AI Maps Six Hidden Structures in Earth's Core

What a machine learning sweep of three decades of earthquakes really found at the base of the mantle, and how much of it we should believe.

A single study just multiplied the world's catalog of a rare seismic signal by roughly ten, from a few thousand entries to 174,929, and used the result to trace six continuous bands of hidden structure across the boundary between Earth's rocky mantle and its molten iron core. The signal is called a PKP precursor, a faint echo that arrives seconds ahead of a stronger core-piercing wave, and it is normally so weak that it drowns in noise. The work, led by Yurui Guan at the Institute of Geology and Geophysics, Chinese Academy of Sciences, trained a deep neural network to hunt those echoes across more than 2 million seismic recordings gathered worldwide between 1990 and 2024 - Journal of Geophysical Research: Solid Earth.

The framing that traveled fastest, that an AI scanned about thirty years of earthquake data and found six previously hidden structures at Earth's core, is close to right and worth stating precisely. The structures sit at the core-mantle boundary (CMB), roughly 2,890 kilometres down, not inside the core proper. They are not six freshly discovered lumps of rock. They are six coherent scattering regions that earlier, sparser data had shown only as scattered patches or had never sampled at all - phys.org. The novelty is one of scale and resolution, not of a single object suddenly appearing where nothing was known before.

This is the kind of result Frontier exists to weigh rather than cheer. The headline is genuinely large, the method is genuinely new for this problem, and the interpretation is genuinely uncertain, all at once. This guide separates those three things. It explains what the study did, walks through Earth's interior from first principles so the finding is legible to a non-specialist, examines how the deep learning worked and where its confidence comes from, looks hard at what seismic imaging cannot tell you, and closes with Frontier's own assessment, a current Frontier Score of 72 on our 0-100 scale of evidence, impact, and novelty. You can see where that sits among the year's results on the Frontier Index.

Contents

1. What the study actually found
2. Earth from the inside out: a first-principles tour
3. The only probe we have: seismic waves and PKP precursors
4. What the deep learning actually did
5. The six bands, examined
6. Why the scale is the breakthrough, not the pictures
7. What seismic imaging cannot tell you
8. Frontier's assessment: scoring a deep-Earth discovery
9. What to watch next

A note on reading this. The sections build on each other but each stands alone, so if you already know why an S-wave cannot cross the outer core you can skip ahead to the method. Where a number is firm, it is stated plainly and linked to its source. Where it is a model-dependent estimate, and much of deep-Earth science is, that uncertainty is named rather than rounded away. The single most important habit for reading any tomographic result is to keep asking what the data can support versus what the interpretation adds, and this guide tries to model that habit rather than just recommend it.

1. What the study actually found

The paper is titled "Global Distribution of PKP Precursors Derived From Three Decades of Seismic Data With Deep Learning," published in the American Geophysical Union's Journal of Geophysical Research: Solid Earth with the DOI 10.1029/2025JB033195. Its central product is a catalog: 174,929 individual precursor measurements, extracted by a trained classifier from over 2 million waveforms produced by roughly 5,000 earthquakes of magnitude 6.0 or greater recorded on seismometers around the world - Yahoo News. That count is described by the authors as about ten times the size of all previous PKP-precursor catalogs combined - ScienceAlert.

From that catalog the team built two probability maps. As the paper's own summary puts it, one map shows "how consistently precursors appear" across the globe, and the other shows "where strong scatterers are most likely located" - phys.org. Reading those maps together, the authors identify six continuous bands of elevated scattering at the core-mantle boundary. Press coverage places them beneath the North Atlantic, northern Eurasia, the South Atlantic, southern Africa, the Pacific, and the Antarctic region, spanning both well-studied zones and areas that prior surveys had barely touched - The Watchers.

The reason this matters is that a PKP precursor is one of the few signals capable of resolving small-scale roughness right at the CMB, and until now the catalog was too thin and too geographically lopsided to see continent-scale pattern in it. A larger, more evenly distributed catalog turns a handful of local case studies into something closer to a global survey. The authors are explicit that this is a foundation for future work rather than a finished map, noting that as the catalog "continues to expand, its high-resolution spatiotemporal coverage will provide a solid foundation for characterizing the spatial and temporal variations in CMB heterogeneity" - phys.org.

