Feed aggregator

New research shows surface melting across Antarctica is set to intensify and spread dramatically over the 21st century, with melt increasing 10-fold and the affected area growing by more than 10% by 2100 if global temperatures continue to rise.

Albert Einstein postulated in his 1905 theory of special relativity that the speed of light in a vacuum is constant. Ever since, that’s been one of the fundamental assumptions of physics.

Now Enbang Li, a physicist at the University of Wollongong in Australia, has challenged this idea by building a machine he says is capable of detecting changes in the speed of light as it crosses Earth’s surface. The findings suggest that light is, in fact, sped up by gravity, which could have implications for Earth science applications ranging from climate monitoring to mineral resource exploration.

An Old Conundrum

The idea that light is influenced by gravity is not new. Einstein’s ideas, which were further developed with his theory of general relativity in 1915, predicted massive objects in space would bend light with their gravitational grab. This theory was famously proven in 1919 when two independent teams measured starlight passing a solar eclipse at two different points on Earth’s surface and found the results matched Einstein’s predictions.

This bending of light’s path, according to general relativity, is achieved by a warping of the space-time fabric. Under this scenario, the speed of light remains constant—it just has to travel farther as it navigates the warped space-time around celestial bodies, so to a distant observer, it appears to have been slowed.

But what if light doesn’t navigate warped space-time and actually is slowed down or sped up by the gravity of large objects?

Li pointed out that Einstein himself was not always convinced the speed of light was constant. In 1911, he wrote a paper postulating that light speed changed depending on the gravity of objects it passed by. However, “when he published his general theory,” said Li, “he just abandoned this model.”

If the movement of light can be affected by gravity, Li reasoned, it might be possible to detect variations in its speed on a local level—such as an elevator shaft in a building on the campus of the University of Wollongong.

Raising the Big Issues

Gravity on Earth varies locally, depending on altitude, underground density, and topography. Gravity at the top of a tall building, for example, is measurably weaker than it is at the bottom.

With these variations in mind, Li installed an experiment in an elevator. It consisted of a coil of fiber-optic cable that if stretched out in one direction, would be 10 kilometers (6.2 miles) long. Laser beams were fired through the cables and then reflected back, thus traveling 20 kilometers (12.4 miles) before reaching an ultrafast photodetector. An oscilloscope measured the time it took for the beam to travel that distance. The experiment was run at the top of the shaft and at the bottom.

The biggest challenge, Li said, was filtering out all the surrounding environmental “noise,” such as changing temperature and humidity, electromagnetic disturbance, and building vibrations. Li designed a temperature control system, and the experiment was sealed in an enclosure with electromagnetic shielding to isolate air flows. Li ran the experiment and found light moved minutely faster at the bottom of the shaft than at the top.

Gravity Sensing on the Go

Next, Li took his research a step further by building a small, portable machine he claims can detect changes in the speed of light as it nears more gravitationally dense objects.

In this second experiment, Li positioned a moveable 72-kilogram (159-pound) weight near the machine. Light, he found, moved faster when the weight was near the machine than when it was farther away.

The results, which were published in Scientific Reports, are consistent with the variable speed of light model Einstein proposed in 1911, although Li’s preliminary results are much larger than that model predicts.

If proven, the findings would present a fundamental challenge to our understanding of both general and special relativity.

In the world of Earth sciences, they could lead to greatly improved gravity-sensing technologies. Because of their sensitivity to changes in mass, gravity sensors are used to map the seafloor and to locate underground mineral reserves. Gravity sensing can also improve our understanding of Earth’s climate as variations in the gravity field can be linked to factors like changes in ice mass and shifts in groundwater.

Currently, gravimeters are vulnerable to vibrations and movement, whereas Li’s machine, which has no moving parts, could even be used on board a plane or submarine.

“A Striking Claim”

Chris Stevens, a numerical relativist with the University of Canterbury in New Zealand, called the work “intriguing and ambitious.” While Stevens, who was not involved in the research, said that Li’s work is “well founded,” he noted that any observable effects of gravity on light on Earth would be “extraordinarily small” and therefore these results must be treated with caution.

“In my own research on observable gravitational phenomena,” he explained, “I usually require a few black holes colliding somewhere in the universe. Separating genuine gravitational signatures from environmental and instrumental noise will therefore be exceptionally demanding.”

“The work is exciting because it pushes precision photonic measurement techniques into a regime where relativistic effects may become practically useful for geophysics and sensing applications.”

Stevens said the implications of Li’s research, if validated, would be far-reaching. “The work is exciting because it pushes precision photonic measurement techniques into a regime where relativistic effects may become practically useful for geophysics and sensing applications.”

John Norton, an historian of physics at the University of Pittsburgh who was also not involved in the research, called the findings a “striking claim.” He was, however, skeptical of them, saying “if there is a coupling between light and gravity of magnitude greater than general relativity predicts, it is hard to see how the 1919 eclipse test and later studies of gravitational lensing would not have found it.”

Li acknowledged there is a long way to go before his device finds everyday use. Disentangling the intricacies of space and time, he said, is a vast challenge. “In physics, people still say gravity is a mystery. Light is another mystery. So if you put these two mysteries together, that’s going to be a giant mystery.”

—Bill Morris, Science Writer

Citation: Morris, B. (2026), How Einstein’s lost theory could help us find minerals, Eos, 107, https://doi.org/10.1029/2026EO260189. Published on 12 June 2026. Text © 2026. The authors. CC BY-NC-ND 3.0
Except where otherwise noted, images are subject to copyright. Any reuse without express permission from the copyright owner is prohibited.

