Arctic sea ice occupies a pivotal position in the Earth’s climate machinery. Its bright, reflective surface bounces solar energy back into space, whereas the darker ocean water that becomes exposed when ice melts absorbs that energy instead. This contrast makes sea ice a powerful amplifier of climate change: less ice leads to more absorbed heat, which in turn leads to still less ice. The implications do not stop at the Arctic Circle. Shifts in sea ice extent can influence atmospheric circulation and oceanic patterns far beyond the polar region, which is why scientists, policymakers and industries alike place a premium on anticipating how much ice will cover the Arctic months before it happens.

Forecasting sea ice on seasonal timescales, however, has long been a stubborn problem. The Arctic is changing rapidly, and the processes that govern ice growth and melt span the atmosphere, the ocean and the ice itself, interacting across scales that challenge even sophisticated physics-based simulations. Francesco Paparella, inaugural director of Mubadala ACCESS at NYU Abu Dhabi and senior author of the study, described the challenge directly. “Forecasting Arctic sea ice several months ahead is a difficult problem, and an increasingly important one as the Arctic continues to change,” he said. “Our approach is deliberately simple, but it performs competitively with much more complex forecasting models.”

The core idea behind RAP is rooted in one of the oldest techniques in the forecasting tradition: analog methods. Rather than simulating the physics of the atmosphere, ocean and ice, the algorithm looks backward. It searches the historical record of Arctic sea ice extent for past episodes in which conditions resembled the present state of the system. Once it has identified those analogous situations, it examines what actually happened next in each case and uses those outcomes to construct a set of possible futures. The result is not a single deterministic prediction but an ensemble of forecasts, each representing a plausible trajectory grounded in what the Arctic has actually done before under similar circumstances.

That ensemble structure is what gives RAP its second major advantage: an honest accounting of uncertainty. When the individual forecasts in the ensemble diverge widely, the spread signals that the current situation is genuinely hard to predict, and users can treat the forecast with appropriate caution. When the forecasts cluster tightly, the method is effectively saying that history offers a clearer lesson about what comes next. This built-in uncertainty estimate is rare among simple statistical tools and addresses a persistent weakness in seasonal prediction, where a single number without a confidence range can mislead more than it informs. Paparella emphasized that this transparency is central to the design. “Importantly, it also provides an estimate of its own uncertainty, making it a useful and transparent benchmark for evaluating future forecasting methods,” he said.

The performance evidence comes from direct comparison with the operational forecasting community. The researchers found that RAP produced forecasts with a level of skill comparable to models used by the Sea Ice Prediction Network, the collaborative effort that coordinates seasonal sea ice outlooks from teams around the world. For September, the month when Arctic sea ice reaches its annual minimum and the most closely watched target in seasonal forecasting, RAP’s forecast error was comparable to that of 34 models used for seasonal prediction. Matching the median performance of a few dozen dedicated forecasting systems with a method that relies on nothing more than the historical ice record is a striking result, and it raises pointed questions about how much of the available predictive signal the complex models are actually capturing.

The contrast with physics-based approaches is worth unpacking. Conventional seasonal sea ice models simulate the dynamics of the atmosphere, the ocean and the ice pack, solving equations for momentum, heat and mass transfer across a gridded representation of the Arctic. These simulations demand substantial computational resources, expert maintenance and careful calibration, and their outputs can be difficult to interpret when they go wrong. RAP, by comparison, uses only the historical record of sea ice extent. Its logic can be inspected directly: here are the past situations that resemble today, here is what followed, and here is how spread out those outcomes were. That interpretability, the researchers argue, is not a cosmetic virtue but a scientific one, because it makes the method’s assumptions visible and its failures diagnosable.

This is precisely why the team proposes RAP as a benchmark for the next generation of forecasting systems, including the rapidly proliferating artificial intelligence models now being applied to climate prediction. Machine learning approaches to sea ice forecasting have grown in number and ambition, but evaluating them rigorously requires a baseline that is simple, reproducible and honest about its uncertainty. “The value of RAP is not that it replaces more sophisticated models, but that it gives us a clear standard against which they can be tested,” Paparella said. “If a much more complex model cannot outperform such a simple approach, that tells us something important about how much additional predictive information that complexity is providing.” In other words, any model claiming an edge over RAP must demonstrate that its added machinery translates into measurably better forecasts, not just added sophistication.

The study, published in Scientific Reports under the title describing random analog prediction as a benchmark for seasonal Arctic sea ice extent forecasting, was based on an observational analysis of the sea ice record rather than new field measurements. The work also carries a regional significance that extends beyond the science itself. The United Arab Emirates has been expanding its investment in polar and Arctic research, and Mubadala ACCESS was established as a center dedicated to climate and environmental science in that context. Paparella noted that RAP could support the country’s growing polar research activities by providing a simple, low-cost tool for seasonal sea ice forecasting, one that institutions without access to supercomputing infrastructure could still run and interpret.

The broader lesson of the study resonates across the data-driven sciences. As models grow larger and more intricate, the community needs reference points that are easy to understand and hard to game, and a method that predicts the future by remembering the past provides exactly that. RAP does not claim to capture the physics of a warming Arctic, and its skill will ultimately depend on how the region’s behavior continues to evolve as conditions move beyond the range of the historical record. But as a measuring stick, it sets a bar that any seasonal forecasting system, whether built on fluid dynamics equations or neural networks, must now clear. For a problem as consequential as anticipating the retreat and advance of Arctic sea ice, a transparent tool that matches dozens of complex models while quantifying its own uncertainty is a meaningful addition to the forecasting toolkit, and a reminder that in climate science, elegance and rigor are not opposites.