The table below summarizes the six bands. One column deserves a word of explanation before you read it. Every band sits at essentially the same depth, the core-mantle boundary near 2,890 km, because that is where this particular signal is generated. What distinguishes the bands is location and the strength and consistency of scattering, not depth. The numbering here is this guide's own, added for reference; the paper describes them as continuous regions of elevated scattering probability, not as six discrete, individually named objects.

Read that table as a set of hypotheses about location, not confirmed compositions. The study tells you where scattering is strong and consistent. It does not, on its own, tell you what the scatterers are made of, and Section 5 unpacks the candidate readings with the appropriate caution. The value here is that six coherent targets now exist for follow-up work, several of them in places nobody had a global picture of before.

2. Earth from the inside out: a first-principles tour

To understand why a faint echo at 2,890 km is worth a decade of archived earthquakes, you have to know what is down there and why it is so hard to see. Earth is layered like an onion, and the layers were not assembled but sorted by density while the young planet was molten, the heaviest material sinking to the center and the lightest floating to the top. That single fact, gravitational differentiation, explains the whole structure: a thin rocky crust, a thick rocky mantle, a liquid outer core of iron and nickel, and a solid inner core of nearly pure iron at the center - Structure of Earth.

The numbers are worth fixing in mind because the deep-Earth story is really a story about depth. The crust reaches down only tens of kilometres. The mantle, solid rock that nonetheless flows over geological time, extends to about 2,890 km. Below that the outer core runs from 2,890 to roughly 5,150 km and is genuinely liquid. The inner core, a ball with a radius of about 1,220 km, occupies the space from 5,150 km to Earth's center at 6,371 km, and although it is hotter than the outer core it is solid, held rigid by the immense pressure - Structure of Earth. The chart below puts those boundaries on one scale.

The core-mantle boundary is the most violent interface on the planet, and the reason this study camps out there. It separates solid silicate rock from liquid metal, and across it the temperature jumps by around 1,000 degrees Celsius into a molten outer core at roughly 5,500 degrees Celsius - phys.org. That is a larger temperature contrast than exists between Earth's surface and the base of the crust. Rock behaves very differently on either side, one part rigid enough to fracture in earthquakes near the surface and to conduct certain waves at depth, the other a fluid that cannot support the same waves at all. This mismatch is what makes the boundary so seismically loud and so physically important: it is where the mantle dumps heat into the core and where the core drives the mantle's slow churn in return.

The lowest few hundred kilometres of the mantle, just above the boundary, carry their own name, the D-double-prime layer (written D″). It is not a uniform shell. It is a rubble zone of structure at every scale, from continent-sized provinces down to blobs a few kilometres across, and it is thought to be where sinking tectonic plates finally come to rest and where the roots of rising mantle plumes are anchored - Generation of Reflections and PKP Precursors From a Scattering Layer in D″. Two families of structure dominate the D″ layer and the CMB, and telling them apart is central to reading the new study.

The first family is the large low-shear-velocity provinces, or LLVPs, sometimes called "blobs." There are two of them, one beneath the central Pacific and one beneath Africa and the southern Atlantic, each larger than a continent, reaching up to 900 km in height, and slowing shear waves by only about 1 to 3 percent relative to normal mantle - Large low-shear-velocity provinces. They are enormous but subtle. The second family is the ultra-low velocity zones, or ULVZs, which are the opposite: small patches only tens of kilometres thick, but dramatic, dropping compressional wave speed by roughly 5 to 25 percent and shear wave speed by as much as 10 to 45 percent - Ultra-low velocity zone. The next chart shows how different those magnitudes are, because the contrast is the whole point.

Why this matters for the reader is simple: the new study does not measure these velocity drops directly. It detects scattering, the fingerprint of sharp, small-scale roughness, which is exactly what the edges of ULVZs and the debris of sunken slabs are expected to produce. So when the six bands are discussed, the honest translation is "here the boundary is rough and heterogeneous," with ULVZs and slab fragments as the leading, but unproven, explanations. How to apply that: whenever a headline says an AI "found structures at the core," ask whether the instrument measured composition or measured roughness. Here it measured roughness, which is real and useful and not the same thing as knowing what the rock is.