SummaryThis study leverages a new, improved and densified GNSS velocity field of the western Aegean region to quantify slip rates, strain localisation, and interseismic loading within the upper plate of the Hellenic subduction zone, including the deformation systems associated with the subduction–collision transition around Cephalonia, the Hellenic forearc extension, and the southwestern termination of the North Anatolian Fault system. We examine several active tectonic domains, comprising four extensional regions (Corinth–Patras Rift, Evia Gulf, Argolic Gulf, southwestern Peloponnese) and three strike-slip systems (Cephalonia Transform Fault, Katouna–Stamna Fault System, Movri Fault Zone). Across the Corinth Rift, NS extension increases westward from ≈ 7 to ≈ 15 mm.yr$^{\hspace{1.0pt}\text{--}1}$, of which up to ≈ 6 mm.yr$^{\hspace{1.0pt}\text{--}1}$ is accommodated offshore. Velocity profiles indicate combined elastic loading and aseismic creep on a limited number of crustal-scale faults. The Evia–Boeotia sector undergoes NS extension at up to 8 mm.yr$^{\hspace{1.0pt}\text{--}1}$, but deformation is distributed across multiple structures, each accommodating creep or elastic loading at rates <2 mm.yr$^{\hspace{1.0pt}\text{--}1}$. In the southern Peloponnese, diffuse EW extension of up to 6 mm.yr$^{\hspace{1.0pt}\text{--}1}$ occurs alongside 1–2 mm.yr$^{\hspace{1.0pt}\text{--}1}$ of NS extension in the central Peloponnese. A portion of this EW deformation may be accumulating interseismically as elastic strain on prominent structures, such as the Sparti and East Messenia faults. No measurable strain is detected across the Argolic Gulf, suggesting substantially lower present-day loading rates on the Astros Fault than previously inferred. Strike-slip systems display contrasting behaviours. The Katouna segment accommodates transtensional left-lateral creep of ≈ 13 mm.yr$^{\hspace{1.0pt}\text{--}1}$ within a zone <6 km wide, whereas slip decreases to ≈ 8 mm.yr$^{\hspace{1.0pt}\text{--}1}$ on the Stamna segment, consistent with strain transfer through the Trichonida pull-apart basin and the Nafpaktia diffuse shear zone. In contrast, the Movri Fault appears locked down to at least 10 km depth, accumulating ≈ 4 mm.yr$^{\hspace{1.0pt}\text{--}1}$ of right-lateral strain. Onshore velocities near the Cephalonia Transform Fault indicate an onshore half-rate of elastic loading of ≈ 8 mm.yr$^{\hspace{1.0pt}\text{--}1}$, suggesting that the full transpressional right-lateral motion (≈ 16 mm.yr$^{\hspace{1.0pt}\text{--}1}$) accumulates interseismically, highlighting considerable seismic hazard potential.

SummaryDistributed Acoustic Sensing (DAS) data often contain various types of noise, including random noise, coherent noise (e.g., coupling or linear noise), and common mode noise, which significantly degrade seismic signal quality. Conventional denoising methods struggle to effectively suppress diverse noise components while preserving important seismic signals. To address this issue, we propose a denoising self-supervised cascade network (DAS-DSCnet), a multi-stage neural network designed to progressively denoise DAS data without requiring external labels or synthetic training data generation. The network consists of three stages: Stage 1 targets random noise using a Noise2Noise-based approach; Stage 2 suppresses dataset-specific coherent noise using a denoising convolutional neural network (DnCNN)-based network trained with internally extracted noise patches; and Stage 3 predicts and removes common mode noise through trace shuffling and a Noise2Noise-based model. Training data for each stage are generated directly from the input DAS data by exploiting the data’s inherent characteristics, enabling efficient learning that reflects field-specific noise features. The model was evaluated using two distinct field DAS datasets with different noise patterns. The results demonstrate that DAS-DSCnet achieves superior noise suppression compared to conventional approaches, enhancing signal continuity while minimizing leakage. The denoising performance remains stable across different stacking configurations and hyperparameters, confirming the model’s robustness. Therefore, DAS-DSCnet offers a scalable and practical framework for improving seismic data quality in DAS applications, demonstrating the potential for fully automated, data-driven denoising in large-scale seismic monitoring.

SUMMARYWe present a comprehensive dataset of 920 coordinates and 509 velocities for geodetic points in central Greece and the Peloponnese, an area characterised by intense tectonic deformation. The points, with observation periods within the 1990–2024 range, are organised into three categories: permanent stations, triangulation pillars, and markers. The latter two categories are subdivided according to whether or not they feature self-centring. Most of the triangulation pillars belong to the Greek national network originally surveyed in the 1960s–70s. The GNSS data were processed using the GIPSY 6.4 software. To assess the secular velocities, we corrected for co-seismic and post-seismic displacements using earthquake parameters constrained by the time series of the permanent stations. Self-centring systems improve precision, reducing the average horizontal coordinate residual variability from 6.15 to 4.45 mm. The velocity uncertainties stabilise below 0.15 mm yr−1 when the time series exceed twenty years. Points with self-centring achieve 0.2 mm yr−1 accuracy after twelve years of data, compared to twenty years for those without self-centring. After twenty-five years, campaign points observed eight to ten times match the precision of permanent stations. The velocities at the campaign points further validate the HELVEL model previously developed using permanent stations only. We calculate a seven-parameter transformation from the original coordinates of 424 triangulation pillars to their GNSS-based ITRF2020 coordinates at epoch 2020.0. The lowest mean scatter after the transformation is 0.134 m when 1965 is used as the mean epoch for the triangulation data. We then apply this transformation to all 9,729 pillars of the study area. At the 424 resurveyed pillars, the GNSS ellipsoidal heights agree with the sum of the levelled heights and the official HG2023 geoid heights to within 0.184 m root-mean-square. Our dataset is entirely referenced to ITRF2020 at epoch 2020.0, which enables interoperability with previous and future geodetic studies. Dense campaign point arrays are critical for resolving the strain distribution at the scale of individual active faults, beyond the reach of arrays of permanent stations alone.