Subject of Research: Seasonal forecasting of Arctic sea ice extent using an analogue-based statistical algorithm

Article Title: New NYU Abu Dhabi algorithm could predict Arctic sea ice up to nine months ahead, offering new climate insights

Article References: New NYU Abu Dhabi algorithm could predict Arctic sea ice up to nine months ahead, offering new climate insights. (n.d.). Original publication

Image Credits: AI Generated

DOI: Not provided

Keywords: Arctic sea ice, seasonal forecasting, Random Analogue Predictor, NYU Abu Dhabi, climate science, uncertainty quantification, Sea Ice Prediction Network, Scientific Reports, analog forecasting, polar research, climate modeling, benchmark models

The bibliometric half of the study is a feat of patient data work. The team searched the Scopus database for English-language articles and reviews containing the terms “cloud physics” or “cloud microphysics” published between 1970 and 2020, filtered them to Earth and planetary science journals, and winnowed the initial 8,199 candidates down to 6,987 papers. For each one, they extracted author names, affiliations, countries, and funding information, then used an automated Python script together with the genderize.io service to assign a probable binary gender to first authors and co-authors, accepting only inferences with a certainty estimate above 75 percent. The authors are candid about the limits of this approach: name-based inference cannot capture non-binary or gender-diverse identities, and the survey itself confirmed that at least three respondents identify outside the gender binary. Still, the scale of the dataset makes the broad patterns difficult to dismiss.

Those patterns are stark. Women first authors barely appear in the cloud physics literature before the mid-1990s; only after roughly 1997 does their presence as lead authors become clearly visible, and even by 2020 just about 17 percent of studies in the field were led by women. The geographic skew is even more pronounced. Seventy-seven percent of all analyzed papers were led by authors from just seven countries: Canada, China, France, Germany, Japan, the United Kingdom, and the United States. Scientists based in the tropics, defined as those living and working between 23.5 degrees north and 23.5 degrees south, accounted for only about 5 percent of publications, with most of those coming from India, Brazil, Taiwan, Mexico, Indonesia, and Thailand. For a region that drives much of the planet’s precipitation and convection, the numbers amount to a scientific blind spot with colonial roots.

The most provocative finding concerns what the authors call parachute science. Of the 6,987 papers analyzed, only 44 were explicitly performed in the African tropics or used data from that region, and just seven of those, a mere 16 percent, included a tropics-based scientist as a co-author. In other words, when wealthy institutions send researchers to sample the atmosphere of lower-income countries, the local experts who make that work possible are frequently left off the byline. The pattern echoes a broader analysis of the Nature Index, which found that between 2015 and 2022 only 2.7 percent of articles across 82 natural science journals featured collaborations between higher-income and lower-income countries, and that in those articles richer countries contributed three authors for every one from a poorer nation. The study’s authors caution that excluding non-English papers may have shaped this conclusion, but the signal is consistent with decades of documented extractive practice.

The survey data add a human dimension to the statistics. Distributed before the 2021 virtual International Commission on Clouds and Precipitation meeting hosted from Pune, India, the 42-question instrument drew 231 responses, reduced to 198 after quality checks, covering an estimated 30 to 47 percent of the community’s active membership. Respondents skewed young, with 64 percent under 40, and 57.1 percent identified as men, 41.4 percent as women, and 1.5 percent outside the binary. Twenty-three percent of respondents identified as belonging to a historically and currently excluded group, and nearly half of those, 21 of 46, said that identity had negatively affected their careers through collaborations, promotions, publishing, funding, salary, or citations. Ten respondents reported disabilities, and six of them described feeling excluded because of it.

The numbers on harassment and hostility are difficult to read. One third of all respondents said they had experienced bullying or harassment from someone in the cloud physics community, and the burden fell unevenly: 44 percent of women reported such experiences compared with 27 percent of men, a statistically significant difference. Similarly, 37 percent of women versus 10 percent of men reported working in a hostile environment tied to their gender, ethnicity, sexual orientation, or disability, a moderate statistical association. Among LGBTQIA+ respondents, the share who did not feel included in the community rose to 44 percent, compared with 30 percent overall. The authors interpret these patterns through the lens of intersectionality, the framework developed by legal scholar Kimberlé Crenshaw, in which overlapping marginalized identities compound the likelihood of discrimination, while noting that small subgroup sizes limit statistical confidence in some of the finer breakdowns.

Not every result is grim. When the team traced individual Scopus author IDs to measure how long researchers stayed active in the field, they found that after 1995 there was no meaningful difference between men and women in how many years they remained, judging by publication activity. Women are not dropping out faster than men, contrary to the leaky-pipeline narrative that dominates discussions of academic gender gaps. The problem, instead, is at the entrance: even in the most recent five-year cohort, between 2015 and 2020, only about a third of scientists newly entering the field were women. The authors also found that papers led by women were published in journals with impact factors statistically comparable to those led by men, whereas papers from tropics-based scientists landed in lower-impact journals on average, a gap the authors link partly to language barriers, editorial desk rejections, and pressure to publish in local, non-indexed venues.

One intriguing secondary finding suggests a practical lever for change. Papers led by women included a better balance of women co-authors than papers led by men, mirroring a recent result from ice-core science showing that women-led studies had 20 percent more women co-authors. Tropics-based first authors likewise tended to bring in other tropics-based collaborators. Whether this reflects deliberate mentorship or homophily, the implication is the same: who holds the lead-author position shapes who else gets included. Putting more women and more scientists from the Global South into senior and leading roles may be one of the most direct mechanisms for diversifying the authorship base of the field, which in turn feeds into who designs the field campaigns and whose data enter the models.

