Surface mass balance is the net result of accumulation and ablation at a glacier’s surface. Direct observations remain sparse and unevenly distributed, making it difficult to apply one calibrated model across large numbers of glaciers. Temperature-index models are widely used because they estimate accumulation from precipitation and melt from near-surface air temperature, often with separate degree-day factors for snow and ice. Their small parameter sets are interpretable and practical, but their accuracy can depend strongly on glacier-specific calibration. That dependence limits transfer to glaciers with different geometries and climate regimes, particularly because global geodetic mass-change products can be noisy at the scale of an individual glacier and do not represent exactly the same quantity as surface mass balance.

GlaUnTI retains the temperature-index structure but adds a shallow convolutional neural corrector. The underlying model represents solid precipitation, snow accumulation, fractional snow cover and temperature-driven melt using smooth functions that support automatic differentiation. The corrector learns spatial and temporal adjustments to five key quantities: precipitation sensitivity, the snow degree-day factor, the relative ice degree-day factor, the centre of the rain-to-snow transition and a temperature offset. Because these adjustments modify interpretable parameters rather than replacing the mass-balance equations, the model keeps a physical backbone while allowing data-driven compensation for systematic errors. In the comparison with the data-driven baseline, the temperature-index backbone used seven trainable parameters, whereas the gated recurrent unit had 961.

The neural component has separate spatial and temporal branches. The spatial branch uses an initial mass-balance estimate, sensitivities of that estimate to small temperature and precipitation perturbations, glacier outlines, surface elevation, elevation uncertainty and, in one variant, satellite-derived glacier facies. The temporal branch processes monthly temperature and precipitation totals. Their outputs are combined and resampled to generate glacier- and grid-cell-specific parameter corrections. The facies maps classify surface conditions such as ice, snow, debris, firn and refrozen-like material at the end of the ablation season when available; the study assembled 163 such maps for 68 glaciers. The two GlaUnTI variants used identical structural changes in this respect, with the facies-enabled version adding classified surfaces and classification confidence as predictors where the maps existed.

The study compiled 10,483 annual, 15,421 winter and 3,482 summer point mass-balance measurements, alongside 1,637 annual, 1,278 winter and 1,279 summer glacier-wide estimates. These records were combined with daily ERA5-Land temperature and precipitation, annual elevation estimates and time-varying glacier outlines. The data covered 78 glaciers, with 38 in the Alps, 21 in Scandinavia, nine in Svalbard, eight in Iceland and two in the Pyrenees. A glacier-level stratified split assigned 52 glaciers to training, 13 to validation and 13 to an entirely independent test set spanning contrasting climates and geometries. The test set contained 793 annual, 756 winter and 314 summer point measurements, together with 312 annual, 235 winter and 233 summer glacier-wide estimates. The authors used the glacier-wide estimates as weaker constraints than direct point observations because monitoring programmes derive them by extrapolating measurements across each glacier.

At point locations, the conventional temperature-index model produced an annual correlation of 0.854, a root mean square error of 1.707 metres water equivalent and a mean bias of +0.298 metres water equivalent. The purely data-driven gated recurrent unit performed worse, with a correlation of 0.798 and root mean square error of 2.066 metres. GlaUnTI without facies maps raised the correlation to 0.933 and reduced error to 1.095 metres, while the facies-enabled version achieved a correlation of 0.940, an error of 1.068 metres and a mean bias of just +0.009 metres. Seasonal results generally followed the same ranking, although the data-driven model retained the lowest winter bias. Per-glacier distributions showed generally lower errors and narrower spreads for the GlaUnTI models, indicating that the aggregate improvement was not confined to one glacier.

At the glacier-wide scale, the differences were also pronounced. The temperature-index baseline achieved an annual correlation of 0.606 and root mean square error of 0.805 metres water equivalent, while the data-driven recurrent network fell to a correlation of 0.331 and an error of 1.009 metres. The two GlaUnTI variants improved the correlation to 0.684 and 0.700 and reduced error to 0.651 and 0.627 metres, respectively. Including facies information produced moderate benefits for summer and annual glacier-wide estimates, although the evidence was limited: the facies-available evaluation contained 27 annual glacier-wide estimates and 32 annual point measurements. The researchers therefore treated the maps as useful auxiliary information rather than evidence of a large general gain. Their results are also consistent with the physical expectation that facies information can act as a proxy for processes such as surface-albedo contrasts, firn retention and debris cover, rather than serving as a causal driver of mass balance in its own right.