3. The only probe we have: seismic waves and PKP precursors

Nobody has ever sampled the deep mantle, let alone the core. The deepest hole ever drilled, the Kola Superdeep Borehole, reached about 12 km, roughly 0.2 percent of the way to the center. Everything else is inference, and almost all of it rides on seismic waves, the vibrations that large earthquakes send coursing through the whole planet. When a magnitude-6 quake ruptures, it acts like a flashbulb going off inside an opaque object, and the thousands of seismometers on the surface are the film. The way those waves speed up, slow down, bend, bounce, and split as they pass through the interior encodes the structure they passed through.

Two wave types do most of the work, and the difference between them is one of the most useful facts in all of Earth science. P-waves (primary, or compressional) push and pull the rock in the direction they travel, like sound, and they can move through solids and liquids alike. S-waves (secondary, or shear) wobble the rock sideways, and they cannot travel through a liquid at all, because a fluid has no rigidity to spring back a sideways push. That single property is how we know the outer core is liquid: S-waves from a distant quake simply fail to arrive on the far side of the planet, leaving a shadow zone that can only be explained by a molten layer swallowing them - Seismic tomography: A window into deep Earth.

Seismologists label the paths waves take with a compact notation. A P is a compressional wave in the mantle, a K is a compressional wave through the liquid outer core (from the German Kern, core), and an I is a wave through the solid inner core. So a PKP wave is one that leaves an earthquake as a P-wave, dives all the way down and refracts through the outer core as a K, then climbs back out as a P to reach a distant station. It is the workhorse phase for studying the deepest Earth because it necessarily passes through the core region, and it shows up clearly at source-to-station distances of roughly 125 to 145 degrees around the globe.

The precursor is the subtle part. Along with the main PKP wave that refracts cleanly through the core, a tiny fraction of the wave's energy hits small irregularities near the core-mantle boundary and scatters, taking a slightly different, faster geometric path to the same station. That scattered energy arrives a few seconds before the main PKP phase, which is why it is called a precursor. The interpretation that precursors are scattered energy from CMB heterogeneity goes back decades - Seismic Wave Scattering near the Core-Mantle Boundary: a New Interpretation of Precursors to PKP, and it was placed on a firm scattering-theory footing shortly after - Evidence for scattering of seismic PKP waves near the mantle-core boundary. The diagram below shows why the geometry produces an early arrival.

The practical problem, and the reason this was a job for machine learning, is that precursors are faint and inconsistent. They ride in the noisy window ahead of a much larger wave, their amplitude varies from quake to quake and station to station, and a single station cannot pin down where along the boundary the scattering happened. Modern work has tied precursor strength to specific small-scale structure in the D″ layer, including scatterers only about 10 km across with velocity perturbations near 5 percent - Generation of Reflections and PKP Precursors From a Scattering Layer in D″, and has begun linking precursor observations to the strange anomalies mapped elsewhere in the deep mantle - ScienceDaily. But finding each precursor by hand, across millions of records, was never realistic. That is the bottleneck the new study attacks.

4. What the deep learning actually did

It helps to be precise about the sense in which this is an "AI" result, because the phrase covers everything from a spam filter to a chatbot. The system here is a supervised deep-learning classifier, a neural network trained to answer a narrow question, "does this waveform contain a PKP precursor, and if so where," rather than a general-purpose model. The authors describe integrating deep learning with iterative manual optimization: they trained the network on human-labeled examples of precursors, ran it across the archive, then had experts check and correct batches of its output, feeding the corrections back to sharpen the next pass - Yahoo News. That loop is the important design choice, and it is worth dwelling on.

The reason a classifier beats a human here is throughput, and the reason human-in-the-loop review is non-negotiable is trust. A network can scan 2 million waveforms in the time it takes a graduate student to work through a few hundred, but a network trained only once will confidently reproduce its own systematic errors two million times over. By interleaving automated scale with expert correction, the team keeps the throughput while catching the failure modes that pure automation would bake in. The output is not a raw pile of detections but the two probability maps described earlier, one for how reliably precursors appear in a region and one for where the strongest scatterers most likely sit - phys.org.

The scale of the jump is the single most concrete claim in the paper, and it is the kind of number worth visualizing rather than asserting. Prior to this work, the global stock of catalogued PKP precursors accumulated over decades of painstaking manual and semi-automated analysis. The new catalog of 174,929 detections is described as roughly ten times that entire prior stock combined, which implies earlier catalogs summed to something on the order of 17,000 to 18,000 entries - ScienceAlert. The chart shows the contrast, with the prior figure marked approximate because it is inferred from the stated multiple rather than tallied directly.

