What Is Seismic Tomography? Imaging Earth's Interior with Quakes

The deepest hole ever drilled into the Earth — the Kola Superdeep Borehole in northwestern Russia, begun in 1970 and abandoned in 1994 — reached 12.26 kilometers before heat and pressure defeated the drilling equipment. The Earth's radius is 6,371 kilometers. Humanity's deepest physical penetration of the planet reached 0.19% of the way to the center. Everything known about the remaining 99.81% — the layered structure of the mantle, the liquid iron outer core, the solid inner core, the subducted slabs descending through the transition zone, the mantle plumes rising from near the core-mantle boundary, the continent-sized anomalies lurking at 2,900 kilometers depth — has been inferred from seismic waves. Earthquakes, in this sense, are the only X-ray machine capable of imaging a planet's interior.

Seismic tomography is the technique that turns those X-rays into three-dimensional images. Like medical CT scanning, which constructs images of tissue density from thousands of X-ray paths through the body at different angles, seismic tomography constructs images of the Earth's interior from thousands of seismic wave paths recorded at different source-receiver geometries. The source is an earthquake. The receiver is a seismograph. The measurement is the travel time of the wave — how long it took to travel from earthquake to seismograph along a specific path through the Earth. The anomaly is the difference between the observed travel time and the time predicted by a reference Earth model. And from millions of those anomalies, assembled from decades of global seismograph records, the three-dimensional velocity structure of the planet can be reconstructed.

The method has revealed a planet far more structurally complex than the simple layered sphere envisioned by early seismologists. Subducting oceanic plates — cold, fast, and seismically distinct from the surrounding mantle — can be traced descending through the upper mantle and transition zone and, in some cases, pooling at the base of the lower mantle or penetrating into it. Mantle plumes — columns of anomalously hot material rising from deep in the mantle — connect the surface volcanism of Hawaii, Iceland, and Yellowstone to source regions near the core-mantle boundary. And at that boundary, 2,900 kilometers down, continent-sized regions of anomalously slow seismic velocity — the Large Low Shear Velocity Provinces — sit beneath Africa and the Pacific, their origin and composition still debated and their relationship to the surface history of continents and oceans still being worked out.

The Basic Principle: Travel Time as a Probe of Structure

The fundamental observable in seismic tomography is the travel-time residual — the difference between the observed arrival time of a seismic wave at a seismograph station and the arrival time predicted by a spherically symmetric reference Earth model. If a seismic wave travels through a region where the rock is colder and stiffer than the average Earth model — a subducting oceanic slab, for instance — it travels faster than the reference model predicts, and it arrives early. The travel-time residual is negative. If it travels through hotter, softer rock — a mantle plume, or the partially molten root beneath a hotspot — it travels slowly and arrives late. The residual is positive.

A single travel-time residual for a single earthquake-station pair provides almost no useful information about where along the ray path the anomaly occurs — the wave might have been slowed by a small anomaly anywhere along its path. But when millions of residuals from thousands of earthquakes recorded at hundreds of stations are combined, the rays from different source-receiver pairs sample the Earth's interior at many different angles. A region that is consistently fast for all rays that pass through it — regardless of the direction they come from or the station they arrive at — stands out as a coherent velocity anomaly. The tomographic inversion problem is to find the three-dimensional velocity model that best explains the full set of observed travel-time residuals.

Reference Earth Models: The Baseline for Comparison

Every travel-time residual is computed relative to a reference Earth model — a one-dimensional (radially symmetric) velocity profile that describes how P-wave and S-wave velocities vary with depth in an average, spherically symmetric Earth. The most widely used reference model is the Preliminary Reference Earth Model (PREM), published by Dziewonski and Anderson in 1981. PREM was constructed from a global dataset of free oscillation frequencies (the planet's normal modes, excited by great earthquakes), surface wave dispersion measurements, and body wave travel times, and remains the standard reference for most global tomographic studies despite being more than four decades old.

PREM defines the expected travel time for any source-receiver geometry as a function of epicentral distance and depth. Positive residuals (late arrivals) indicate slower-than-average velocity along the ray path; negative residuals (early arrivals) indicate faster-than-average velocity. The amplitudes of these residuals — typically a fraction of a second to a few seconds for teleseismic P-waves — are small relative to the total travel time (600–1,200 seconds for teleseismic distances) but large relative to the measurement precision achievable with modern digital seismographs, which can time arrivals to within 0.01–0.1 seconds.