The hybrid model also approached the performance obtained when the conventional model was calibrated separately for each test glacier. In that deliberately favourable comparison, each temperature-index model used the target glacier’s observations from 1995–2012, was validated on 2013–2018 and was evaluated on 2019–2024. The locally calibrated model achieved an annual point-level correlation of 0.948, a mean bias of −0.308 metres water equivalent and a root mean square error of 0.905 metres. GlaUnTI, evaluated without glacier-specific fine-tuning, achieved a correlation of 0.944, a bias of +0.107 metres and an error of 0.883 metres. It also had lower absolute bias in winter and summer and comparable seasonal errors. The comparison shows that GlaUnTI approached local calibration in this test while its original spatial training used no observations from the target glaciers.

GlaUnTI nevertheless has important limitations. Its temperature-index backbone does not explicitly resolve radiation, debris effects, snow ageing, terrain shading or the transport of snow by ice flow, so some physical processes remain embedded in learned corrections rather than represented mechanistically. Point stakes may also fail to represent conditions across an entire 100-metre grid cell, particularly on glaciers with strong microclimate or topographic variation. Elevation observations contain gaps and uncertainty, and climate forcing has its own errors. The study’s sensitivity analysis addressed the choice of how elevations were extended outside the 2000–2019 observation window: replacing quadratic temporal extrapolation with backfilled values changed mass-balance estimates by less than ±2.5%, indicating that this extrapolation strategy was secondary to other error sources. The European scope and limited facies observations also mean that transfer beyond the sampled climates and glacier types remains to be tested.

Beyond prediction, the model’s differentiability enables analytical applications demonstrated in the study. It supports gradient-based calibration, input attribution, transfer learning, inverse optimisation and the propagation of climate-forcing uncertainty through long mass-balance trajectories. In one example on Grosser Aletsch in the Swiss Alps, the researchers optimized a temperature offset for 2015–2024 that made the model’s average annual glacier-wide surface mass balance for 2020–2024 equal zero. The resulting effective offset was −2.42°C. The authors emphasized that this was a reduced inverse experiment with fixed glacier geometry, not a complete climate-change scenario, and that the inferred offset could compensate for both forcing discrepancies and residual model biases. In another example, integrated gradients separated the temperature- and precipitation-associated contributions to annual anomalies for Schwarzberg. A separate uncertainty calculation propagated temporally correlated temperature and precipitation errors through cumulative mass-balance trajectories using first-order automatic differentiation.

The study’s broader implication is methodological rather than a claim that all local glacier processes have been resolved. The fully data-driven recurrent network had the weakest spatial transfer, while the physically structured temperature-index model generalized more reliably and benefited from a learned corrector that could adjust its key parameters. End-to-end differentiability also makes it possible to connect mass balance with differentiable ice-flow models, perform input attribution and ask what effective forcing perturbation would be needed to reach a specified balance target. The authors identify possible extensions involving albedo products, transient snowlines, satellite indicators of melt, more explicit energy-balance forcing and regional geodetic mass-change records. They caution that geodetic observations integrate processes beyond surface mass balance, including dynamic thinning, frontal ablation and basal melt, and that climate-forcing uncertainty should be represented explicitly rather than hidden within learned structural corrections.

Subject of Research: Transferable estimation of glacier surface mass balance using a hybrid physics–machine learning model

Article Title: GlaUnTI: a hybrid physics–machine learning model enables transferable glacier surface mass balance estimation

Article References: Maslov, K. A., Schellenberger, T., Persello, C., & Stein, A. (2026). GlaUnTI: a hybrid physics–machine learning model enables transferable glacier surface mass balance estimation. The Cryosphere, 20(10), 6045-6070. https://doi.org/10.5194/tc-20-6045-2026

Image Credits: AI Generated

DOI: 10.5194/tc-20-6045-2026

Keywords: glacier surface mass balance, hybrid modelling, machine learning, temperature-index model, European glaciers, climate forcing, remote sensing, neural corrector, snow accumulation, glacier melt, GlaUnTI, hybrid

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