This design is not unique to Guan's team, which is a point in its favor rather than against it: independent groups converging on machine methods for the same signal is how a technique matures. A separate 2025 effort applied a graph neural network to the global search for PKP precursors, explicitly aimed at scatterers in the lowermost mantle - Global Search of PKP Precursors With Graph Neural Network. The broader field of machine-learning seismology has been building toward exactly this for years, moving from hand-tuned detectors to trained networks that pick, classify, and locate seismic phases at scale, a shift chronicled in recent reviews of the discipline - The evolution of seismic tomography in earth sciences. What Guan's team contributes is not the invention of the approach but its application at global scale to a signal that had defeated manual cataloging, and that distinction matters when we score novelty later.

5. The six bands, examined

With the method understood, the six bands can be read more carefully. The first thing to hold onto is that the study maps where scattering is strong, and everything past that is inference layered on top of the data. The scatterers themselves are consistent with several deep-mantle structures we already know exist, and the geography of the six bands lines up suggestively, but not conclusively, with those structures. The authors and the coverage float three main candidate origins, and each is worth weighing on its own evidence.

The strongest candidate is subducted slab material. When an oceanic plate dives into the mantle at a subduction zone, it does not vanish; cold, dense slab sinks over tens of millions of years and can eventually pile up at the core-mantle boundary, creating exactly the kind of sharp compositional and thermal contrasts that scatter seismic waves. Several of the six bands, notably those beneath northern Eurasia and around the Pacific, sit above regions with long subduction histories, making a "slab graveyard" reading plausible. This mechanism is also invoked for the ultra-low velocity zones, with recent work arguing that ULVZs can be generated in downwelling slabs as they reach the boundary - Evidence for Ultra-Low Velocity Zone Genesis in Downwelling Subducted Slabs, and mega-scale ULVZs have been tied directly to slab control beneath the North Pacific - Slab control on the mega-sized North Pacific ultra-low velocity zone.

The second candidate is partial melt, the ULVZ interpretation in its purest form. If pockets of rock at the boundary are hot enough to be partly molten, or are enriched in dense iron-bearing phases, they would both slow waves dramatically and present rough, scattering edges. This reading connects the new scattering bands to the older, painstaking case studies of individual ULVZs, including the landmark discovery of a mega-ULVZ beneath the Marquesas Islands that itself came out of an early machine-learning analysis (more on that in Section 6). The bands beneath southern Africa, the South Atlantic, and the Pacific overlap the margins of the two great LLVPs, and ULVZs are frequently found clustered along those margins, so a melt-related or iron-enriched origin is credible for at least part of the pattern - A review of large low shear velocity provinces and ultra low velocity zones.

The third candidate is the most speculative and should be flagged as such: primordial material, possibly including debris from the protoplanet Theia, the Mars-sized body whose giant impact is thought to have formed the Moon. Some coverage of the study raised Theia's remnants as a possible source of deep heterogeneity - Research Snipers, and the idea has genuine currency in the wider literature about the LLVPs. But it is an interpretation reaching well beyond what a scattering map can establish, and the honest position is that this study provides no direct evidence for it. Naming it is fine; leaning on it would be hype.

What ties the three candidates together, and what the reader should take away, is that they are not mutually exclusive and none is settled. The data say the boundary is rough in six coherent regions. The geography is consistent with slabs in some places and LLVP-margin melt in others, and frankly under-constrained beneath the newly sampled Antarctic region. The paper's real contribution to this debate is not a verdict but a map of where to look, six well-defined targets, several of them global firsts, that future multi-station and waveform-modeling studies can now attack with the location already narrowed down.

6. Why the scale is the breakthrough, not the pictures

It is tempting to treat the six bands as the discovery and the AI as the tool that produced them. From a first-principles view that has it backwards. The durable contribution is the catalog and the method that made it, because a ten-fold expansion of a scarce signal changes what questions can be asked, while any single map is provisional and will be revised. The right frame is not "AI found six structures" but "AI turned an intractable manual task into a repeatable global survey, and the first survey shows six bands."

