About 1,800 miles beneath your feet, Earth’s mantle meets the liquid outer core. That contact is not a smooth seam. It is a landscape. For decades seismologists have known that the base of the mantle is lumpy, with patches of odd material that scatter passing earthquake waves. Mapping those patches has always been a fight against faint signals and uneven coverage. A new effort from Chinese Academy of Sciences researchers uses a deep-learning system, a mountain of old seismograms, and a lot of human cleanup to argue that the lumps are more connected, and more numerous, than earlier catalogs showed.
The study is in the Journal of Geophysical Research: Solid Earth. Its claim is not that the planet has suddenly grown new guts. It is that the guts were always there, and that a machine trained to hear a whisper in earthquake records can finally draw them as belts instead of scattered islands. Six of those features, in regions the old networks barely sampled, now look like major hidden heterogeneities. If the interpretation holds, they are not trivia. Deep structures can steer mantle flow, magma, volcanoes, and even the magnetic field.
The floor of the rocky world
The mantle is solid on human timescales and still flows over geologic ones. The outer core beneath it is liquid metal. The boundary between them is one of the most extreme interfaces in the planet: rock against metal, solid against liquid, a jump in density and chemistry, a temperature that belongs to a furnace. About 1,800 miles down, that interface is where descending slabs can stall, where heat from the core can pool, and where ancient material might linger for a very long time.
How faint waves become a map
Seismology is how that world is seen. Earthquakes shake the planet. Waves race through rock and metal. Stations at the surface record wiggles. The travel times and amplitudes of those wiggles encode the interior. Some waves dive through the core. Some skim the boundary. Some, the ones this paper cares about, scatter off lumps near the core-mantle boundary and arrive just before stronger waves. Those early arrivals are PKP precursor signals. They are faint. They are easy to miss. They are also, if you can collect enough of them, a map.
The problem has always been catalog size. A handful of precursors in a handful of regions can look like isolated patches because that is all the lighting in the room. Turn on more lights and the patches may join. That is the bet behind training a model on more than two million earthquake recordings from 1990 to 2024.
A model, then a human with a red pen
The Chinese Academy of Sciences team trained a deep-learning system on that archive, then hand-corrected the model’s mistakes. That second step is not a footnote. Machine-learning papers in seismology often sell the network as an autonomous listener. This one treats the network as a first pass. The model proposes. People still decide.
Why the catalog had to be checked by hand
That workflow matches the nature of the signal. PKP precursors are not loud, clean spikes. They are small energy that shows up slightly early, scattered from heterogeneities near the core-mantle boundary, riding ahead of the main PKP arrival. A classifier that has seen the full archive can flag candidates a tired analyst would skip. A classifier can also flag noise, processing artifacts, and lookalikes. Hand correction is how a catalog becomes something other than a list of the model’s moods.
The harvest is a number that changes the scale of the problem. The system pulled out 174,929 faint PKP precursor signals — more than ten times all earlier catalogs. An order of magnitude is not a modest improvement in completeness. It is the difference between a sketch and a survey. With that many scattered arrivals, the map is no longer forced to treat every blob as a lonely island simply because the next station was too far away.
From isolated patches to belts
The new map shows patches once thought isolated actually join into larger belts. That sentence is the geological payoff. Heterogeneity at the base of the mantle has been described for years in terms of discrete piles, ultra-low velocity zones, and scattered scatterers. Some of those descriptions were always going to be resolution artifacts. If your catalog is small, everything looks isolated.
Belts imply organization. Material at the core-mantle boundary is not sprinkled at random. It is arranged, at least in places, into elongated structures. Elongation asks for a process: flow that stretches piles, slabs that drape along a front, or chemical domains that have been sheared by mantle circulation. The paper does not need to pick a single artist to make the observational point. The point is connectivity. What used to look like dots can look like lines once the catalog is an order of magnitude larger.
Why belts matter more than dots
Connectivity also changes how the features might interact with the rest of the deep Earth. A isolated patch is a local oddity. A belt is a boundary condition for flow. Fluid-adjacent, high-contrast structures that run for long distances can deflect currents in the mantle the way a submerged ridge deflects ocean water. They can also focus heat.
Six zones that used to be in shadow
The most striking additions are not just bigger versions of known piles. Six poorly sampled zones, labeled B1 to B6, including high-latitude Eurasia, Central Asia, and the South Atlantic, now look like major hidden heterogeneities. Poor sampling is the villain of global seismology. Earthquakes are not evenly spaced. Stations are not evenly spaced. Oceans are quiet in the wrong way. Polar paths are awkward. Entire swaths of the core-mantle boundary have been lit only by a trickle of useful rays.
B1 to B6 are the team’s names for six of those dim rooms after the lights came on. High-latitude Eurasia, Central Asia, and the South Atlantic are named explicitly as members of that set. They are not the only interesting places in the deep Earth, but they are the ones this catalog newly elevates from “we did not have the rays” to “there is something large down there.” Calling them major hidden heterogeneities is a claim about amplitude and extent, not a claim that anyone has touched them.