The stakes extend well beyond fairness. The tropics host extraordinary biodiversity, biomass burning, dust emissions, and the vast convective systems of the Intertropical Convergence Zone, yet input observations for climate models are often calculated rather than actually sampled there. Excluding the scientists best positioned to collect those measurements is both an ethical failure and a scientific one, degrading the very data needed to validate satellite retrievals and improve projections for the regions already being hit hardest by climate change. The authors argue that an equity-centered approach, grounded in ethical responsibility rather than a business case for diversity, is what the moment demands, a stance aligned with the International Association for Promoting Geoethics.

The paper closes with concrete recommendations: institutional diversity training that extends beyond recruitment, funded participation of Global South early-career scientists in major conferences, unconscious-bias training for reviewers, gender-balanced hiring and editorial boards, journal policies against parachute science, and funding priorities for projects conceived and led by tropics-based researchers, alongside support for networks such as the Latin America Early Career Earth System Scientist Network and the cloudsfor.africa initiative. The journal’s editor went further, writing that the findings concern all of geoscience and deserve to be read by every geoscientist. For a community whose models must describe the whole planet, the message is unambiguous: the science of clouds cannot be done well by a fraction of the world’s scientists sampling a fraction of its sky.

Subject of Research: Diversity, equity, and inclusion in the cloud physics research community

Article Title: The state of diversity, equity, and inclusion in the cloud physics community

Article References: Ladino, L. A., Ardon-Dryer, K., Pereira, D. L., Proske, U., Ramirez-Diaz, Z., Velicu, A., & Kanji, Z. A. (2026). The state of diversity, equity, and inclusion in the cloud physics community. Geoscience Communication, 9(3), 415-436. https://doi.org/10.5194/gc-9-415-2026

Image Credits: AI Generated

DOI: 10.5194/gc-9-415-2026

Keywords: cloud physics, diversity, equity, inclusion, geosciences, parachute science, gender gap, Global South, tropics, bibliometrics, harassment, climate science

The research, led by Malak Sadki of Magellium in France with colleagues from CLS, Météo-France, the European Space Agency Climate Office, and the University of Toulouse, centers on products from ESA’s Climate Change Initiative River Discharge project. These datasets span the years 2000 to 2020 and come in three flavors. The first is water surface elevation, measured directly by radar altimetry from a fleet of missions including ERS-1 and -2, Envisat, SARAL, the Topex-Poseidon and Jason series, Sentinel-3, Sentinel-6A, and CryoSat-2. The second is discharge derived from those altimetry measurements through rating curves calibrated against historical gauge data. The third, and in some ways the most intriguing, is discharge estimated from multispectral imagery, exploiting the fact that the reflectance of water differs sharply from that of land, so the wetted area near a river channel tracks how much water the river is carrying.

To test these products rigorously, the team assimilated them into two fundamentally different modeling frameworks. The first is CTRIP, the river routing component of the ISBA land surface system used in the CNRM-CM6 climate model that contributed to the IPCC Sixth Assessment Report. CTRIP runs on a global grid at one-twelfth of a degree resolution and, crucially, is uncalibrated: its parameters come from physical reasoning and global datasets rather than fitting to observations, which makes it consistent everywhere but vulnerable to structural errors. The second is MGB, a semi-distributed, calibrated hydrological model designed for regional applications and operational flood forecasting, which explicitly simulates floodplains, wetlands, and the evaporation that occurs when floodwaters spread across them. Each model carries its own ensemble Kalman filter data assimilation system, HyDAS for CTRIP and HYFAA for MGB, giving the researchers two independent laboratories in which to evaluate the same satellite data.

The experimental stage was deliberately chosen for contrast. The Niger Basin, draining roughly 1.5 million square kilometers across nine countries, is dominated by strong monsoon-driven seasonality and the Inner Niger Delta, a vast floodplain in Mali that dissipates enormous volumes of water through absorption and evaporation. The Congo Basin, the second largest on Earth at over 3.7 million square kilometers, behaves almost oppositely: its mean annual flow of around 40,500 cubic meters per second is remarkably stable from year to year, buffered by an immense network of tributaries, lakes, and wetlands. Crucially, the Congo has lost most of its in-situ gauging stations since the 1960s, making it an ideal proving ground for satellite-only approaches, while the Niger retains a comparatively dense gauge network for validation.

The headline result is that assimilating discharge data generally outperformed assimilating water surface elevation anomalies, because discharge is the variable the models actually compute. When observations speak the model’s native language, corrections propagate smoothly and physically consistently through the river network. In the Niger Basin, altimetry-derived discharge assimilation into MGB produced the strongest gains of the entire study, lifting the median Nash-Sutcliffe Efficiency to 0.83 and the correlation coefficient to 0.94. Water surface elevation assimilation, by contrast, sometimes introduced noise, with visible jumps in simulated discharge after each update, a phenomenon the authors call the hashed effect. This occurs when corrections are applied intermittently and the model drifts back toward its biased state between updates, and it worsens when the relationship between water level and discharge, the rating curve, differs between the model and the observations.

