Gong Cheng
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Previous Organizations / Institutes
The organizations / institutes from the scholar's other articles in the past ten years
Research Subjects
Research Interests
Gong Cheng has a total of 133 co-authors. Here are the top co-authors by number of shared publications:
- Mathieu Morlighem
Department of Earth Sciences, Dartmouth College, Hanover, NH, USA
50 shared publications - Mansa Krishna
Department of Earth Sciences, Dartmouth College, Hanover, NH 03755, USA
13 shared publications - Shfaqat Abbas Khan
DTU Space, Technical University of Denmark, Kgs. Lyngby, 2800, Denmark
13 shared publications - Martin Rückamp
Bavarian Academy of Sciences and Humanities, Munich, Germany
11 shared publications - Helene Seroussi
Thayer School of Engineering, Dartmouth College, Hanover, NH, USA
10 shared publications - Per Lötstedt
Uppsala Univ, Dept Informat Technol, SE-75105 Uppsala, Sweden
10 shared publications
Latest Publications
- Open Access
- 11 November 2026
- Published by Copernicus GmbH in The Cryosphere
- Vol. 20 (11) , 6453-6475
The implementation and effectiveness of Calving Algorithms in numerical ice models (CalvingMIP)
Ice calving plays a significant role in ice sheet mass loss. Predictions of future ice sheet mass balance require accurate representations of the calving process in numerical ice models to determine rates of future global mean sea level rise. Whilst calving has recently begun to be implemented in numerical models there has not yet been a systematic investigation into how this implementation compares between different ice flow models. The Calving Model Intercomparison Project (CalvingMIP) has been established to address this question by providing a framework of experiments to investigate the accuracy of simulated calving rates and how simulated properties at the calving front evolve over time. Our initial focus has been on how calving is implemented in models, therefore focusing on calving algorithms, rather than on how much ice should be calved at a particular time via a calving law. Our results, from thirteen different numerical modelling groups, show that the majority of calving algorithms implemented are able to accurately implement a given calving rate with an ice front that evolves smoothly throughout the calving process, albeit with slightly more variation of results in a more complex domain. These results show that we can have confidence in the models capacity to accurately implement calving laws in the future.
- 11 November 2026
- Published by Copernicus GmbH in The Cryosphere
- Vol. 20 (11) , 6453-6475
Open AccessThe implementation and effectiveness of Calving Algorithms in numerical ice models (CalvingMIP)
Ice calving plays a significant role in ice sheet mass loss. Predictions of future ice sheet mass balance require accurate representations of the calving process in numerical ice models to determine rates of future global mean sea level rise. Whilst calving has recently begun to be implemented in numerical models there has not yet been a systematic investigation into how this implementation compares between different ice flow models. The Calving Model Intercomparison Project (CalvingMIP) has been established to address this question by providing a framework of experiments to investigate the accuracy of simulated calving rates and how simulated properties at the calving front evolve over time. Our initial focus has been on how calving is implemented in models, therefore focusing on calving algorithms, rather than on how much ice should be calved at a particular time via a calving law. Our results, from thirteen different numerical modelling groups, show that the majority of calving algorithms implemented are able to accurately implement a given calving rate with an ice front that evolves smoothly throughout the calving process, albeit with slightly more variation of results in a more complex domain. These results show that we can have confidence in the models capacity to accurately implement calving laws in the future.
- Open Access
- 3 November 2026
- Published by Copernicus GmbH
- Vol. 2026, 1-18
Reply on AC1
Machine learning (ML) models are increasingly used in glaciology to replace computationally expensive model components or emulate specific processes. These approaches, however, are commonly developed as standalone tools or within application specific software and cannot be readily deployed within established ice-sheet modeling frameworks. We present ISSM-ML, a modular framework for integrating externally trained, Python-based ML models into the Ice-sheet and Sea-level System Model (ISSM). Model development and training remain within the Python ecosystem, while ISSM invokes selected ML modules at runtime and retains control of model initialization, numerical solvers, physical analyses, time stepping, and other procedures in its original workflow. We demonstrate the framework using a multilayer perceptron to replace the nonlinear velocity-dependent term in the Weertman basal-friction law. Two training strategies are evaluated, one using conventional scaling and the other using distribution-aware preprocessing. When the scaling-only model is coupled to ISSM, the stress balance solver fails to converge. The distribution-aware model, in contrast, allows ISSM to converge reliably throughout the transient simulation, with a good agreement with the analytical-friction reference. These results demonstrate that offline prediction accuracy alone is not sufficient for evaluating an ML component intended for numerical-model coupling. Their behavior must also be assessed within the simulation environment in which they operate, including convergence, stability, and feedback with evolving model states.