Names for rooms that used to be dark
The labeling — B1 to B6 — is itself a sign that the field is still in a naming-and-mapping stage. These are not yet household structures like a named mountain range. They are catalog objects. If later work confirms them with other wave types, other frequencies, and other inversion styles, the labels will either stick or be absorbed into a broader geography of the core-mantle boundary. For now they are the paper’s way of saying: look here, here, and here, where earlier maps were quiet.
What the piles might be
The team thinks they may be thermochemical piles. That is a compact phrase for a messy menu of origins. The paper’s own list is the right one to keep in view: leftover subducted slabs, partial melt, mineral transitions, or leftovers from the object that made the Moon, interacting with giant low-shear-velocity blobs.
Leftover subducted slabs are the most familiar. Oceanic plates dive into the mantle at trenches. Some of that cold, chemically distinct rock may reach the base of the mantle and stall, stacking into piles that still differ in temperature and composition from the surrounding rock. Those piles would scatter waves. They would also be denser or lighter than their surroundings depending on how much they have warmed and what minerals they now hold.
Partial melt is a different kind of heterogeneity. A small amount of liquid, perched at the hottest, lowest part of the mantle, can drop seismic velocity sharply and scatter energy. Mineral transitions — changes in crystal structure under core-mantle pressures — can do similar work without any liquid at all, if the new mineral pack is slower or faster than the old one. Each of those options is a thermochemical story. Temperature and chemistry are entangled. A pile that is both hotter and chemically distinct is not required to choose a single identity.
Slabs, melt, minerals, or a leftover world
The most ancient option on the list is leftovers from the object that made the Moon. The standard giant-impact story says a large body struck the early Earth and lofted the material that became the Moon. Some models allow residue from that impactor, or from the magma ocean the impact created, to sink and survive at the base of the mantle. If that residue is still there, it would be among the oldest chemical domains in the planet, sitting at the core-mantle boundary like a stain that never washed out.
Those piles, whatever their recipe, are not imagined in isolation. They interact with giant low-shear-velocity blobs. Those blobs — large, slow regions in the deep mantle already known from tomography — are a central fact of deep-Earth geophysics. Piles and blobs together make a basal landscape with highlands and lowlands of seismic speed. The new belts and the B1-to-B6 zones are additional topography on that landscape.
Why a lump 1,800 miles down can matter at the surface
It is fair to ask why anyone who does not invert seismograms should care. The team’s answer is causal, not aesthetic. Those deep structures can steer mantle flow, magma, volcanoes, and even the magnetic field.
Mantle flow is the slow convection that moves plates, lifts plumes, and mixes the interior. A thermochemical pile at the base is a barrier and a heat engine. Flow has to go around it or over it. Heat from the core has to leak through it or around it. Over geologic time that steering can decide where upwellings form. Upwellings that reach the upper mantle feed magma. Magma feeds volcanoes. A map of hidden basal belts is, in that sense, a possible map of why volcanism is not randomly sprinkled across the planet.
Heat, plumes, and the dynamo
The magnetic field is a core story, but the mantle still has a vote. The liquid outer core dynamo is sensitive to how heat leaves the core. If the core-mantle boundary is a patchwork of insulating piles and leaky windows, the pattern of heat flow can organize core convection. Organized core convection can organize the field, including the way it wanders and reverses. The paper is not a claim that B1 to B6 flip the poles. It is a claim that major hidden heterogeneities belong on the list of things that might shape the thermal boundary condition of the dynamo.
All of that is downstream of a measurement. The measurement is 174,929 PKP precursors, harvested from more than two million earthquake recordings spanning 1990 to 2024, after a deep-learning system proposed detections and people corrected its mistakes. The waves that matter are the ones that scatter off lumps near the core-mantle boundary and arrive just before stronger waves. The map those waves support shows isolated patches joining into larger belts, and six poorly sampled zones — including high-latitude Eurasia, Central Asia, and the South Atlantic — stepping forward as major hidden features.
A catalog is not the last word
Seismic maps are arguments. They depend on which phases you pick, how you define a detection, how you locate the scatterer, and how you draw contours through uneven coverage. A tenfold jump in catalog size is a powerful argument, not a photograph. Other groups will try to see B1 to B6 with different tools. Some of the belts may fragment again at higher frequency. Some of the thermochemical interpretations may lose to simpler thermal ones, or the reverse.
What will not reverse is the location of the mystery. About 1,800 miles down, mantle meets liquid outer core. The interface is lumpy. The lumps scatter. For a long time the catalog of those scattered whispers was too small to tell islands from archipelagos. A deep-learning pass through a generation of earthquake recordings, checked by hand, has made the catalog large enough to prefer belts, and to name six hidden structures in the planet’s poorly lit basement. Whether they are slabs, melt, mineral shifts, or Moon-making leftovers, they sit in a conversation with giant low-shear-velocity blobs, and they are in a position to steer the slow machinery that links the core to volcanoes and the magnetic field. That is what the Journal of Geophysical Research: Solid Earth paper puts on the table: not a new Earth, but a sharper one.