That rating curve mismatch proved to be the crux of the elevation-versus-discharge debate. At Kinshasa on the lower Congo, the MGB model produces a nearly flat rating curve, while the altimetry observations show water level swings of up to seven meters for the same range of flows. Converting observed elevation anomalies into discharge corrections under these conditions introduced serious errors, at times driving the filter to diverge and even produce physically impossible negative discharge values. Yet the story is not one-sided. In the uncalibrated CTRIP model, whose open-loop simulation over the Niger overestimates high flows by up to 2.3 times because it lacks evaporation from the Inner Delta, elevation anomaly assimilation actually delivered better statistical scores, largely because its lower uncertainty gave those observations more weight in the Kalman filter when the model’s own rating curves happened to align well with the satellite data.

Perhaps the most striking finding concerns temporal density rather than accuracy. The multispectral discharge product is noisier than its altimetry-based counterpart, but it is available far more frequently, and that frequency transformed its value. In the Niger Basin, assimilating multispectral discharge cut the median bias ratio from 1.2 to nearly 1 in MGB and from 2.3 to 1.78 in CTRIP, and pushed the Kling-Gupta variability component toward 1.0, meaning the models finally captured the short-term ebb and flow of the rivers. The mechanism is intuitive: frequent updates prevent the model from reverting to its biased state, steadily correcting both high-flow and low-flow errors throughout the hydrological year. At Koulikoro on the upper Niger, this dense sampling reduced bias by 17 to 20 percent compared with the open loop. For climate studies concerned with floods, droughts, and the evolution of water resources, capturing that internal variability may matter as much as getting the average right.

The Congo Basin, however, delivered a sobering counterpoint. Improvements there were modest and heavily dependent on product quality. The available CCI stations provide weak spatial coverage of a basin whose outlet discharge integrates contributions from four major tributary systems, and at Bangui the multispectral product failed to reproduce the hydrological signal at all, with low correlation and no meaningful seasonality. Assimilating that noisy, phase-shifted data degraded model performance rather than improving it, a reminder that temporal density is only an asset when the underlying observations are trustworthy. The authors also note that uncertainty handling matters in subtle ways: because discharge uncertainty is proportional to flow magnitude, high-flow outliers are naturally down-weighted by the filter, but during low flows the absolute uncertainty shrinks and the filter may overfit isolated erroneous observations.

The dual-basin, dual-model design allowed the team to separate what is universal from what is context-dependent. Discharge assimilation is more physically consistent than elevation assimilation in models whose purpose is to simulate flow, and high-frequency data reliably reduces bias and improves variability when quality permits. But the magnitude of improvement depends on the interplay between model structure, observation sampling, and basin hydrology. In the calibrated MGB, assimilation refined already decent simulations; in the uncalibrated CTRIP, it reduced bias but could not conjure missing physics such as floodplain evaporation. In the strongly seasonal, well-observed Niger, gains were clear and diagnosable; in the stable, sparsely gauged Congo, they were harder to detect and more fragile. The researchers also caution that Congo validation scores deserve careful interpretation, since the same long-term records served both product generation and model evaluation.

Looking forward, the team outlines a roadmap for the next phase of the CCI Discharge project: merging altimetry and multispectral discharge streams to combine spatial reach with temporal density, exploiting the SWOT mission’s detailed measurements of water surface slope, refining localization schemes that currently rely on 20-year spatial correlation statistics, and exploring joint state-and-parameter estimation using river width and slope observations. They also plan smoother Kalman filter methods to suppress the hashed effects that plague intermittent updates, and dynamic localization distances that could adapt to human interventions like dams and irrigation. For a planet where climate change is reshaping river regimes faster than gauges can be installed, the message of this study is quietly revolutionary: the satellites already overhead, properly assimilated, can give the world’s water models the memory and the fidelity they have long lacked.

Subject of Research: Assimilation of ESA Climate Change Initiative satellite discharge and water surface elevation data into large-scale river routing models over the Niger and Congo basins

Article Title: Improving large-scale river routing models with ESA long-term CCI discharge data assimilation: advancing accuracy for climate studies

Article References: Sadki, M., Noual, G., Munier, S., Pedinotti, V., Verma, K., Albergel, C., Biancamaria, S., & Andral, A. (2026). Improving large-scale river routing models with ESA long-term CCI discharge data assimilation: advancing accuracy for climate studies. Hydrology and Earth System Sciences, 30(19), 6159-6187. https://doi.org/10.5194/hess-30-6159-2026

Image Credits: AI Generated

DOI: 10.5194/hess-30-6159-2026

Keywords: data assimilation, river discharge, ESA Climate Change Initiative, satellite altimetry, hydrological modeling, Niger Basin, Congo Basin, ensemble Kalman filter, CTRIP, MGB model, climate studies, remote sensing

The core insight of the method, published in the journal Nonlinear Processes in Geophysics, is deceptively simple. Instead of manipulating the inputs or outputs of a trained network, as most previous approaches to the out-of-distribution problem have done, the Potsdam team manipulates the operator itself, meaning the network’s weights and biases. Their technique unfolds in three steps. First, a trained network, which the authors call the parent model, is fine-tuned on individual data points or subsets of its own training data, and the resulting deviations in every weight and bias are carefully recorded. Second, a regression is established between these weight anomalies and suitable predictors drawn from the data. Third, that regression is extrapolated to the application data, generating a new network, the child model, whose weights are tailored to the conditions of the target domain.

The reasoning behind the approach rests on a subtle property of neural networks. When a trained model is fine-tuned on a single training example, its weights shift only slightly, but those shifts are meaningful: they encode the network’s sensitivity to that particular sample. Ideally, retraining on many individual samples produces a cloud of related models scattered around the parent in parameter space, and that cloud encodes how the network’s internal parameters respond to changes in the data. By fitting a regression to this cloud and extrapolating it beyond the training distribution, the researchers can generate networks whose weights are adapted to regimes that were never part of the original training. In effect, the method turns the network’s own training sensitivity into a compass pointing toward uncharted territory.