AbstractMantle viscosity remains one of the largest outstanding uncertainties in global geodynamics. Time-dependent mantle circulation models that assimilate tectonic histories (MCMs) provide a way to test viscosity by assessing their present-day predictions against observations. This approach allows for the influence of viscosity on mantle density structure to be accounted for, which is not possible using instantaneous modelling approaches. Here we present the first systematic test of lower mantle viscosity against dynamic topography, the geoid, and seismic heterogeneity using high-resolution MCMs. Model density structure depends strongly on the assumed viscosity profile, which in turn controls the fit to seismic heterogeneity. The fit to dynamic topography and the geoid is further influenced by the instantaneous transmission of stresses to the surface. These two effects can either reinforce or counteract each other at different depths, which must be considered when attempting to match dynamic topography and geoid amplitudes. MCMs typically overestimate dynamic topography amplitudes. We find that it is possible to reduce these amplitudes by lowering viscosity in the upper lower mantle (≈660-2000 km), though this comes at the expense of a reduced fit to the geoid and/or seismic heterogeneity. Our preferred viscosity profile provides an excellent fit to observed geoid amplitudes and the seismic heterogeneity of S40RTS. We also tested an alternate tectonic reconstruction with tomography-based refinements around the Pacific which improved the correlation with the observed geoid by ≈20%. Our results show that MCMs can now reach a level of resolution and realism sufficient for comparison to multiple independent data sets, opening the door to systematic assessment of uncertain parameters which govern convection in the mantle.

SUMMARYThe excitation and propagation of multiple wave types, including seismic waves, ocean acoustic waves, and tsunamis triggered by earthquakes within the oceanic wavefield, constitute a problem of substantial scientific and practical challenge. This phenomenon involves complex interactions of waves within a fluid-solid coupled system, which is critical for both fundamental geophysical understanding and enhancing hazard assessment. While several numerical methods have been developed to simulate the full wavefield, few studies have systematically explored the crucial influence of complex seafloor topography on coupled wave dynamics. This study introduces a novel earthquake-tsunami coupling simulation method within a 2-D spectral-element method (SEM) framework, leveraging its flexibility to handle complex geometries and accuracy for long-range wave propagation. To validate the accuracy of the simulated seismic waves, ocean acoustic waves, and tsunamis, we quantitatively evaluate the permanent seafloor displacement, yielding a high correlation coefficient of 0.997 and a negligible error of 5 × 10−3 compared with the analytical solution. The mean relative error of the calculated tsunami phase velocities of the proposed method with those from the propagator matrix method is only 0.12%. Furthermore, we establish two distinct numerical models—one incorporating irregular bathymetry and another with an idealized flat bathymetry—to systematically investigate the effects of bathymetry on the oceanic wavefield. Our results demonstrate that the irregular bathymetry significantly influences the propagation characteristics of both seismic waves and tsunamis, altering wave amplitudes, travel times, and spatial patterns. We further decompose the contributions of seawater and seafloor geometry, highlighting their respective roles in shaping the overall wavefield. Additionally, we examine the influence of varying earthquake source locations on wave propagation paths, emphasizing the importance of accurately modelling bathymetry for offshore seismic events. Overall, our proposed 2-D earthquake-tsunami coupling simulation framework provides a powerful tool for comprehensively understanding the oceanic wavefield under gravity and offers significant potential for improved earthquake and tsunami hazard assessment, particularly when combined with seismological and oceanographic observations.

body {background-color: #D2D1D5;} Research & Developments is a blog for brief updates that provide context for the flurry of news that impacts science and scientists today.

The key word here is could. Experts including Ken Graham, the director of NOAA’s National Weather Service, all emphasize that no two El Niños are alike.

“Each one is unique with its own imprint on our weather,” Graham said in a NOAA press release. However, scientists have learned a few things from watching the ways that this warm phase of a natural climate cycle over the tropical Pacific has affected our weather patterns in the past.

“Advanced monitoring and an improved understanding of El Niño patterns allow the NWS to better predict and better prepare the public and our core partners for what is to come,” Graham said.

 Related

•  El Nino forms, expected to strengthen, say NOAA forecasters
•  Our Warming Waters Press Briefing
•  NOAA Forecasts a Below Average Hurricane Season
 

This morning, NOAA released an El Niño Advisory, announcing that the climate phenomenon (the warm phase of the El Niño–Southern Oscillation) has officially arrived in the tropical Pacific. The agency forecasts a 63% chance of a “very strong” El Niño from November 2026 to January 2027 that “would rank among the largest El Niño events in the historical record.”

NOAA defines a “very strong” El Niño as when the Pacific’s surface waters are more than 2°C warmer than average. The agency doesn’t use the phrase “Super El Niño,” but there have only been three such “super” or “very strong” El Niño events since 1980. The last one was in 2015.