The Inversion Problem: From Residuals to Images

Translating millions of travel-time residuals into a three-dimensional velocity model is a large-scale linear inverse problem. The Earth is divided into a grid of cells — typically tens to hundreds of thousands of cells in global models — and the velocity anomaly in each cell is treated as an unknown. The observed travel-time residuals are expressed as linear combinations of the velocity anomalies in all the cells the ray passes through (weighted by the time the ray spends in each cell). The system of equations is then solved for the cell velocity anomalies that minimize the misfit between the predicted and observed residuals, subject to regularization constraints that prevent the solution from becoming unstable or geologically implausible.

The practical challenges of this inversion are substantial. Global P-wave tomographic datasets contain tens of millions of measurements; global S-wave datasets contain millions. The system is heavily overdetermined (far more equations than unknowns) but also rank-deficient (some combinations of unknowns are completely unconstrained because no ray paths sample them). The solution requires either iterative algebraic methods or conjugate gradient approaches that avoid ever explicitly constructing the enormous matrix of the full system, instead operating on it implicitly through successive approximations. Modern global tomographic models are produced on supercomputer clusters, with individual runs taking weeks of compute time to converge to a stable solution.

Ray Theory vs. Finite-Frequency Kernels

Classical seismic tomography uses ray theory — the assumption that seismic waves travel along infinitely thin geometric rays, the same approximation used for light rays in optics. In ray theory, a wave is sensitive only to the velocity structure directly along its geometric ray path, and the travel-time sensitivity is uniform along the entire path. This is a good approximation when the velocity anomalies are large compared to the seismic wavelength, but it becomes inaccurate for long-period waves (which have wavelengths of tens to hundreds of kilometers) sampling anomalies of comparable scale.

Finite-frequency tomography, developed by Dahlen, Tromp, and collaborators in the late 1990s and 2000s, improves on ray theory by computing the actual sensitivity of a wave's travel time to velocity perturbations at every point in the medium — a three-dimensional sensitivity kernel (or Fréchet kernel) that accounts for the finite bandwidth of the wave. For a teleseismic P-wave, the sensitivity kernel resembles a hollow banana or donut shape: the wave is most sensitive to structure off the geometric ray path, in the Fresnel zone around it, and least sensitive at the ray itself (the "banana-doughnut" kernel). This counterintuitive result arises from wave diffraction and interference effects that are ignored in ray theory but matter significantly for long-period waves sampling large-scale anomalies at great depths.

Types of Seismic Tomography

Modern seismic tomography encompasses several distinct methodological approaches that use different wave types, different observables, and different mathematical frameworks. Each has particular strengths and weaknesses in terms of depth sensitivity, horizontal resolution, and computational cost.

P-wave Travel-Time Tomography

P-wave tomography uses the travel times of compressional body waves from teleseismic earthquakes, recorded at global seismograph networks. P-waves are the fastest-traveling seismic waves, arriving first at the seismograph and providing the cleanest, least contaminated arrival time measurements. Global P-wave datasets are also the largest: the International Seismological Centre (ISC) bulletin contains hundreds of millions of P-wave arrival time picks accumulated over more than a century of global seismic monitoring. P-wave tomography provides relatively high horizontal resolution (100–300 km in well-sampled regions) and resolves structures from the uppermost mantle to the inner core, but its sensitivity to temperature is indirect — P-wave velocity is less sensitive to temperature than S-wave velocity and is also affected by compositional variations.

S-wave and Surface Wave Tomography

S-wave tomography uses the travel times of shear waves, which are more sensitive to temperature than P-waves (because the shear modulus is more temperature-dependent than the bulk modulus) and cannot propagate through the liquid outer core. S-wave models therefore provide the clearest thermal images of the mantle, with the largest anomalies at hotspots (slow) and subduction zones (fast) being most prominent in S-wave models rather than P-wave models. Surface wave tomography uses the dispersion of Love and Rayleigh waves — the variation of their phase and group velocity with period — to constrain shear velocity structure from the crust to the upper mantle (depths of 0–400 km). Long-period surface waves sample the deepest upper mantle and transition zone, while short-period surface waves constrain the crust and uppermost mantle.

Ambient Noise Tomography

A transformative development in the 2000s was the recognition that the coherent seismic noise always present in seismograph records — generated by ocean wave interaction with the seafloor, wind loading of the crust, and anthropogenic sources — could be used as a source for tomography in the absence of earthquakes. Cross-correlating the ambient seismic noise between pairs of seismograph stations extracts a signal that approximates the Green's function between the two stations — essentially the response that would be recorded at one station if the other were an earthquake source. This technique, ambient noise tomography, opened up tomographic imaging to scales and depth ranges that were previously inaccessible because of poor earthquake coverage at short periods.