To demonstrate the technique, the authors designed three demanding test cases drawn from climate and Earth sciences. The first tackles one of the most consequential problems in modern climate science: the potential tipping of the Atlantic Meridional Overturning Circulation, the vast system of ocean currents that includes the Gulf Stream. Using simulation data from a complex climate model in which the circulation collapses under freshwater forcing, the researchers trained a convolutional neural network to predict the current’s strength from two-dimensional maps of ocean velocity. Crucially, the network was trained only on data from before the tipping point, which occurs around year 1800 of the simulation, yet it had to produce accurate predictions during and after the collapse, a genuinely severe out-of-distribution challenge.

The results were striking. The parent model, trained conventionally, remained stubbornly biased toward the conditions it had learned, producing outputs that lagged far behind the actual collapse of the circulation. The child model, whose weights had been extrapolated using predictors based on the leading empirical orthogonal functions of the velocity fields, tracked the tipping and its aftermath far more faithfully. Because the experiment involves an element of randomness inherent to neural network training, the authors repeated it several hundred times with varying training windows. Across the ensemble, the child models consistently showed lower root mean squared errors than their parents, and the advantage grew larger as conditions diverged further from the training data. Notably, simple linear regressions of the weights performed as well as or better than higher-order polynomials, which tended to assign noisy, spurious slopes to terms that barely varied within the training range.

The second experiment addressed a spatial rather than temporal challenge: estimating the density of seawater, a highly nonlinear function of salinity, temperature, and pressure governed by an empirical international standard known as TEOS-10. The researchers trained a small feed-forward network on gridded observations from the World Ocean Atlas, but only above 2000 meters depth, the limit of the Argo float array that dominates modern ocean monitoring. Below that boundary, the network had to operate entirely on its own. Here the weight-prediction method proved remarkably stable: in the deep ocean, the child models reduced the error of the parent models by 50 to 100 percent, an average improvement of about 2 kilograms per cubic meter that the authors describe as substantial by oceanographic standards. The improvement was largest precisely where the parent model performed worst, in the cold, high-pressure depths far removed from the training data.

The third experiment pushed the method into cross-regime territory. The researchers trained a network to estimate uncertainties in global wind velocity reanalyses, using two widely used atmospheric products, ERA5 from the European Centre for Medium-Range Weather Forecasts and CFSv2 from the United States National Centers for Environmental Prediction. The training data came from a single oceanic grid point in the Southern Ocean, but the child models were asked to predict wind uncertainty over the continents, where boundary-layer processes, surface roughness, and convection differ fundamentally from conditions over water. Even in this extreme scenario, the approach delivered: the global mean error over land dropped by roughly 10.7 percent in the headline experiment, with repeated trials showing improvements between minus 2 and 14 percent and an average of 5.25 percent. The largest gains appeared over southeastern Australia, eastern Brazil, the East African Highlands, Indonesia, and southeastern Russia.

Perhaps the most philosophically intriguing result came from the third experiment, where the researchers used a neural network itself to generate the child models, rather than a conventional regression. Using one network to fix the out-of-distribution failures of another might seem circular, since the terrestrial data is equally foreign to both. The authors argue that the trick works because it splits the problem into two parts: solving the task and adapting the task to a new domain. That separation, they suggest, amounts to a structural implementation of prior knowledge that a single network cannot yet develop on its own. They also note that nonlinear regime shifts in a network’s input and output space can translate into simpler, even linear, changes in weight space, because nonlinear activation functions allow complex output changes to be realized through modest parameter adjustments distributed across layers.

The method is not without caveats. Because it inherits the notorious stochasticity of neural network training, not every parent model produces a better-behaved offspring; some parents sit in loss landscapes so flat or so narrow that fine-tuning barely moves their weights, and occasionally a child model deteriorates on data resembling the training conditions. The authors recommend ensemble approaches for now and identify parent-model selection, predictor pruning, and regression validation as priorities for follow-up research. They also caution that the choice of predictors and regression methods must be guided by domain knowledge, and that for very large networks, low-rank adaptation techniques may be needed to keep the per-weight regressions computationally feasible. In their appendix comparisons, none of the standard extrapolation methods they tested, including Gaussian processes, random forests, and support vector machines, outperformed the child models on the out-of-distribution tasks.

Even so, the implications reach well beyond oceanography and climate science. The authors suggest the approach could benefit any field where training data are spatiotemporally limited and conditions at the application site differ from those at the training site, from astrophysics and biology to medicine, where learned diagnostic skills might be transferred to patient groups underrepresented in the training data. They add that the technique may even improve networks that are undertrained, overfitted, or deliberately simple, even when no distribution shift is present. In an era when artificial intelligence is being asked to forecast climates it has never observed, monitor oceans it has never sampled, and anticipate tipping points that have not yet arrived, methods that let neural networks rewrite their own internal parameters for the world they actually face may prove essential.