What does this mean for climate, for humans, and marine species? Here’s a roundup of some potential forecasted effects—some good, some bad—of the weather pattern that’s been making headlines over the past few months.

1. More rain and snow in the southern U.S.

In a typical year, a warm pool of water in the equatorial Pacific would be transported westward—away from the western coast of the Americas—by trade winds. But during an El Niño event, those trade winds weaken, and the warm pool of water extends east, explained Ariel Cohen, the meteorologist in charge of the National Weather Service’s Los Angeles and Oxnard Office in a press briefing at the Aquarium of the Pacific in Long Beach, Calif.

This warm water “causes jet energy in the atmosphere to bring disturbed weather southward across the southern United States, which can bring wetter than normal conditions to our area with drier conditions farther to the north,” Cohen said.

The southward shift of the storm track could also lead to drier conditions over the northern Rockies and as far east as the Ohio and Tennessee Valleys.

2. More shark and whale sightings off the Southern California coast

In the past, strong El Niños have led to decreased amounts of plankton in the Pacific, particularly the open ocean, forcing species that rely on plankton (and the species that rely on the species that rely on plankton, and so forth) to widen their net when searching for food.

“[Plankton] is important because that’s the base of the food web,” explained Andrew Leising, a research oceanographer at NOAA, at the Aquarium of the Pacific. “Marine mammals and other migratory species end up being closer to shore, because they’re going to where their food is.”

Whales in particular rely on the upwelling of cold water to bring them krill to eat. As they are driven nearer to the coast in search of food, they also grow more likely to become entangled in fishing nets.

3. A milder Atlantic hurricane season

Warm water is a key ingredient in a hurricane, so it might seem, at first thought, that the Pacific’s unusually warm waters might augur a more extreme hurricane season. But another effect of El Niño is that it strengthens vertical wind shear over the Atlantic. When winds are too strong, they can tear a storm apart before it picks up the momentum to become a hurricane.

“Wind shear is good for us, bad for the hurricanes,” Phil Klotzbach, a hurricane forecaster at Colorado State University and lead author of the university’s 2026 Atlantic Hurricane Forecast, told Eos.

NOAA’s 2026 Atlantic Hurricane Forecast suggests that the 2026 season has a 55% chance of being below normal, and will likely include 8 to 14 named storms with winds of at least 39 miles per hour.

4. Fewer squid along the California coast

Past El Niño events have shown that warmer Pacific waters can increase the likelihood of harmful algal blooms. Among other effects, these blooms can lead to a lower abundance, and a northward shift, of market squid. Market squid and Dungeness crab bring the most volume and value to California’s commercial fisheries.

In 2014, a large mass of hot water in the Pacific known as the Blob was followed up by an El Niño event. That year, “we had several closures of crab and shellfish fisheries due to harmful algal blooms,” Leising said.

However, Leising also explained that the warm patch of water in the Pacific this year is much smaller and farther from shore than the Blob was in 2014. So, though we may see effect similar those in 2014, they’re likely to be less extreme.

In addition, the same conditions driving sharks and whales toward the coast could also drive tuna toward the coast, leading to increased opportunities for that fishery.

5. More high-tide flooding on U.S. coasts

With El Niño shifting the Pacific jet stream south of its usual position, sea levels along the U.S. West Coast may rise, exacerbating the existing sea level rise linked to climate change. On the East Coast, the jet stream shift can lead to more storm surges, which combine with higher-than-typical precipitation levels.

“It usually ends up being a double whammy,” said NOAA oceanographer and high tide flooding expert William Sweet, in a NOAA news story. “The first punch is decades of sea level rise, which has waters close to the brim in many coastal communities. And now with this second punch—a strong El Niño—coastal communities face more frequent, deeper and widespread high tide flooding along both the West and East Coasts.”

6. A bad year for sea lions

El Niño events can have harmful effects on sea lions. Algal blooms can lead to severe illness, or even death, for the pinnipeds. Algal blooms can also kill off fish and cephalopod species (such as market squid) that sea lions rely on for food. During past El Niño events, California sea lions have also experienced lower rates of reproduction and produced smaller pups, Leising said.

“California sea lions are indicator species, meaning they will be one of the first species which may show signs of domoic acid toxicity, respond to changes in their ecosystem, and signal to the public how our oceans and ecosystem are doing,” said Brett Long, vice president of animal care at the Aquarium of the Pacific.

—Emily Gardner (@emfurd.bsky.social), Associate Editor

These updates are made possible through information from the scientific community. Do you have a story about science or scientists? Send us a tip at eos@agu.org. Text © 2026. AGU. CC BY-NC-ND 3.0
Except where otherwise noted, images are subject to copyright. Any reuse without express permission from the copyright owner is prohibited.

Planetary heating is intensifying and key climate indicators are deteriorating, top scientists said Thursday, warning that funding decisions affecting Earth observation systems in the United States and other countries threaten efforts to track global warming.

In 2023, Cyclone Gabrielle triggered an estimated 800,000 landslides across the North Island, making it one of the most extreme landslide events ever recorded. New research by Te Whare Wānanga o Waitaha | University of Canterbury (UC) and Earth Sciences New Zealand suggests that under a warmer climate, future storms similar to Cyclone Gabrielle could be even more extreme, triggering tens of thousands more landslides across parts of the North Island and highlighting the need for targeted planning in vulnerable areas.

A new study finds that burning 500,000 acres (202,000 hectares) of California conifer forests each year with prescribed fire could cut deadly pollution from wildfire smoke by roughly 10% over a decade.