This becomes clearer against the history of machine learning in deep-Earth seismology, which is short but instructive. In 2020, a team led by Doyeon Kim at the University of Maryland applied a manifold-learning algorithm called the Sequencer, originally built for astrophysics, to thousands of seismograms of waves diffracting along the core-mantle boundary. It recovered the known plume root beneath Hawaii and, more strikingly, revealed a previously unrecognized mega-ULVZ beneath the Marquesas Islands, publishing the result in Science - Sequencing seismograms: A panoptic view of scattering in the core-mantle boundary region. The Maryland group framed the lesson in words that read as a preview of the 2026 work: approaches "flexible enough to detect robust patterns with little to no user supervision can reveal distinctive insights into the deep Earth" - University of Maryland.

The 2020 study was a proof of concept across a single region, the Pacific. The 2026 study is the global-survey version of the same idea, and the jump from one to the other is a jump in scope, not in kind. That is genuine progress, but it is progress along an established line, which is exactly the sort of distinction an evidence-based score has to make rather than blur. A parallel example of squeezing new structure out of old signals comes from the inner core, where a different technique, stacking faint waves that reverberate up to five times through Earth's center, strengthened the case for a distinct innermost inner core with its own anisotropy - Nature Communications. That work relied on advanced signal processing rather than deep learning, so it is a cousin rather than a twin, but it points at the same shift: the frontier increasingly lies in mining decades of archived recordings with new algorithms, not only in deploying new instruments.

Why this matters for how you value the result: the six bands could be revised, merged, or split by the next study without touching the core achievement, because the achievement is the ability to generate and regenerate such maps as the archive grows and the network improves. How to apply that judgment: when you read that an AI made a discovery in science, look past the specific finding to whether a reusable capability was created. Here one was, and that is the part most likely to still matter in five years. We placed this result in context alongside the year's other AI-accelerated science in our scored ranking of 2026's breakthroughs, where the pattern of machine methods reopening old datasets recurs across fields.

7. What seismic imaging cannot tell you

Every map of the deep Earth is the answer to an inverse problem, and inverse problems have a built-in humility that gets lost in headlines. The forward problem, "given a structure, what signals arrive," has one answer. The inverse problem this study lives in, "given the signals, what structure produced them," generally has many. This is the property seismologists call non-uniqueness: multiple different Earth models can fit the same observations equally well within measurement error - The evolution of seismic tomography in earth sciences. A scattering map is a best-supported reconstruction, not a photograph, and the distinction is not pedantic.

The study's own authors are commendably direct about the limits, which is one reason it scores as well as it does on evidence. They stress that their maps show "probable scattering regions, not exact outlines," and that a single seismic station cannot precisely locate a scatterer from a precursor alone - phys.org. Location comes from aggregating many observations, which is powerful but blurs fine detail. And a scattering detection reports that something is rough at the boundary; it does not report the rough thing's composition, temperature, or origin, which is why Section 5's candidate readings stay candidates.

Two further limits are structural rather than incidental, and they shape how much of the "global" picture to trust. The first is uneven coverage. Earthquakes and seismometers are not spread evenly over the planet; both cluster around plate boundaries and populated, instrumented regions, leaving vast areas, much of the southern hemisphere's ocean floor, the polar regions, thinly sampled. Even a ten-fold larger catalog inherits that geography, so confidence in the six bands is not uniform, and the newly resolved Antarctic band rests on sparser support than the well-instrumented northern ones. The general point, that ray paths sample the Earth unevenly and leave large volumes poorly constrained, producing smearing and potential artifacts, is a standing feature of tomographic imaging - The evolution of seismic tomography in earth sciences.

The second is a subtler risk specific to machine methods: a trained classifier can learn the biases of its training data and its coverage. If precursors were easier to label in certain distance ranges or noise conditions, the network may over- or under-report in ways that mimic real structure. The human-in-the-loop design is precisely the guard against this, and it is a good one, but it does not eliminate the concern; it manages it. A veteran of this field, commenting on an earlier core-mantle study, put the epistemic situation memorably: "we are always up against the issue of knowing if our favoured solution model is [both] unique and the real earth" - Physics World. That caution applies here in full. None of it makes the result weak. It makes the result honest, which is a different and more valuable thing.