Subject of Research: A weight-prediction method for improving neural network performance on out-of-distribution data in the geosciences

Article Title: Conditional updates of neural network weights for increased out of training performance

Article References: Saynisch-Wagner, J., & Sari, S. R. (2026). Conditional updates of neural network weights for increased out of training performance. Nonlinear Processes in Geophysics, 33(4), 503-520. https://doi.org/10.5194/npg-33-503-2026

Image Credits: AI Generated

DOI: 10.5194/npg-33-503-2026

Keywords: neural networks, out-of-distribution, machine learning, climate science, AMOC tipping, oceanography, weight prediction, seawater density, wind reanalysis uncertainty, GFZ Potsdam, extrapolation, geosciences

A new study by Jennifer Jacquet and Loredana Loy of the University of Miami, published in PLOS Climate, examines how the National Cattlemen’s Association, the oldest and largest trade association for U.S. cattlemen, responded to this emerging science between 1986 and 1996. Drawing on 154 internal organizational documents from the association’s archives at the American Heritage Center in Wyoming, along with media coverage and government reports, the researchers reconstructed a decade-long campaign in which the industry treated global warming not as an environmental problem to be mitigated, but as a public relations threat to consumer demand. Strikingly, the authors found no evidence that the association ever researched or promoted any of the emissions-reduction options that federal scientists had laid out.

The stakes were scientifically significant. Methane absorbs thermal infrared radiation far more efficiently than carbon dioxide, making it roughly 80 times stronger per unit mass over a 20-year timescale. In 1989, the Environmental Protection Agency convened a workshop on methane emissions from livestock and published a report concluding that managed livestock, particularly ruminants, accounted for some 15 to 20 percent of annual methane emissions. Crucially, a representative of the cattlemen’s association, Richard Neal, attended that workshop, and the report stated that its findings reflected consensus among all attendees. The EPA report also catalogued practical opportunities to reduce emissions: alternative feeding practices, feed additives, nutrient supplements, manure management, and even revisions to the U.S. meat grading system so that leaner, lower-emission meat would not be penalized.

Yet the real threat to the industry, as the association saw it, came not from regulation, which never materialized, but from voluntary dietary change. Civil society campaigns were urging Americans to eat less beef, partly on climate grounds. John Robbins’s bestselling 1987 book Diet for a New America linked meat consumption to global warming, and Jeremy Rifkin launched the Greenhouse Crisis Foundation in 1989 and the Beyond Beef campaign in 1992, which called on consumers in 16 countries to cut their beef consumption in half. Internal association documents show the organization understood exactly what was at stake. Its 1989 Strategic Plan on the Environment, a 17-page document listing 22 environmental issues relevant to the cattle industry, including atmospheric emissions, laid out a plan to counter the Greenhouse Crisis Foundation by circulating what it called factual information, working with legislative influencers, and commissioning position papers from scientific experts.

The centerpiece of the industry’s response was what public relations scholars call third-party validation. At a 1991 convention in Dallas, an executive from Edelman Public Relations advised the association to disseminate technical information to academics and policymakers who would become third-party spokespersons, repeating industry messages with a credibility the industry itself could not achieve. But the association was already ahead of that advice. By the end of 1989 it had contracted Floyd M. Byers, an animal scientist at Texas A&M University, who produced reports, op-eds, and media statements that reframed the science of cattle methane. His reports contained no methods section, were not reproducible, and do not appear to have been peer-reviewed, yet they generated a statistic that would echo through mainstream media for years.

That statistic was the claim that methane from all U.S. beef cattle amounted to less than 0.5 percent of global methane emissions. The figure appeared in a 1990 New York Times article that opened with the line, In America, the cow is on trial, and in a USA Today op-ed titled Eating meat won’t harm the environment. The mechanics of the number reveal its fragility. Byers appears to have used as his denominator the total of all methane sources, both anthropogenic ones like rice paddies and natural ones like tundra and wetlands, whereas the EPA had focused only on the anthropogenic emissions humans could actually control. His numerator counted only U.S. beef cattle, excluding dairy cows, and multiplied by an undisclosed emissions factor that appears to have been lower than reliable estimates of the time. Michael Gibbs, lead author of the EPA report, told the Times that the figure was rather low and emphasized the importance of global emissions. Notably, the association’s own representative had sat in the workshop where the EPA’s much larger consensus estimate was produced, and there is no record of anyone contesting it there.

The association amplified the 0.5 percent figure through newsletters, press releases, and paid media analysts. It hired the firm CARMA, computer-aided research and media analysis, to track how its talking points were landing. A 1991 CARMA report found that among 196 news articles referencing methane emissions, 25 had adopted the claim that U.S. beef accounted for only 0.5 percent of global methane. The association even commissioned a report from Harvard economist Robert Stavins and colleague David Reishus, The Real Economics of Beef Production, which opened with the statement that calls for curtailing domestic beef production were unwarranted, and repeated the 0.5 percent figure while citing Byers. The association also spent $100,000 of Beef Board funds on a full-page advertisement, Every Day is Earth Day for American Cattlemen, which ran in USA Today and the New York Times and reached an estimated 20 million readers around Earth Day 1990.

Beyond minimizing its own emissions, the association deployed a four-pronged discursive strategy that researchers classify as challenging the problem, the causation, the policy, and the messengers. Its newsletters ran articles titled No Global Warming and No Global Warming Threat, and spokespeople told reporters that scientists continued to debate whether warming was occurring. When causation could not be denied outright, the scale of analysis was shifted: termites, rice paddies, and wetlands were invoked to make U.S. cattle look trivial, and the efficiency of intensive American production systems was contrasted with emissions in developing countries said to be five to seven times greater. The association also went after its critics directly, calling Robbins’s book laced with conjecture and hearsay, describing Rifkin as an anti-science activist, and founding the Food Facts Coalition, which asserted that Beyond Beef showed a total disregard for science. In one remarkable episode, the association persuaded publisher Random House not to reprint a Teenage Mutant Ninja Turtles children’s book whose entry for the letter M explained that cattle expel methane and encouraged readers to eat less meat.