Ambient noise tomography has been particularly revolutionary for imaging crustal and upper mantle structure, where short-period surface waves (5–40 seconds period) provide sensitivity to structures at depths of 10–200 km with horizontal resolutions of 10–50 km — far higher than is achievable with teleseismic body waves in the same depth range. The technique requires only a dense network of seismographs with long continuous records — no earthquakes needed — making it applicable in regions of low seismicity and particularly powerful when combined with USArray or similar large portable network deployments.

Full Waveform Inversion

The most computationally demanding and potentially most powerful approach to seismic tomography is full waveform inversion (FWI) — using not just the travel times of specific wave arrivals but the entire recorded seismogram, including all reflected, refracted, converted, and scattered wave phases. FWI requires solving the full seismic wave equation numerically for each earthquake in the dataset, computing synthetic seismograms for the current model, comparing them to the observed seismograms, computing the gradient of the misfit with respect to the model parameters using the adjoint method, and iterating toward a model that fits the full waveform character of the observations.

FWI was originally developed for exploration seismology — imaging subsurface structure in oil and gas exploration at kilometer scales — but has been progressively scaled up to regional and global applications as computational power has grown. Global FWI models, now regularly published by groups at ETH Zurich, Princeton, the University of Texas, and elsewhere, achieve significantly higher resolution than classical travel-time models and reveal structural details — slab edges, plume conduits, lower mantle heterogeneity — that are smeared or missed in ray-theory inversions. A global FWI using data from thousands of earthquakes and hundreds of stations requires billions of floating-point operations per iteration on petascale computing systems, representing one of the largest computational challenges in solid earth geophysics.

Major Discoveries: What Tomography Has Revealed

Over four decades of global seismic tomography — from Dziewonski's pioneering 1984 model to modern high-resolution FWI models — the technique has produced a series of discoveries that have fundamentally reshaped our understanding of mantle dynamics, plate tectonic history, and Earth's deep structure.

Subducting Slabs: Cold Fingers in the Mantle

The most visually striking and scientifically consequential discovery of seismic tomography is the imaging of subducting oceanic slabs as high-velocity anomalies extending deep into the mantle. Cold oceanic lithosphere — several hundred degrees cooler than the surrounding mantle — has higher seismic velocity than the ambient mantle, making it appear as a blue (fast) anomaly in tomographic models color-coded by velocity deviation from the reference Earth.

Early tomographic models revealed that slabs do not simply sink to a fixed depth and stop — they descend through the upper mantle and in many cases penetrate the 660-kilometer discontinuity (the boundary between the upper and lower mantle) and continue into the lower mantle. In other cases, slabs appear to stagnate at the 660 km boundary, spreading laterally across the transition zone in a flattened geometry before eventually sinking further. The difference between penetrating and stagnating slabs appears to reflect the competition between the viscosity jump at the 660 km phase transition and the kinetic energy of the descending slab — fast-subducting, old, dense slabs tend to penetrate, while slowly subducting, young slabs tend to stagnate.

Some tomographic models resolve slab remnants — "slab graveyards" — in the deep lower mantle, at depths of 1,500–2,800 km, that correspond to ancient subduction zones no longer active at the surface. The Farallon plate, which largely subducted beneath North America between 200 and 30 million years ago, is imaged as a massive high-velocity anomaly in the lower mantle beneath the eastern United States — a fossil record of ancient tectonics preserved in the thermal structure of the lower mantle.

Mantle Plumes: Hot Columns Rising from the Deep

Hotspot volcanoes — Hawaii, Iceland, Yellowstone, the Galapagos, Réunion — have long been suspected to overlie mantle plumes: columns of anomalously hot material rising from deep in the mantle and impinging on the base of the lithosphere. Seismic tomography has provided the most direct evidence that these plumes exist, but imaging them has proven far more difficult than imaging cold subducting slabs, for a simple reason: plumes are hot, and hot rock is slow. Slow anomalies are harder to resolve than fast ones because the reference Earth model already includes lateral velocity variations from continental roots and ocean basin structure that partially mask the plume signal.

The best-imaged plume system is Iceland — where a combination of dense regional seismograph coverage, proximity to the Mid-Atlantic Ridge, and the exceptional depth extent of the anomaly has allowed imaging of a low-velocity conduit extending from the Iceland surface to depths of at least 400 km in most models, and possibly to 650 km or deeper in the most recent high-resolution FWI results. The Hawaii plume is imaged as a slow anomaly to depths of 500–1,500 km in different models — the range reflecting genuine model uncertainty rather than inconsistency in the observations. Whether the Hawaii plume originates at the core-mantle boundary (~2,900 km) or in the mid-mantle remains one of the most actively debated questions in mantle dynamics.

