanism, or whether it offers only a description of underlying components. Equally, one should assess whether a higher-level generalization is robust, causally informative, or merely classificatory. Neither greater detail nor greater generality is an automatic index of explanatory superiority. What matters is the dependency the question requires.

The Challenge of Black-Box Machine Learning

Machine learning systems can achieve powerful predictive performance while leaving their operation or the relationships they exploit difficult to interpret. Scientific use of such models makes the distinction between prediction, explanation, and understanding especially urgent. Rudin (2019) argues in her critique of black-box explanations that high-stakes decisions warrant stronger attention to models that are interpretable by design. The original article discusses risks of relying on post hoc interpretations of opaque systems.

The central problem for philosophy of science is evidential and representational. A predictor trained on data may exploit stable correlations in one setting while failing under distributional shifts or interventions. A feature-attribution visualization can describe aspects of model behavior, but by itself does not establish the physical mechanism or causal dependency that produced the observed phenomenon. This does not render machine-learning-assisted explanation impossible. It clarifies the additional work required to show that a learned relation represents the scientific explanandum appropriately.

A defensible study should separate three questions: why did the algorithm output this result, why does the observed phenomenon occur, and how reliably does the algorithm's behavior generalize? These questions can be related while retaining distinct evidence requirements. Model architecture, training provenance, perturbation tests, domain knowledge, and external validation contribute to the assessment. The philosophical point applies beyond any particular technology: the explanation of a representation's operation is one object of inquiry, and the explanation of the represented world is another.

The Public Form of Scientific Explanations

Scientific explanation is communicated through papers, datasets, code, diagrams, mathematical models, and the practices by which claims are revised. Its public form matters because reasons and evidence must be inspectable by others. A scientific account becomes more accountable when it identifies its explanandum, its assumptions, its supporting studies, the status of its inferences, and the conditions under which it could be corrected. This is consistent with the emphasis on intersubjective testability in the SEP account of Hempel and with contemporary attention to models and scientific practice.

The publication of an explanation introduces further responsibilities. A hyperlink should lead to the precise document or source asserted, rather than to an institution that happens to publish related material. A full bibliography should give enough information to distinguish books, articles, editions, and revisions. A publisher's description of a restricted book can verify its bibliographic identity and general thesis; it cannot substitute for inspection of inaccessible detailed passages. Explicitly recording that difference improves scholarly reliability.

These requirements become especially important when research is written, indexed, and reused by computational systems. Machine-readable source registries and cryptographic file digests make provenance auditable. They do not establish the truth of philosophical conclusions, but they can expose whether the cited evidence was altered, omitted, or mismatched. Scientific explanation as a public practice depends on the continuing availability of responsible criticism and correction as well as the initial plausibility of its arguments.

Comparative Assessment of Major Approaches

The deductive-nomological approach examines whether a phenomenon can be derived from laws and conditions; it disciplines inference but encounters asymmetry and relevance objections. Statistical approaches illuminate probability distributions and relevant factors, especially where the explanandum is intrinsically probabilistic; they require further care with correlation and causal structure. Causal and interventionist explanations illuminate difference-making under counterfactual variation; they depend on defensible assumptions about intervention and domain. Mechanistic explanations reveal productive organization in particular systems; they remain selective with respect to level and detail. Unificationist theories evaluate explanatory integration across phenomena; they must distinguish informative economy from artificial compression. Mathematical explanations can disclose constraints and impossibilities beyond causal narratives.

These comparisons are grounded in distinct philosophical contributions: Hempel and Oppenheim (1948) in logical explanation, Salmon (1984) in causal and statistical accounts, Woodward (2003) in interventionist difference-making, Machamer, Darden, and Craver (2000) in mechanistic organization, and Lange (2016) in noncausal explanatory constraints. None supplies a single uncomplicated criterion that resolves every successful scientific practice.

My conclusion from this comparison is that disagreements are most productive when the explanandum and intended explanatory relation are specified before virtues are ranked. The claim is methodological and interpretive. It recognizes the irreducible importance of evidence and argument while resisting the demand for one universal structure. It is compatible with the possibility that future scholarship could develop stronger general theories, without presenting this monograph as such a theory.