- 3 November 2026
- Published by Copernicus GmbH
- Vol. 2026, 1-18
Open AccessReply on AC1
Machine learning (ML) models are increasingly used in glaciology to replace computationally expensive model components or emulate specific processes. These approaches, however, are commonly developed as standalone tools or within application specific software and cannot be readily deployed within established ice-sheet modeling frameworks. We present ISSM-ML, a modular framework for integrating externally trained, Python-based ML models into the Ice-sheet and Sea-level System Model (ISSM). Model development and training remain within the Python ecosystem, while ISSM invokes selected ML modules at runtime and retains control of model initialization, numerical solvers, physical analyses, time stepping, and other procedures in its original workflow. We demonstrate the framework using a multilayer perceptron to replace the nonlinear velocity-dependent term in the Weertman basal-friction law. Two training strategies are evaluated, one using conventional scaling and the other using distribution-aware preprocessing. When the scaling-only model is coupled to ISSM, the stress balance solver fails to converge. The distribution-aware model, in contrast, allows ISSM to converge reliably throughout the transient simulation, with a good agreement with the analytical-friction reference. These results demonstrate that offline prediction accuracy alone is not sufficient for evaluating an ML component intended for numerical-model coupling. Their behavior must also be assessed within the simulation environment in which they operate, including convergence, stability, and feedback with evolving model states.
- Open Access
- 3 November 2026
- Published by Copernicus GmbH
- Vol. 2026, 1-18
Reply on CEC1
Machine learning (ML) models are increasingly used in glaciology to replace computationally expensive model components or emulate specific processes. These approaches, however, are commonly developed as standalone tools or within application specific software and cannot be readily deployed within established ice-sheet modeling frameworks. We present ISSM-ML, a modular framework for integrating externally trained, Python-based ML models into the Ice-sheet and Sea-level System Model (ISSM). Model development and training remain within the Python ecosystem, while ISSM invokes selected ML modules at runtime and retains control of model initialization, numerical solvers, physical analyses, time stepping, and other procedures in its original workflow. We demonstrate the framework using a multilayer perceptron to replace the nonlinear velocity-dependent term in the Weertman basal-friction law. Two training strategies are evaluated, one using conventional scaling and the other using distribution-aware preprocessing. When the scaling-only model is coupled to ISSM, the stress balance solver fails to converge. The distribution-aware model, in contrast, allows ISSM to converge reliably throughout the transient simulation, with a good agreement with the analytical-friction reference. These results demonstrate that offline prediction accuracy alone is not sufficient for evaluating an ML component intended for numerical-model coupling. Their behavior must also be assessed within the simulation environment in which they operate, including convergence, stability, and feedback with evolving model states.
- 3 November 2026
- Published by Copernicus GmbH
- Vol. 2026, 1-18
Open AccessReply on CEC1
Machine learning (ML) models are increasingly used in glaciology to replace computationally expensive model components or emulate specific processes. These approaches, however, are commonly developed as standalone tools or within application specific software and cannot be readily deployed within established ice-sheet modeling frameworks. We present ISSM-ML, a modular framework for integrating externally trained, Python-based ML models into the Ice-sheet and Sea-level System Model (ISSM). Model development and training remain within the Python ecosystem, while ISSM invokes selected ML modules at runtime and retains control of model initialization, numerical solvers, physical analyses, time stepping, and other procedures in its original workflow. We demonstrate the framework using a multilayer perceptron to replace the nonlinear velocity-dependent term in the Weertman basal-friction law. Two training strategies are evaluated, one using conventional scaling and the other using distribution-aware preprocessing. When the scaling-only model is coupled to ISSM, the stress balance solver fails to converge. The distribution-aware model, in contrast, allows ISSM to converge reliably throughout the transient simulation, with a good agreement with the analytical-friction reference. These results demonstrate that offline prediction accuracy alone is not sufficient for evaluating an ML component intended for numerical-model coupling. Their behavior must also be assessed within the simulation environment in which they operate, including convergence, stability, and feedback with evolving model states.