In a new paper published in Science, leading scientists and climate policy experts show that 15% of current global warming (0.3°C) from human emissions stems from pollutants that fall outside most existing climate policy frameworks. Most of these overlooked pollutants are called "indirect greenhouse gases" and include carbon monoxide, non-methane volatile organic compounds, nitrogen oxides and molecular hydrogen.

Editors’ Highlights are summaries of recent papers by AGU’s journal editors. Source: Journal of Geophysical Research: Solid Earth

Earthquakes release energy and result in source properties defined across a wide range of scales that are not represented in conventional frictional laws. Norisugi and Noda [2026] introduce a new rate- and roughness-dependent friction (RRF) law which incorporates both effects from fault slip rate and multi-scale variation in fault topography. By limiting the number of state variables in the RRF formulation, the authors show with efficient earthquake cycle simulation that this multi-scale approach can reproduce a key observed relationship between fracture energy and fault slip.

Although further refinement is needed to better represent roughness evolution, this study marks a major advance in earthquake modeling by demonstrating the necessity and feasibility of incorporating multi-scale fault topography in the characterization of earthquake source process.  

Citation: Norisugi, R., & Noda, H. (2026). Multi-scale rate- and roughness-dependent frictional constitutive law and dynamic earthquake sequence simulation. Journal of Geophysical Research: Solid Earth, 131, e2025JB033580. https://doi.org/10.1029/2025JB033580

—Yajing Liu, Associate Editor, JGR: Solid Earth

Text © 2026. The authors. CC BY-NC-ND 3.0
Except where otherwise noted, images are subject to copyright. Any reuse without express permission from the copyright owner is prohibited.

Researchers combining two methods to reconstruct the rupture evolution of the July 2025 magnitude 8.8 Kamchatka earthquake found the rupture from the megathrust event extended about 500 kilometers (311 miles) from its epicenter.

Solar activity affecting Earth and its planetary neighbors encompasses a wide range of phenomena, from the steady solar wind and the interplanetary magnetic field to extreme events like solar flares, coronal mass ejections (CMEs), and solar energetic particle (SEP) events. These space weather phenomena interact in complex ways with planetary magnetospheres and atmospheres. On Earth, we see the results in the dancing lights of stunning auroras and in less frequent but sometimes severe disruptions to telecommunications, navigation, and energy infrastructure.

Forecasting conditions throughout the heliosphere (the region influenced by the solar wind), understanding the variety of Sun-Earth interactions, and predicting arrivals of space weather events—both benign and potentially hazardous—are a grand challenge.

The Sun-Earth challenge requires tracking and predicting conditions—from routine and quiet to rare and extreme—across tens of millions of kilometers of interplanetary space.

Solar flares emit electromagnetic radiation that spreads in all directions. In contrast, the propagation of CMEs and SEP events depends on their source location on the Sun and on the heliospheric magnetic field, which is carried outward by the solar wind. The impacts these events have on magnetosphere systems further vary depending on particle energies and intensities in SEPs and on particle speeds and the magnetic field orientation in CMEs. The Sun-Earth challenge thus requires tracking and predicting conditions—from routine and quiet to rare and extreme—across tens of millions of kilometers of interplanetary space.

This tracking and prediction is powered by petabyte-scale datasets from solar observatories and spacecraft measurements that provide rich observational archives. Researchers use these data to deduce physically meaningful quantities describing the heliosphere and to identify patterns to distinguish quiet from active conditions. The resulting insights not only answer fundamental science questions but also provide critical prediction time frames needed by space weather forecasters.

Even with all these data, the enormity of space between the Sun and Earth presents a major obstacle to our predictive capabilities. Another obstacle is that the data are obtained by different instruments operating at different locations and times. These factors combine to create a unique data sparsity challenge that complicates large-scale analysis.

These fundamental issues—the massive yet still insufficient supply of data available, the extreme differences in the scales of the processes we must illuminate, and the need for actionable predictions—suggest opportunities for artificial intelligence (AI) and machine learning (ML) to complement traditional physics-based analytical approaches [Camporeale, 2019]. In a series of workshops—insights from which inform the discussion below—scientists explored such opportunities and how they can advance heliophysics research and operational space weather forecasting.

The Need for Space Weather Forecasting

Space weather events can have significant impacts on infrastructure and humans. They can disrupt satellite operations (e.g., by enhancing atmospheric drag on satellites), damage electronics in space, interfere with radio communications and GPS, and even affect power grids (e.g., through geomagnetically induced currents) during the most severe events. They can also pose risks to people, especially astronauts beyond the protection of Earth’s atmosphere and airline crews and passengers on long-distance polar flights, during which exposure to energetic particles is elevated. Forecasting offers a first line of defense in preparing for or preventing damaging and hazardous effects of space weather.

In assessing major CMEs, forecasters consider whether and when events will reach Earth and whether they will trigger geomagnetic storms and substorms. For SEP events, predictions must include arrival times, peak intensities, durations, and energy characteristics.

Predicting extreme space weather phenomena is vital, but equally important is forecasting periods when no significant activity is expected, which is critical information for satellite operators and other stakeholders. Making such predictions requires understanding physics spanning 8 orders of magnitude in space and time, from subsecond processes in Earth’s magnetic environment to multiday solar eruptions propagating across the 150 million kilometers between the Sun and Earth (Figure 1) and long-term interactions at scales associated with the 11-year solar cycle.