The Large Low Shear Velocity Provinces

Among the most striking and puzzling discoveries of global S-wave tomography are two enormous regions of anomalously slow shear wave velocity at the base of the mantle — the Large Low Shear Velocity Provinces, or LLSVPs. One sits beneath Africa, the other beneath the central Pacific, each roughly continent-sized in lateral extent (several thousand kilometers across) and rising 500–1,000 km above the core-mantle boundary. They appear in virtually all global S-wave tomographic models and are among the most robust large-scale features in deep Earth imaging.

The LLSVPs are slower than the surrounding mantle by 2–3% in shear wave velocity — a substantial anomaly at depth where the reference velocity is ~7 km/s. This slow velocity could reflect anomalously high temperature (partial melt, perhaps, or simply hotter rock), anomalous composition (iron-enriched material left over from an ancient magma ocean or subducted oceanic crust accumulated over billions of years), or both. The two leading compositional hypotheses — primordial undegassed mantle material surviving since Earth's formation, and accumulated subducted oceanic crust — make very different predictions about the long-term thermochemical evolution of the mantle and have very different implications for Earth's differentiation history.

Ultra-Low Velocity Zones at the Core-Mantle Boundary

At finer scales, seismic tomography and related waveform analysis techniques have identified Ultra-Low Velocity Zones (ULVZs) — patches at the very base of the mantle, directly atop the liquid iron outer core, where shear wave velocities drop by 10–30% and P-wave velocities drop by 5–10% relative to the surrounding lower mantle. ULVZs are typically hundreds of kilometers in lateral extent and a few tens of kilometers thick — thin slivers of extraordinarily anomalous material at the planet's most dramatic compositional boundary.

The velocity drops in ULVZs are so large that temperature alone cannot explain them — no geologically plausible temperature increase could reduce seismic velocity by 10–30%. The leading hypotheses involve partial melting of the base of the mantle (the lowering of the melting point by iron enrichment could allow a thin layer of silicate melt to persist just above the core) or iron-rich solid phases whose unusual crystal structure produces anomalously low seismic velocities. ULVZs appear to be concentrated at the margins of LLSVPs, suggesting a genetic relationship between these two deep-mantle structures that is still being worked out.

The Inner Core: Anisotropy and the Eastern-Western Hemisphere Dichotomy

The solid iron inner core — 1,221 km in radius, surrounded by the liquid outer core — reveals unexpected complexity in seismic tomographic and waveform studies. The inner core is seismically anisotropic: P-waves traveling parallel to Earth's rotation axis (north-south direction in the deep interior) travel roughly 3% faster than P-waves traveling in the equatorial plane. This anisotropy is thought to reflect the alignment of iron crystals with the stress field of the inner core, either from preferential crystal growth or from slow plastic deformation driven by convection or magnetic torques.

More surprising is the observation that the inner core is divided into two hemispheres with distinct seismic properties: the western hemisphere (beneath the Americas) has lower P-wave velocity and less anisotropy than the eastern hemisphere (beneath Asia and the Indian Ocean). This asymmetry — robust in multiple independent datasets and models — has been interpreted as evidence for asymmetric inner core growth: the eastern hemisphere grows faster (solidification of the liquid outer core adds to the inner core preferentially on one side), while the western hemisphere loses material to the liquid outer core on the other side — a net translation of the inner core relative to the rotation axis over geological time.

Resolution and Limitations: What Tomography Cannot See

Seismic tomography has transformed our understanding of Earth's interior, but it has fundamental limitations that are important to understand when interpreting tomographic models. Every tomographic image is a blurred, incomplete, and potentially artifact-contaminated approximation of the true velocity structure — and honest interpretation requires knowing where the model is reliable and where it is not.

Uneven Data Coverage

The global distribution of seismograph stations is far from uniform. The vast majority of stations are on continents, particularly in North America, Europe, Japan, and Australia. The ocean basins — covering 70% of Earth's surface — are sparsely instrumented, and the southern hemisphere has far fewer stations than the northern hemisphere. The distribution of earthquakes is similarly uneven: most seismicity occurs at plate boundaries, leaving large volumes of the mantle under stable continental interiors and ocean basins poorly sampled by crossing ray paths.

The consequence is that tomographic resolution is highly variable across the globe. Well-sampled regions like the western Pacific subduction zones, the western United States (sampled by USArray), and Europe (sampled by AlpArray and related networks) may achieve horizontal resolutions of 100–200 km in the upper mantle. Poorly sampled regions like the South Atlantic, the central Pacific Ocean, and the Antarctic mantle may have resolutions of 500–1,000 km or worse — adequate for continent-scale structures but insufficient for plume conduits or slab edges.