- Open Access
- 2 November 2026
- Published by Copernicus GmbH
- Vol. 2026, 1-18
Comment on egusphere-2026-6193 - No compliance with the policy of the journal
Machine learning (ML) models are increasingly used in glaciology to replace computationally expensive model components or emulate specific processes. These approaches, however, are commonly developed as standalone tools or within application specific software and cannot be readily deployed within established ice-sheet modeling frameworks. We present ISSM-ML, a modular framework for integrating externally trained, Python-based ML models into the Ice-sheet and Sea-level System Model (ISSM). Model development and training remain within the Python ecosystem, while ISSM invokes selected ML modules at runtime and retains control of model initialization, numerical solvers, physical analyses, time stepping, and other procedures in its original workflow. We demonstrate the framework using a multilayer perceptron to replace the nonlinear velocity-dependent term in the Weertman basal-friction law. Two training strategies are evaluated, one using conventional scaling and the other using distribution-aware preprocessing. When the scaling-only model is coupled to ISSM, the stress balance solver fails to converge. The distribution-aware model, in contrast, allows ISSM to converge reliably throughout the transient simulation, with a good agreement with the analytical-friction reference. These results demonstrate that offline prediction accuracy alone is not sufficient for evaluating an ML component intended for numerical-model coupling. Their behavior must also be assessed within the simulation environment in which they operate, including convergence, stability, and feedback with evolving model states.
- 2 November 2026
- Published by Copernicus GmbH
- Vol. 2026, 1-18
Open AccessComment on egusphere-2026-6193 - No compliance with the policy of the journal
Machine learning (ML) models are increasingly used in glaciology to replace computationally expensive model components or emulate specific processes. These approaches, however, are commonly developed as standalone tools or within application specific software and cannot be readily deployed within established ice-sheet modeling frameworks. We present ISSM-ML, a modular framework for integrating externally trained, Python-based ML models into the Ice-sheet and Sea-level System Model (ISSM). Model development and training remain within the Python ecosystem, while ISSM invokes selected ML modules at runtime and retains control of model initialization, numerical solvers, physical analyses, time stepping, and other procedures in its original workflow. We demonstrate the framework using a multilayer perceptron to replace the nonlinear velocity-dependent term in the Weertman basal-friction law. Two training strategies are evaluated, one using conventional scaling and the other using distribution-aware preprocessing. When the scaling-only model is coupled to ISSM, the stress balance solver fails to converge. The distribution-aware model, in contrast, allows ISSM to converge reliably throughout the transient simulation, with a good agreement with the analytical-friction reference. These results demonstrate that offline prediction accuracy alone is not sufficient for evaluating an ML component intended for numerical-model coupling. Their behavior must also be assessed within the simulation environment in which they operate, including convergence, stability, and feedback with evolving model states.
- Open Access
- 26 October 2026
- Published by Copernicus GmbH in EGUsphere
- Vol. 2026, 1-18
ISSM-ML: a modular framework for integrating machine learning models into the Ice-sheet and Sea-level System Model (ISSM v2026.4)
Machine learning (ML) models are increasingly used in glaciology to replace computationally expensive model components or emulate specific processes. These approaches, however, are commonly developed as standalone tools or within application specific software and cannot be readily deployed within established ice-sheet modeling frameworks. We present ISSM-ML, a modular framework for integrating externally trained, Python-based ML models into the Ice-sheet and Sea-level System Model (ISSM). Model development and training remain within the Python ecosystem, while ISSM invokes selected ML modules at runtime and retains control of model initialization, numerical solvers, physical analyses, time stepping, and other procedures in its original workflow. We demonstrate the framework using a multilayer perceptron to replace the nonlinear velocity-dependent term in the Weertman basal-friction law. Two training strategies are evaluated, one using conventional scaling and the other using distribution-aware preprocessing. When the scaling-only model is coupled to ISSM, the stress balance solver fails to converge. The distribution-aware model, in contrast, allows ISSM to converge reliably throughout the transient simulation, with a good agreement with the analytical-friction reference. These results demonstrate that offline prediction accuracy alone is not sufficient for evaluating an ML component intended for numerical-model coupling. Their behavior must also be assessed within the simulation environment in which they operate, including convergence, stability, and feedback with evolving model states.