Fig 1. Length scales and Sun-to-Earth transit times vary greatly for different types of space weather (SW), including solar flares, solar energetic particle (SEP) events, coronal mass ejections (CMEs), and interplanetary coronal mass ejections (ICMEs). High-speed particles are the first to arrive, usually within minutes of a flare, whereas CMEs arrive in 2–4 days. Credit: Georgoulis et al. [2026], CC BY-NC-ND 4.0

In addition to operational forecasting, these challenges are fundamental in heliophysics research. Such research includes work to reveal how the Sun generates its magnetic field, how solar wind accelerates and evolves, how planetary magnetospheres respond to external forcing, how particles are accelerated, and how energy transfers across multiple scales and regimes.

Unique Challenges in Heliophysics

Modern AI and ML algorithms excel at analyzing well-curated, extensive datasets that include millions of training examples. For example, AI-aided terrestrial weather forecasting relying on continuous, high-resolution coverage from thousands of ground stations, weather balloons, and satellites has advanced dramatically in recent years.

Fewer than a dozen spacecraft monitor Earth’s magnetosphere, a region spanning tens of Earth radii. Solar wind observations are even sparser.

Heliophysics, however, presents a unique and somewhat opposite scenario. Fewer than a dozen spacecraft monitor Earth’s magnetosphere, a region spanning tens of Earth radii (about 6,371 kilometers). Solar wind observations are even sparser, with just a handful of monitors scattered across the space between the Sun and Earth. This fundamental scarcity poses a challenge for data-driven approaches, which typically depend on abundant observations that are well distributed in space and time to produce trustworthy (i.e., generalizable and reproducible) models.

Data sparsity is further compounded by the relative rarity of intense space weather phenomena such as CMEs, major geomagnetic storms, and extreme substorms, which occur only a few times per solar cycle. Most heliophysical observations capture quiet, low-activity conditions when the solar wind is steady and magnetospheres are calm. Standard ML approaches trained on such imbalanced datasets may achieve high statistical accuracy by simply predicting a “nothing-will-happen” outcome but completely fail when extreme events occur.

Although solar eruptions and geomagnetic storms are relatively rare, they exhibit recurring patterns and consistency in their physical drivers. This regularity suggests that historical observations, when properly clustered and analyzed, can be used to enhance prediction capabilities. The challenge therefore lies in extracting meaningful patterns from sparse measurements of rare events while avoiding models that work well for average conditions but fail when they matter most [Chu et al., 2025].

AI Solutions for Data Sparsity

Heliophysics research employs clever approaches to extract maximum information from the limited available observations. One strategy is to mine multidecade observational records from various satellites and to match and group together measurements collected at times with similar solar wind and geomagnetic activity conditions.

This process clusters tens of thousands of data points from similar magnetospheric states. Such clustering enables reconstruction of dynamic features like nightside magnetic field changes during substorms [Stephens et al., 2019] and the presence of near-Earth magnetotail reconnections [Angelopoulos et al., 2020].

Another, more universal approach is to embed fundamental physical laws directly into ML models through physics-informed neural networks [Raissi et al., 2019], ensuring that predictions respect physical reality even when training data are limited. Data assimilation techniques used in weather forecasting similarly blend sparse observations with physics-based simulations and update models as new measurements arrive.

This animated model shows Earth’s magnetosphere during a powerful May 2024 geomagnetic storm that involved strong solar flares and multiple CMEs. The visualization uses the Multiscale Atmosphere-Geospace Environment (MAGE) model from the Johns Hopkins Applied Physics Laboratory to depict wind rushing toward Earth and disturbing its magnetic field (orange and purple lines). The green cloud represents electric field current intensity; the blue squiggles are tracers of solar wind velocities. Credit: NASA Scientific Visualization Studio and NASA DRIVE Science Center for Geospace Storms

These methods converge on a common theme: building gray box models (so named because they’re less opaque than black box models) that are data driven but grounded in physically real constraints. For data-starved applications, hybrid approaches can outperform purely data-driven or purely physics-based methods [Liu et al., 2025].

Satellite instruments are generating increasingly large solar wind datasets. However, the variables obtained (e.g., solar wind speed and pressure) are highly intercorrelated [Borovsky, 2018], making it difficult to identify which ones truly drive magnetospheric responses. New algorithms are helping to distill datasets without losing critical scientific information [e.g., Camporeale, 2025]. Meanwhile, advanced statistical and ML methods can cut through dataset complexity by reducing dimensionality, identifying causal relationships among variables, and providing clues about dominant drivers.

For instance, information theory provides tools to detect dependencies in complex systems, establish causality, and rank variables that most effectively predict space weather outcomes [Wing et al., 2022]. Such techniques can be paired with other “explainable” tools, such as SHAP (SHapley Additive exPlanations) values, a method inspired by game theory, to pinpoint physical variables (e.g., solar wind speed or magnetic orientation) that drive a prediction [Ma et al., 2023].

Distilling datasets and improving model interpretability help make ML more practical and more scientifically trustworthy and its predictions more robust. But fully trusting ML models in operational environments requires rigorous validation and uncertainty quantification. These models must not only make predictions but also indicate their confidence levels for operational decisionmaking.

When a model forecasts a major geomagnetic storm, operators need to know whether that prediction carries 60% or 95% confidence, for example.

When a model forecasts a major geomagnetic storm, operators need to know whether that prediction carries 60% or 95% confidence, for example. Ensemble approaches, in which multiple models provide a range of outcomes, help quantify this uncertainty, while using standardized, well-documented datasets enables fair model intercomparisons.