What the archive does not contain is perhaps the most damning finding. The researchers searched every document for the emissions-mitigation strategies the EPA had identified: feed, feeding, supplements, hormones, manure, animal waste, and grading. Three of the industry-commissioned reports acknowledged in passing that nutritional management could reduce methane, but there is no mention anywhere in the association’s own materials of researching, encouraging, or implementing any of these changes. The one mention of grading-standard revisions made no connection to emissions. Instead, the association’s strategic plans focused on awards for environmental stewardship, media training for spokespersons, monitoring environmental groups, and distributing pro-industry materials to Congress, where a 1993 media analysis found more than a dozen members among the top 25 favorable sources for the industry.

The consequences of that choice are visible in the emissions record. According to the EPA, U.S. methane emissions fell by 19 percent between 1990 and 2022, driven by mitigation in landfills and the fossil fuel sector, while emissions from agricultural activities increased over the same period. Methane from enteric fermentation in U.S. beef cattle rose from 4,742 kilotonnes in 1990 to 5,018 kilotonnes in 2020, and emissions from manure management more than doubled. Jacquet and Loy argue that the association’s campaign, amplified by news media norms of balance and by headlines that made cow methane a punchline, helped make dietary change seem laughable as climate policy. The study is a case study, the authors suggest, in how public relations can reframe a substantive environmental problem into a mere image problem, and in how an industry facing inconvenient science can choose messaging over mitigation, a strategy whose success is measured in the methane still rising from American pastures four decades after the science first made it visible.

Subject of Research: The U.S. beef industry trade association's strategic communications response to scientific evidence linking cattle methane to global warming from 1986 to 1996

Article Title: Global warming and the National Cattlemen’s Association 1986–1996

Article References: Global warming and the National Cattlemen’s Association 1986–1996. (n.d.). https://doi.org/10.1371/journal.pclm.0001061

Image Credits: AI Generated

DOI: 10.1371/journal.pclm.0001061

Keywords: methane, global warming, beef industry, National Cattlemen's Association, public relations, climate disinformation, EPA, dietary change, trade associations, Texas A&M, climate science, livestock emissions

The international team, led by Min Chen and Zhiyi Zhu of Nanjing Normal University together with Thorsten Wagener, Niklas Boers and more than twenty colleagues across Europe, Asia, Australia and North America, argues that reproducibility must be rethought for the artificial intelligence era. In traditional Earth system modelling, reproducibility largely meant running the same code, on comparable hardware, with the same inputs, and obtaining the same outputs. Hybrid models break that tidy picture. Their numerical behaviour depends not only on deterministic physical solvers but also on stochastic training procedures, opaque learned components, versioned datasets and specialized computing environments that are difficult to document and share.

The authors identify three compounding problems. The first is numerical instability: neural network components embedded in climate and hydrological simulations can behave chaotically across tiny changes in hardware, software libraries or floating-point precision, so that two nominally identical runs produce divergent results. The second is procedural opacity: when a deep learning module replaces a physically derived parameterization of clouds, convection or runoff, the reasoning behind a prediction can become effectively inaccessible, making verification and error diagnosis far harder. The third is asymmetric access to computational resources: training frontier models demands GPU clusters and curated datasets that many institutions, particularly in lower-income countries, simply cannot obtain, narrowing who can check and extend published work.

The stakes extend beyond academic housekeeping. Hybrid Earth system models increasingly inform climate risk assessments, disaster early-warning systems and national adaptation strategies. If their results cannot be traced, tested and re-derived, the authors contend, the models risk becoming weakly verifiable black boxes. That would erode model traceability, weaken the cumulative build-up of knowledge across research communities and ultimately narrow the evidential basis available to bodies such as the Intergovernmental Panel on Climate Change and the policymakers who rely on its assessments.

At the core of the Perspective is a conceptual reframing. Rather than treating reproducibility as a post hoc checklist applied when a paper is published, the team proposes that it should be understood as a structural property of the modelling process itself. They argue for a theory of reproducibility assessment calibrated to the epistemology of hybrid systems, in which the question is not merely whether a numerical output can be replicated bit-for-bit, but whether the full chain of evidence — data provenance, training procedure, physical coupling, uncertainty characterization and code environment — can be independently reconstructed and evaluated.

To translate that principle into practice, the researchers introduce RHEM, short for Reproducibility in hybrid Earth system models, a reference guideline intended to govern transparent, trustworthy and reproducible model development. RHEM is conceived as a shared framework against which modelling groups can design, document and audit their workflows. Its criteria cover the life cycle of a hybrid model, from the definition of research questions and the curation of training data through the coupling of machine learning components with physical cores, and finally to release, versioning and community reuse.

Complementing the framework, the Perspective lays out practical pathways that operationalize its criteria into actionable measures. These include comprehensive documentation strategies for geosimulation experiments, provenance capture using standards such as W3C PROV, containerized software environments that pin library versions and hardware assumptions, and modular architectures that clearly separate learned components from physical ones so each can be tested independently. The authors also highlight physically informed constraints on neural networks, which enforce conservation laws and other analytic requirements to stabilize learned parameterizations, and benchmark datasets such as WeatherBench that allow head-to-head comparison of competing forecasting approaches.

The proposal draws on a wide body of prior work in computational science. The authors situate their argument in the broader reproducibility crisis documented across disciplines, citing surveys in which a majority of surveyed scientists reported failing to reproduce others’ experiments, along with manifestos for reproducible science, guidelines for artificial intelligence transparency, and studies showing that much computational hydrology has historically resisted independent replication. They also note existing community infrastructure — component-based modelling frameworks, virtual laboratories, FAIR principles for research software and open modelling portals — as building blocks on which RHEM-style governance can be constructed rather than starting from scratch.