- 26 October 2026
- Published by Copernicus GmbH in EGUsphere
- Vol. 2026, 1-18
Open AccessISSM-ML: a modular framework for integrating machine learning models into the Ice-sheet and Sea-level System Model (ISSM v2026.4)
Machine learning (ML) models are increasingly used in glaciology to replace computationally expensive model components or emulate specific processes. These approaches, however, are commonly developed as standalone tools or within application specific software and cannot be readily deployed within established ice-sheet modeling frameworks. We present ISSM-ML, a modular framework for integrating externally trained, Python-based ML models into the Ice-sheet and Sea-level System Model (ISSM). Model development and training remain within the Python ecosystem, while ISSM invokes selected ML modules at runtime and retains control of model initialization, numerical solvers, physical analyses, time stepping, and other procedures in its original workflow. We demonstrate the framework using a multilayer perceptron to replace the nonlinear velocity-dependent term in the Weertman basal-friction law. Two training strategies are evaluated, one using conventional scaling and the other using distribution-aware preprocessing. When the scaling-only model is coupled to ISSM, the stress balance solver fails to converge. The distribution-aware model, in contrast, allows ISSM to converge reliably throughout the transient simulation, with a good agreement with the analytical-friction reference. These results demonstrate that offline prediction accuracy alone is not sufficient for evaluating an ML component intended for numerical-model coupling. Their behavior must also be assessed within the simulation environment in which they operate, including convergence, stability, and feedback with evolving model states.
- Open Access
- 16 August 2026
- Published by Copernicus GmbH
- Vol. 2026, 1-39
Reply on RC1
Ice calving plays a significant role in ice sheet mass loss. Predictions of future ice sheet mass balance require accurate representations of the calving process in numerical ice models to determine rates of future global mean sea level rise. Whilst calving has recently begun to have been implemented in numerical models there has not been a systematic investigation into how this implementation compares between different ice flow models. The Calving Model Intercomparison Project (CalvingMIP) has been established to address this question by providing a framework of experiments to investigate the accuracy of simulated calving rates and how simulated properties at the calving front evolve over time. Our initial focus has been on how calving is implemented in models, therefore focusing on calving algorithms, rather than on how much ice should be calved at a particular time (calving law). Our results, from thirteen different numerical modelling groups, show that the majority of calving algorithms implemented are able to accurately implement a given calving rate with an ice front that evolves smoothly throughout the calving process. These results show that we can have confidence in the models capacity to accurately implement calving laws in the future.
- 16 August 2026
- Published by Copernicus GmbH
- Vol. 2026, 1-39
Open AccessReply on RC1
Ice calving plays a significant role in ice sheet mass loss. Predictions of future ice sheet mass balance require accurate representations of the calving process in numerical ice models to determine rates of future global mean sea level rise. Whilst calving has recently begun to have been implemented in numerical models there has not been a systematic investigation into how this implementation compares between different ice flow models. The Calving Model Intercomparison Project (CalvingMIP) has been established to address this question by providing a framework of experiments to investigate the accuracy of simulated calving rates and how simulated properties at the calving front evolve over time. Our initial focus has been on how calving is implemented in models, therefore focusing on calving algorithms, rather than on how much ice should be calved at a particular time (calving law). Our results, from thirteen different numerical modelling groups, show that the majority of calving algorithms implemented are able to accurately implement a given calving rate with an ice front that evolves smoothly throughout the calving process. These results show that we can have confidence in the models capacity to accurately implement calving laws in the future.
- Open Access
- 30 June 2026
- Published by American Geophysical Union (AGU) in Geophysical Research Letters
- Vol. 53 (13)
A Vast Valley Network Beneath the Greenland Ice Sheet
Mapping subglacial topography along the margin of the Greenland Ice Sheet was revolutionized by reconciling measurements of ice thickness and surface velocity using the principle of mass conservation. Despite evidence that many subglacial valleys resolved by that method extend upstream, it cannot be applied to the ice sheet's slower‐flowing interior, where ice thickness observations are sparse. Here we apply Ice Flow Perturbation Analysis to surface elevation and velocity observations, revealing a vast, island‐wide network of subglacial valleys connected to the peripheral valleys resolved by the mass conservation method. The morphology of this valley network further implicates Greenland's southern and eastern highlands as inception points for the ice sheet and also suggests that this network is modulated by both groundwater flow and Greenland's tectonic inheritance.