The research community is developing ML-ready benchmark datasets with consistent formatting and clear metadata to establish such validation procedures [e.g., Angryk et al., 2020]. These resources allow researchers to test new algorithms against common baselines, accelerating progress while ensuring that advances are robust and reproducible rather than artifacts of specific data processing choices.

Notably, one domain in heliophysics that is not affected by severe data sparsity is solar imaging. Decades of continuous, high-resolution observations from the Solar Dynamics Observatory (SDO), which delivers 1.5 terabytes of data every day, have created enormous data archives. Because the Sun drives space weather throughout the heliosphere, these datasets offer an ideal opportunity for use in foundation models, large-scale ML systems trained to learn comprehensive internal representations that can then be easily adapted to specific scientific tasks with minimal additional training.

Surya, a foundation model designed to construct a digital representation of the Sun, represents one such effort. It is still in early development and has yet to be validated, but this approach illustrates how data-rich domains can be leveraged with modern AI techniques to create tools that broadly benefit heliophysics research and space weather forecasting.

Advancing Research and Operational Forecasting Together

In addition to the needs for data and model development and validation, applying AI to address the challenges of heliophysics requires sustained, multidisciplinary collaborations. Fostering those collaborations has been the focus of a series of workshops, with the most recent being 2025’s Machine Learning, Data Mining and Data Assimilation in Geospace (LMAG25) meeting at the Johns Hopkins University Applied Physics Laboratory. The workshops have brought together heliophysicists, machine learning experts, data scientists, and specialists from weather forecasting and applied mathematics to exchange knowledge and establish community standards.

Space weather forecasters need models that are accurate and interpretable and that provide not just statistical metrics but also actionable predictions.

The LMAG forums also serve as gathering spaces for scientists to validate models against diverse datasets, compare physics-based and data-driven approaches, develop performance benchmarks, and discuss how to bridge research and operational requirements. Space weather forecasters need models that are accurate and interpretable and that provide not just statistical metrics but also actionable predictions with known limitations and reliability. Of course, researchers also benefit. These conversations allow them to gain insight into operational constraints that shape how modeling approaches become practical in real-world settings.

LMAG and similar initiatives facilitate direct exchanges among adjacent communities, including by making meeting presentations openly available. These efforts are helping translate cutting-edge AI and ML techniques into practical tools that help protect critical infrastructure and human well-being. They are also deepening our understanding of how the Sun shapes space weather throughout the solar system and its effects—both mundane and major—on Earth.

References

Angelopoulos, V., et al. (2020), Near-Earth magnetotail reconnection powers space storms, Nat. Phys., 16(3), 317–321, https://doi.org/10.1038/s41567-019-0749-4.

Angryk, R. A., et al. (2020), Multivariate time series dataset for space weather data analytics, Sci. Data, 7(1), 227, https://doi.org/10.1038/s41597-020-0548-x.

Borovsky, J. E. (2018), The spatial structure of the oncoming solar wind at Earth and the shortcomings of a solar-wind monitor at L1, J. Atmos. Sol. Terr. Phys., 177, 2–11, https://doi.org/10.1016/j.jastp.2017.03.014.

Camporeale, E. (2019), The challenge of machine learning in space weather: Nowcasting and forecasting, Space Weather, 17(8), 1,166–1,207, https://doi.org/10.1029/2018SW002061.

Camporeale, E. (2025), PARIS: Pruning Algorithm via the Representer theorem for Imbalanced Scenarios, arXiv:2512.06950, https://doi.org/10.48550/arXiv.2512.06950.

Chu, X., et al. (2025), Imbalanced Regression Artificial Neural Network Model for Auroral Electrojet Indices (IRANNA): Can we predict strong events?, Space Weather, 23(5), e2024SW004236, https://doi.org/10.1029/2024SW004236.

Georgoulis, M. K., et al. (2026), Prediction of solar energetic events impacting space weather conditions, Adv. Space Res., in press, https://doi.org/10.1016/j.asr.2024.02.030.

Liu, Y., et al. (2025), Data-driven modeling of electrostatic turbulence by physics-informed Fourier neural operator, Mach. Learn. Sci. Technol., 6(4), 045050, https://doi.org/10.1088/2632-2153/ae19cd.

Ma, D., et al. (2023), Opening the black box of the radiation belt machine learning model, Space Weather, 21(4), e2022SW003339, https://doi.org/10.1029/2022SW003339.

Raissi, M., P. Perdikaris, and G. E. Karniadakis (2019), Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations, J. Comput. Phys., 378, 686–707, https://doi.org/10.1016/j.jcp.2018.10.045.

Stephens, G. K., et al. (2019), Global empirical picture of magnetospheric substorms inferred from multimission magnetometer data, J. Geophys. Res. Space Phys., 124(2), 1,085–1,110, https://doi.org/10.1029/2018JA025843.

Wing, S., et al. (2022), Modeling radiation belt electrons with information theory informed neural networks, Space Weather, 20(8), e2022SW003090, https://doi.org/10.1029/2022SW003090.

Author Information

Savvas Raptis (savvas.raptis@jhuapl.edu), Manolis K. Georgoulis, Mikhail Sitnov, Anthony Sciola, and Simon Wing, Johns Hopkins University Applied Physics Laboratory, Laurel, Md.

Citation: Raptis, S., M. K. Georgoulis, M. Sitnov, A. Sciola, and S. Wing (2026), Vast space, sparse data: An AI answer to twin space weather challenges, Eos, 107, https://doi.org/10.1029/2026EO260188. Published on 11 June 2026. Text © 2026. The authors. CC BY-NC-ND 3.0
Except where otherwise noted, images are subject to copyright. Any reuse without express permission from the copyright owner is prohibited.