8. Frontier's assessment: scoring a deep-Earth discovery

Frontier scores every breakthrough on three pillars held deliberately apart, so that a result can be strong on one and weak on another without the number hiding it. Evidence asks whether the finding is real and independently corroborated. Impact asks whether it moves its field and beyond. Novelty asks whether it combines ideas in a genuinely new way. This deep-Earth result carries a current Frontier Score of 72 on the 0-100 scale, and it is worth walking through why each pillar lands where it does, because the reasoning is more useful than the number.

On evidence, the result is strong for a deep-Earth study, which is a category where certainty is scarce by nature. It is published in a rigorous AGU journal after peer review, the central claim (a ten-fold larger precursor catalog) is a countable, reproducible fact rather than an interpretation, and the authors bound their own confidence carefully, distinguishing probable regions from exact outlines. It loses points only where the field forces it to: the non-uniqueness of the inverse problem and the uneven global coverage mean the six-band map, as opposed to the catalog, is provisional. That is a ceiling imposed by seismology, not a flaw in the work.

On impact, the honest reading is high potential, not yet realized. A global precursor catalog is a new dataset for a whole subfield, and it hands researchers six concrete targets, several in previously blank regions. But its consequences are upstream and scientific: it reshapes how the base of the mantle will be studied, rather than changing anything a non-specialist will feel soon. Compared with a result that alters clinical practice or industrial capability, its impact is real but narrower and slower, and the score reflects that. On novelty, the method is a meaningful advance in application rather than a first-of-its-kind invention, given the 2020 Sequencer precedent and the parallel graph-neural-network work; the novelty lies in global scale and completeness, which is genuine but incremental. Blend a strong-but-capped evidence pillar with high-latent-but-unrealized impact and solid-but-not-unprecedented novelty, and 72 is where the arithmetic settles, comfortably above the middle of a year whose top results cluster near 70. You can inspect the full methodology and every underlying number in Frontier's open dataset.

It is worth naming what a 72 communicates in Frontier's grammar. It is not a claim that this is the most important thing that happened in science this year. It is a claim that, weighed against public evidence signals, this is a well-corroborated, field-shaping, methodologically fresh result whose main uncertainties are inherent to its discipline rather than self-inflicted. That is a high compliment on our scale, and it sits deliberately short of the eighties, which we reserve for results that are ironclad, broadly consequential, and truly without precedent all at once. This is two of those three, strongly, and the score says exactly that.

9. What to watch next

The most telling signal over the next year will be whether the six bands survive independent scrutiny. Because the catalog is public in spirit and the method is reproducible, other groups can test the bands with different techniques, multi-station arrays, full-waveform modeling, complementary phases like ScP that probe ULVZs directly. If several bands hold up under methods that do not share this study's biases, confidence climbs and the evidence score with it. If some dissolve into coverage artifacts, that too is a finding, and precisely the kind of correction an evidence-based index is built to absorb rather than resist.

The second thing to watch is what the bands are made of. Scattering told us where the boundary is rough; the next step is composition, and that will come from combining precursor maps with mineral-physics experiments and with the LLVP and ULVZ literature that already brackets the candidates. Expect the slab-graveyard and partial-melt readings to be tested band by band, and expect the more exotic primordial-material ideas to remain interesting and unproven for a good while yet. The broader arc, machine learning reopening decades of archived recordings to ask questions the original analysts could not, is the pattern to track across the deep-Earth sciences and beyond. It is the same engine now surfacing results in biology and imaging, from an AI-designed cancer vaccine to a whole-animal movie of cellular activity, and it recurs throughout the year's medical breakthroughs we scored.

The lasting lesson of this result is quiet and worth stating plainly. The frontier of knowledge is not only pushed outward by new instruments; it is also pushed inward through old data by better methods of seeing. A generation of seismometers recorded far more than any human could ever read, and the signal that reveals the base of the mantle was sitting in those archives the whole time, ten times more of it than we had ever catalogued, waiting for a method patient enough to find it. That the method turned out to be a neural network with a human editor is fitting. The best deep-Earth science, like the best of Frontier's own scoring, is a careful conversation between what the data can support and what a person is willing to claim on its behalf.

This guide reflects the evidence available as of September 2026 and Frontier's assessment at that date. Seismic imaging of the deep Earth is model-dependent and provisional by nature, and specific findings, including the six bands discussed here, may be revised as independent studies test them. Frontier Scores are recomputed as evidence accumulates.