- 30 June 2026
- Published by American Geophysical Union (AGU) in Geophysical Research Letters
- Vol. 53 (13)
Open AccessA Vast Valley Network Beneath the Greenland Ice Sheet
Mapping subglacial topography along the margin of the Greenland Ice Sheet was revolutionized by reconciling measurements of ice thickness and surface velocity using the principle of mass conservation. Despite evidence that many subglacial valleys resolved by that method extend upstream, it cannot be applied to the ice sheet's slower‐flowing interior, where ice thickness observations are sparse. Here we apply Ice Flow Perturbation Analysis to surface elevation and velocity observations, revealing a vast, island‐wide network of subglacial valleys connected to the peripheral valleys resolved by the mass conservation method. The morphology of this valley network further implicates Greenland's southern and eastern highlands as inception points for the ice sheet and also suggests that this network is modulated by both groundwater flow and Greenland's tectonic inheritance.
- Open Access
- 19 June 2026
- Published by Copernicus GmbH in The Cryosphere
- Vol. 20 (6) , 3533-3558
Inferring subglacial topography using physics informed machine learning constrained by two conservation laws
Subglacial topography beneath the Greenland Ice Sheet is a fundamental control on its dynamics and response to changes in the climate system. Yet, it remains challenging to measure directly, and existing representations of the subglacial topography rely on a limited number of observations. Although the use of mass conservation and the development of BedMachine Greenland substantially improved the representation of the bed topography, this approach is limited to fast-flowing sectors and is less effective in regions with complex, alpine topography. As an alternative to traditional numerical methods, recent work has explored using Physics Informed Neural Networks (PINNs), constrained by only one physical law, to solve forward and inverse problems in ice sheet modeling. Building on this work, we assess three PINN frameworks constrained by distinct conservation laws, showing that PINNs informed with a single conservation law are not sufficient for regions with sparse measurements and complex topographies. To that end, we introduce a novel approach that involves coupling two conservation laws within a PINN framework to infer the subglacial topography and test this approach for three regions with distinct environments in Greenland. This PINN is trained with both the conservation of mass and an approximation of the conservation of momentum (the Shelfy-Stream Approximation), which allows us to simultaneously infer the ice thickness and basal shear stress using observations of ice velocities, surface elevation, surface mass balance, and ice thinning rates in a mixed inversion problem. We compare the predicted ice thickness to ground-truth ice-penetrating radar measurements of ice thickness, showing that the PINN informed with two conservation laws is capable of inferring ice thickness in sparsely surveyed regions. Furthermore, comparisons of predicted bed topographies with BedMachine Greenland show that this approach is capable of discovering new bed features in slower-moving regions and in regions of complex topography, highlighting its potential for better constraining the bed topography of the Greenland Ice Sheet.
- 19 June 2026
- Published by Copernicus GmbH in The Cryosphere
- Vol. 20 (6) , 3533-3558
Open AccessInferring subglacial topography using physics informed machine learning constrained by two conservation laws
Subglacial topography beneath the Greenland Ice Sheet is a fundamental control on its dynamics and response to changes in the climate system. Yet, it remains challenging to measure directly, and existing representations of the subglacial topography rely on a limited number of observations. Although the use of mass conservation and the development of BedMachine Greenland substantially improved the representation of the bed topography, this approach is limited to fast-flowing sectors and is less effective in regions with complex, alpine topography. As an alternative to traditional numerical methods, recent work has explored using Physics Informed Neural Networks (PINNs), constrained by only one physical law, to solve forward and inverse problems in ice sheet modeling. Building on this work, we assess three PINN frameworks constrained by distinct conservation laws, showing that PINNs informed with a single conservation law are not sufficient for regions with sparse measurements and complex topographies. To that end, we introduce a novel approach that involves coupling two conservation laws within a PINN framework to infer the subglacial topography and test this approach for three regions with distinct environments in Greenland. This PINN is trained with both the conservation of mass and an approximation of the conservation of momentum (the Shelfy-Stream Approximation), which allows us to simultaneously infer the ice thickness and basal shear stress using observations of ice velocities, surface elevation, surface mass balance, and ice thinning rates in a mixed inversion problem. We compare the predicted ice thickness to ground-truth ice-penetrating radar measurements of ice thickness, showing that the PINN informed with two conservation laws is capable of inferring ice thickness in sparsely surveyed regions. Furthermore, comparisons of predicted bed topographies with BedMachine Greenland show that this approach is capable of discovering new bed features in slower-moving regions and in regions of complex topography, highlighting its potential for better constraining the bed topography of the Greenland Ice Sheet.