In November 2025, Cyclone Senyar generated extreme rainfall in parts of Sumatra, Indonesia, triggering thousands of landslides. Our new paper in the journal Current Biology demonstrates that these landslides might have a devastating impact on a critically endangered population of Tapanuli orangutan.

In November 2025, Cyclone Senyar brought extreme rainfall to large parts of Sumatra in Indonesia. I have written about this on previous occasions – the rainfall triggered vast numbers of landslides.

In my line of work, we often focus on the landslide impacts on the landscape, on human lives and on infrastructure. We rarely consider the impacts on th eanimal population. This is certainly a weakness that the Cyclone Senyar event brings to focus.

Part of the area devastated by the landslides is that slopes around the Batang Toru rover, an area of forest that is home to a rare species of orangutang. These great apes, Pongo tapanuliensis, live in a habitat known as the West Block of Tapanuli. There are only 800 individuals left in the wild, a situation that is highly precarious. The loss of even a small number of adults could tip the species towards extinction.

I was a part of a consortium of scientists that considered the landslide impacts of Cyclone Senyar on the habitat of these orangutangs. The results have just been published in the journal Current Biology (Meijaard et al. 2026) – the paper is open access and published under a creative commons license.

This image, from the paper, shows the landslide impacts of Cyclone Senyar:-

Before and after satellite imagery of the impacts of Cyclone Senyar. From: Meijaard et al. (2026).

In the study area of 71,161 hectares, the mapping indicates that there were 50, 185 individual landslides, covering a surface area of 8,303 hectares. This is about 11% of the forested area. We then estimate the likely loss of the orangutang population, which is likely to be in the range of 18-120 individuals, with a central estimate of 58 individuals. This is likely to have been a devastating loss for this highly endangered population.

This level of habitat loss might also be placing a severe pressure on the remaining population, so further fatalities are very possible through, for example, reduced food availability.

The intensity of the rainfall was almost certainly supercharged by climate change. The impacts of Cyclone Senyar are being replicated widely – and of course we are now in the northern hemisphere tropical cyclone season again.

Our paper makes some policy recommendations for this population of orangutans. First, the government of Indonesia needs to permanently protect this area of forest against mining , palm oil and hydropower developments. Ideally, the protected area should be expanded. Second, Indonesia needs support for biodiversity-recovery, hazard forecasting and ecological restoration planning.

Reference

Meijaard, E. … Petley. D. … et al. 2026, Extreme rainfall further endangers the world’s rarest great ape. Current Biology. https://doi.org/10.1016/j.cub.2026.05.029

Return to The Landslide Blog homepage Text © 2026. The authors. CC BY-NC-ND 3.0
Except where otherwise noted, images are subject to copyright. Any reuse without express permission from the copyright owner is prohibited.

AbstractSeismicity rates in west Texas and southeast New Mexico have increased over the last nine years and are in large part driven by subsurface wastewater disposal associated with oil-gas operations. Injection-induced seismicity is often explained as the result of fault weakening from fluid pressurization. However, fluid injection also induces poroelastic stresses from fluid-rock coupling, which in some cases are larger than the perturbations induced from pore pressure alone. In this work, we model in-zone changes in pore pressure and poroelastic stressing along four major fault zones that have hosted moderate-large ML 4 – 5 earthquakes in the Delaware Basin (DB) of west Texas. We leverage high-quality downhole pore-pressure measurements to constrain the in-situ hydraulic diffusivity and storativity. The data show that the deep injection interval has a storativity of ∼ 5×10−5 and a diffusivity between 23-65 m2/s, suggesting that this interval can pressurize rapidly and transmit fluid pressure efficiently. We view these as local hydraulic properties and use an ensemble modeling approach that accounts for a large range in diffusivities and storativities to model changes in Coulomb Failure Stresses (CFS). Our 2D fully coupled poroelastic models show that deep subsurface fluid injection can induce between ∼1-1000 kPa in CFS along basement rooted faults that penetrate the injection interval, with the largest values occurring in models that use the hydraulic properties inferred from our in-situ measurements. However, the induced changes in CFS are much smaller (∼ 20-30 kPa) when averaging over a large range in hydraulic properties. Irrespective of the model parameterization, the in-zone perturbations in CFS are dominated by changes in pore pressure, even at distances as far as 20-30 km from the nearest injection source. Our results highlight the importance of obtaining in-situ poromechanical measurements and indicate that such high-resolution measurements are critical to understanding subsurface stressing associated with fluid injection.

Climate change could make historically rare tropical storms in Southern California produce significantly more precipitation in the next few decades, and when they strike, landslides are likely to become a bigger risk across the region, according to new research in Nature Climate Change.

Strong and consistent evidence shows that the entire climate system is continuing to heat, driving rapid global warming. Human activities pushed global warming to 1.37°C in 2025, and its level is projected to surpass 1.5°C in about four years. Crucially, the rate at which heat is accumulating in Earth's system suggests high levels of future warming. These are some of the key findings from the latest Indicators of Global Climate Change (IGCC) report, published in Earth System Science Data.

For almost a century, researchers have known that vertical land motion—the lifting and sinking of the ground—affects sea level locally. As the ground sinks, the sea level rises relative to the land. Scientists also assumed this process generally occurred at a steady rate over time. But a research team that includes Thomas Wahl, a UCF researcher and associate professor in the Department of Civil, Environmental and Construction Engineering, has found that ground subsidence has undergone phases of variable change, creating significant implications for coastal communities.