Trade-offs Between Structure and Earthquake Location

A fundamental ambiguity in travel-time tomography is the trade-off between the velocity model and the earthquake source parameters. A late P-wave arrival could mean the ray passed through slow material, or it could mean the earthquake occurred later than the assumed origin time, or it could mean the earthquake occurred deeper than assumed, or in a slightly different location. Tomographic inversions address this trade-off by jointly solving for both the velocity model and the earthquake locations (and sometimes origin times) simultaneously, but the coupling between the two means that errors in one propagate into errors in the other — a systematic bias in earthquake locations will produce systematic artifacts in the velocity model, and vice versa.

The Next Generation: Ocean Bottom Seismographs and Exascale Computing

Two frontiers are currently expanding the reach of seismic tomography beyond what was achievable with land-based networks and conventional computing.

The deployment of ocean bottom seismograph (OBS) networks is beginning to fill the critical gap in global station coverage over the ocean basins. Programs like the OBS Instrument Pool and the planned expansion of permanent seafloor networks in the Pacific and Atlantic will dramatically improve resolution in the oceanic mantle — precisely the regions where many key questions about plume conduits, slab dynamics, and deep mantle structure remain unresolved because of data gaps. The PLUME experiment off Hawaii and the NoMelt experiment in the central Pacific have already demonstrated that OBS arrays can resolve mantle structure beneath ocean basins at resolutions previously achievable only under continents.

On the computational side, the move to exascale computing — systems capable of 10¹⁸ floating-point operations per second — is enabling global FWI at resolutions and with datasets that were computationally impractical just a decade ago. The SPECFEM3D spectral element method, developed at Princeton and widely used in global seismology, can now simulate seismic wave propagation through a fully heterogeneous 3D Earth at periods as short as 8–10 seconds on modern supercomputers, enabling FWI that is sensitive to structures of 100–200 km scale throughout the mantle. Within the next decade, global FWI at 5-second periods — sensitive to features as small as 50–75 km anywhere in the mantle — is computationally within reach.

Tomography and Earthquake Hazard: The Connection to the Surface

Seismic tomography might appear to be pure basic science — remote from the practical concerns of earthquake hazard and building codes. In several important ways, it is not.

At the regional scale, tomographic models of the crust and upper mantle directly inform seismic hazard assessment by mapping velocity structure that controls ground motion amplification and attenuation. Basin geometry — the three-dimensional shape of a sedimentary basin and the velocity contrast at its boundaries — determined from seismic reflection profiles combined with ambient noise tomography, is a primary input to 3D ground motion simulations used for scenario earthquake planning in cities like Los Angeles, Seattle, Tokyo, and Istanbul. The SCEC (Southern California Earthquake Center) Community Velocity Model — a detailed 3D seismic velocity model of Southern California assembled from decades of tomographic, reflection, and refraction data — is the foundation for the physics-based ground motion simulations used in the ShakeOut scenario planning exercises that have trained millions of Californians in earthquake preparedness.

At the global scale, tomographic imaging of subducting slabs constrains the geometry of the plate interface at depth — information that feeds directly into estimates of where the megathrust is locked and how large future earthquakes might be. The imaging of flat-slab subduction geometry beneath the Andes and beneath parts of North America helps explain the distribution of intraslab seismicity and crustal deformation far from the trench, allowing more accurate seismic hazard assessment for inland cities that are far from the subduction zone trench but directly above the flat-subducting slab.

Conclusion

Seismic tomography has given humanity its only meaningful view of the planet's interior — not through a borehole that reaches 0.2% of the way to the center, but through the accumulated X-ray of ten thousand earthquakes recorded at ten thousand seismographs, their travel times measured to fractions of a second and inverted by algorithms consuming weeks of supercomputer time into images of a churning, heterogeneous mantle that bears no resemblance to the simple onion-shell Earth of introductory textbooks.

The images have earned their authority through reproducibility: the major features — the Farallon slab graveyard, the LLSVPs, the Iceland plume, the stagnating slabs of the western Pacific — appear in models produced by independent groups using independent datasets and independent methods. They are as real as anything in deep Earth science. What they mean — how the LLSVPs formed, whether the Hawaii plume reaches the core-mantle boundary, why the inner core is hemispherically asymmetric — remains actively contested, which is precisely where science should be at the frontier of what it can see.

Every earthquake is an opportunity. Every seismograph is an eye. The planet is constantly illuminating itself from within, and seismic tomography is the technology that turns that illumination into knowledge about a world that no human being will ever directly observe.