- Open Access
- 22 May 2026
- Published by Copernicus GmbH in The Cryosphere
- Vol. 20 (5) , 2999-3024
Future Retreat of Great Aletsch Glacier and Hintereisferner – application of a full-Stokes model to two valley glaciers in the European Alps
We simulate the future evolution of two valley glaciers in the European Alps over the course of the 21st century. The model setup combines a numerical realization of full-Stokes ice dynamics coupled to a surface energy balance model forced with the sustained (inline with the Paris Agreement) and highest climate emission scenarios based on CMIP5 and CMIP6 data pools. The initialization of the three-dimensional glacier flow model is based on data assimilation, where a detailed observed ice surface velocity map serves as reference for constraining unknown parameters by means of inversion techniques. This setup is applied to Great Aletsch Glacier (GAG) and Hintereisferner (HEF) to assess their individual responses to climate change in the western and eastern European Alps, respectively. The model results of both glaciers are calibrated with comprehensive glaciological observations over several years to ensure a realistic glacier response in the observation period. The end-of-the-century projections reveal a substantial volume loss of both glaciers: HEF is projected to vanish in the middle of the 21st century regardless of the climate emission scenario. GAG is likely to disappear at the end of the 21st century under high-emission scenarios RCP 8.5 and SSP5-8.5, whereas low-emission scenarios RCP 2.6 and SSP1-2.6 predict a median glacier volume reduction of 67.7 % [62.2 % to 77.6 %] and 86.4 % [76.2 % to 89.4 %], respectively (values in brackets correspond to the 17th to 83rd percentile range). Our individual and detailed results of glacier evolution provide well-constrained estimates to complement large-scale modelling efforts. In general, our findings of substantial volume loss at the end of the 21st century align with large-scale modelling outcomes; however, a rough model-intercomparison study reveals a large spread of volume projections with the different glacier models.
- 22 May 2026
- Published by Copernicus GmbH in The Cryosphere
- Vol. 20 (5) , 2999-3024
Open AccessFuture Retreat of Great Aletsch Glacier and Hintereisferner – application of a full-Stokes model to two valley glaciers in the European Alps
We simulate the future evolution of two valley glaciers in the European Alps over the course of the 21st century. The model setup combines a numerical realization of full-Stokes ice dynamics coupled to a surface energy balance model forced with the sustained (inline with the Paris Agreement) and highest climate emission scenarios based on CMIP5 and CMIP6 data pools. The initialization of the three-dimensional glacier flow model is based on data assimilation, where a detailed observed ice surface velocity map serves as reference for constraining unknown parameters by means of inversion techniques. This setup is applied to Great Aletsch Glacier (GAG) and Hintereisferner (HEF) to assess their individual responses to climate change in the western and eastern European Alps, respectively. The model results of both glaciers are calibrated with comprehensive glaciological observations over several years to ensure a realistic glacier response in the observation period. The end-of-the-century projections reveal a substantial volume loss of both glaciers: HEF is projected to vanish in the middle of the 21st century regardless of the climate emission scenario. GAG is likely to disappear at the end of the 21st century under high-emission scenarios RCP 8.5 and SSP5-8.5, whereas low-emission scenarios RCP 2.6 and SSP1-2.6 predict a median glacier volume reduction of 67.7 % [62.2 % to 77.6 %] and 86.4 % [76.2 % to 89.4 %], respectively (values in brackets correspond to the 17th to 83rd percentile range). Our individual and detailed results of glacier evolution provide well-constrained estimates to complement large-scale modelling efforts. In general, our findings of substantial volume loss at the end of the 21st century align with large-scale modelling outcomes; however, a rough model-intercomparison study reveals a large spread of volume projections with the different glacier models.