Criteria for Evaluating a Concrete Scientific Explanation

An appraisal can begin with a series of connected questions. What exactly is to be explained, and against which alternatives? Does the proposed account provide a lawlike dependency, a probabilistic structure, a causal relation, a mechanism, a unifying pattern, or a mathematical constraint? Why are its selected factors relevant? What empirical observations or theoretical results warrant the dependencies asserted? Within what conditions and degree of approximation are the claims intended to hold? Which serious rival accounts would produce different expectations? These questions focus assessment on the concrete work an explanation is asked to perform.

The comparison of scientific models, causal explanation, and mechanistic accounts shows why the answers vary by case. A mathematical impossibility argument can be adequate without an intervention experiment; a causal treatment effect demands careful identification; a mechanism sketch requires evidence for its organized parts and productive relations. Evaluation should retain those distinctions rather than flatten them into one checklist of formal properties.

A second set of questions addresses scholarly conduct. Does the bibliography identify the actual source? Is the claim supported by that work at the level at which it has been accessed? Are uncertainties and disputed interpretations distinguished from established observations? Does the argument separate the author's reconstruction from the position of the historical scholar? An explanation can be conceptually strong yet poorly sourced, or impeccably cited yet scientifically unpersuasive. The two dimensions demand independent but coordinated scrutiny.

Original Analytical Synthesis

The central result of this inquiry is an account of explanatory assessment rather than a new theoretical system. Scientific explanation is best evaluated by considering the fit between a specified explanandum, the dependencies an account claims to disclose, the available evidence for those dependencies, and the alternatives against which the account is offered. The comparison is warranted by the limitations of reducing explanation to deduction, probability, causation, mechanism, unification, or mathematical necessity alone. Each historical program reveals a genuine explanatory dimension and a distinct set of possible failures.

This yields three connected arguments. First, explanatory claims can be assessed meaningfully only when the phenomenon is described at the relevant scale and contrast. Second, the evidence required for an explanation depends on the kind of dependence claimed: logical validity, robust statistical association, intervention support, mechanistic organization, or mathematical proof establish different things. Third, explanatory adequacy is sensitive to scope. A successful idealization is sufficient for some questions while insufficient for others. These arguments synthesize long-standing discussions, including van Fraassen's context-sensitive account, Woodward's interventionist framework, and Potochnik's analysis of idealization.

The synthesis also clarifies the relationship between explanatory truth, explanatory usefulness, and explanatory communication. A scientifically useful account may be restricted, approximate, and revisable while providing reliable understanding of a dependency. A clear presentation cannot compensate for a false dependence, and an accurate dependence cannot rescue an answer to the wrong question. Scientific explanation reaches its strongest form when its claims, limitations, and support remain available for public correction.

Conclusion

Scientific explanation is a central practice through which research connects phenomena to their relevant dependencies, processes, regularities, and constraints. Its philosophical investigation has moved from explicit logical models through probability and causality to mechanistic, unificationist, pragmatic, and mathematical accounts. This history has produced a richer understanding of explanatory success. Explanations can disclose why an event was expected, why a probability changed, how a process is organized, what difference an intervention makes, why disparate phenomena fall under a shared pattern, or why an alternative is structurally impossible.

The original analytical contribution of this monograph is a sustained comparison showing why the explanatory target must be defined before the adequacy of an answer can be judged. The account's dependencies must then be evaluated through the kind of evidence those dependencies require, with rival explanations, scope conditions, and representational assumptions made explicit. This conclusion is supported by the established literature and does not claim the invention of a novel philosophical doctrine. It is an argument about how the principal traditions can be read together while retaining their real disagreements.

A scientific explanation earns intellectual authority through its relation to evidence and publicly assessable reasoning. Its quality is tested in the alignment of question, argument, world, and documented support. The advance of computational science intensifies these obligations, especially where simulations and predictive models can produce persuasive outputs without immediately revealing the relevant dependencies. Scientific inquiry progresses when explanatory proposals are precise enough to be evaluated, rich enough to illuminate, and open enough to be revised.

Research Note and Scholarly Scope

This is a source-stage scholarly manuscript prepared for private batch publication. The principal theoretical claims were compared against original publisher documentation, journal abstracts and records, accessible scholarly articles, and living reference entries. Publisher metadata supports authorship, dates, bibliographic details, and documented general positions; it is not treated as proof of detailed arguments in restricted chapters. Examples are analytical reconstructions rather than original empirical studies. Scientific claims are interpreted within their documented source scope. Scientific Explanation and Understanding is a separate planned dossier with a different explanandum. No external peer review, public publication, or site deployment is asserted.