What distinguishes the new contribution, the authors argue, is the explicit coupling of a conceptual structure with operational guidance tailored to hybrid systems. Earlier reproducibility efforts addressed either classical numerical models or machine learning research in isolation. Hybrid Earth system models inherit the pathologies of both: the chaos sensitivity and tuning culture of climate simulation, and the nondeterminism, dataset dependence and interpretability gaps of deep learning. By treating these as an integrated design problem, the team hopes reproducibility can move from an aspirational virtue to an engineered characteristic, embedded in how models are built, tested and documented.

The Perspective closes with a call for coordinated community action. Modelling centres, funding agencies, journals and cyberinfrastructure providers each hold part of the solution, from mandating reproducibility statements at submission to sustaining shared computing resources that level the field between well-funded and resource-constrained groups. If such measures take hold, the authors conclude, reproducibility can be established as a foundational principle of Earth system science in the artificial intelligence era — ensuring that as models grow more powerful, they also grow more accountable, traceable and worthy of the public trust placed in the climate science they support.

The reproducibility dilemma confronting hybrid models has deep historical roots in Earth science. As early as 1994, philosophers and modellers cautioned that numerical models of the Earth could never be verified in the strict sense familiar from laboratory science, because the systems they describe are open, unique and only partially observable. That insight, once confined to philosophical debates about validation, has acquired new urgency now that learned components trained on finite observational records sit inside simulation codes whose outputs shape billion-dollar adaptation decisions. The Perspective effectively extends an old epistemological caution into the computational present, arguing that the looseness of verification once tolerated in physical modelling cannot be compounded by the additional opacity of machine learning.

Computational studies illustrate how fragile reproducibility can be even without neural networks. Research on the EC-Earth3 climate model showed that a change in computing environment could alter simulation outputs in ways that complicated replication, while work on atmospheric regional modelling identified hydrodynamic noise as a fundamental limit on reproducibility, distinguishing irreducible chaos from avoidable procedural failures. These findings matter for the hybrid era because they supply a baseline: some divergence between runs is physically expected, and a sensible reproducibility standard must therefore separate meaningful nondeterminism from instability introduced by sloppy engineering, unstated library versions or uncontrolled floating-point behaviour. The roadmap proposed by Chen and colleagues builds directly on this distinction, seeking to pin down everything that is controllable so that remaining variability can be interpreted scientifically rather than dismissed as noise.

The timing of the Perspective reflects how quickly artificial intelligence has penetrated the field. Differentiable programming frameworks now allow entire Earth system processes to be embedded in gradient-based training loops, generative adversarial networks constrained by physics have been used to sharpen precipitation fields, and scale-adaptive generative models can downscale coarse model output with quantified uncertainty. Commercial and academic systems alike have demonstrated skillful medium-range global weather forecasting learned purely from data, and a foundation model for the Earth system has been proposed as a general-purpose substrate for downstream tasks. Each of these advances widens the surface area of reproducibility risk: more pretrained weights, more data pipelines, more hardware dependencies and more points at which undocumented choices silently enter the scientific record.

The authors’ emphasis on asymmetric access also connects to a broader equity conversation in geospatial and climate research. Surveys of reproducibility practices across the sciences have repeatedly found that most researchers regard replication as difficult, and geospatial researchers have documented how data licensing, proprietary platforms and uneven computational capacity compound the problem in their field. Explainable artificial intelligence for Earth observation has been examined partly through regulatory and societal lenses, underscoring that transparency is not merely a technical preference but increasingly a governance expectation. Framing reproducibility as a structural property rather than a publication-time obligation aligns with these developments, since transparency measures embedded early in a workflow are far cheaper and more effective than forensic reconstruction after the fact.

There is also a constructive reading of the moment. The same machine learning techniques that create reproducibility hazards can support the remedy: automated provenance capture, differentiable pipelines that make couplings explicit, and uncertainty-aware generative methods that state their own limitations. The Perspective’s insistence on coupling a theory of assessment with practical pathways acknowledges this duality. By treating documentation, containerization, modular separation of learned and physical components, and physically informed constraints as parts of a single design discipline, the framework aims to make trustworthiness a property that can be engineered, audited and incrementally improved — a stance the authors present as essential if the accelerating capabilities of hybrid Earth system models are to translate into durable, verifiable scientific knowledge.

Subject of Research: Reproducibility challenges and governance frameworks for hybrid Earth system models that integrate artificial intelligence

Article Title: Enhancing reproducibility in hybrid Earth system models

Article References: Chen, M., Zhu, Z., Wagener, T., Boers, N., Müller, R. D., Strobl, J., Camps-Valls, G., Batty, M., Jakeman, A. J., Kolditz, O., Nativi, S., Brovelli, M. A., Creutzig, F., Kumar, P., Whitehead, P., Barton, C. M., Liu, D., Ma, P., Ma, Z., … Lü, G. (2026). Enhancing reproducibility in hybrid Earth system models. Nature Machine Intelligence. https://doi.org/10.1038/s42256-026-01299-5

Image Credits: AI Generated

DOI: 10.1038/s42256-026-01299-5

Keywords: Earth system models, artificial intelligence, reproducibility, hybrid modelling, machine learning, climate science, Nature Machine Intelligence, RHEM framework, scientific transparency, model documentation, climate risk, computational reproducibility