- Open Access
- 18 May 2026
- Published by American Geophysical Union (AGU) in Journal of Advances in Modeling Earth Systems
- Vol. 18 (5)
DJ4Earth: Differentiable, and Performance‐Portable Earth System Modeling via Program Transformations
Differentiable Earth system models (ESMs) enable powerful applications such as sensitivity analysis, gradient‐based calibration, state estimation, boundary flux inversions, uncertainty quantification, and online machine learning. Reverse‐mode automatic differentiation (AD) efficiently provides gradients for such tasks, yet models have rarely included this capability because of complex, bespoke numerical algorithms. As part of the Differentiable programming in Julia for Earth system modeling (DJ4Earth) initiative, we present improved capabilities of the AD tool Enzyme.jl and the new compiler transpilation tool Reactant.jl, augmented by sophisticated checkpointing algorithms, which, together make general‐purpose AD tractable and efficient for full‐fledged ESM components written in Julia. Operating at the low‐level virtual machine intermediate representation or multi‐level intermediate representation compiler levels, these frameworks support mutable memory, custom kernels, and compiler optimizations before and after differentiation. Julia‐specific challenges related to just‐in‐time compilation and garbage collection are handled efficiently. Reactant further enables automatic performance portability across central processing units, graphics processing units, and tensor processing units, facilitating use of emerging AI‐customized high‐performance computing architectures. We demonstrate these frameworks on four Julia‐based ESM components featuring diverse spatial discretizations and numerical algorithms: the rotating‐sphere shallow water model ShallowWaters.jl, the finite‐volume ocean model Oceananigans.jl, the finite‐element ice sheet model DJUICE.jl, and the spectral atmospheric model SpeedyWeather.jl. Across these ESM components, our tools compute efficient and correct gradients. These results establish a foundation for differentiable, high‐performance and performance‐portable ESMs that can integrate neural networks for unresolved processes, trained online, enabling next‐generation hybrid physics–machine learning ESMs constrained by physical dynamics and observations.
- 18 May 2026
- Published by American Geophysical Union (AGU) in Journal of Advances in Modeling Earth Systems
- Vol. 18 (5)
Open AccessDJ4Earth: Differentiable, and Performance‐Portable Earth System Modeling via Program Transformations
Differentiable Earth system models (ESMs) enable powerful applications such as sensitivity analysis, gradient‐based calibration, state estimation, boundary flux inversions, uncertainty quantification, and online machine learning. Reverse‐mode automatic differentiation (AD) efficiently provides gradients for such tasks, yet models have rarely included this capability because of complex, bespoke numerical algorithms. As part of the Differentiable programming in Julia for Earth system modeling (DJ4Earth) initiative, we present improved capabilities of the AD tool Enzyme.jl and the new compiler transpilation tool Reactant.jl, augmented by sophisticated checkpointing algorithms, which, together make general‐purpose AD tractable and efficient for full‐fledged ESM components written in Julia. Operating at the low‐level virtual machine intermediate representation or multi‐level intermediate representation compiler levels, these frameworks support mutable memory, custom kernels, and compiler optimizations before and after differentiation. Julia‐specific challenges related to just‐in‐time compilation and garbage collection are handled efficiently. Reactant further enables automatic performance portability across central processing units, graphics processing units, and tensor processing units, facilitating use of emerging AI‐customized high‐performance computing architectures. We demonstrate these frameworks on four Julia‐based ESM components featuring diverse spatial discretizations and numerical algorithms: the rotating‐sphere shallow water model ShallowWaters.jl, the finite‐volume ocean model Oceananigans.jl, the finite‐element ice sheet model DJUICE.jl, and the spectral atmospheric model SpeedyWeather.jl. Across these ESM components, our tools compute efficient and correct gradients. These results establish a foundation for differentiable, high‐performance and performance‐portable ESMs that can integrate neural networks for unresolved processes, trained online, enabling next‐generation hybrid physics–machine learning ESMs constrained by physical dynamics and observations.