Energy Exascale Earth System Model (E3SM)

Updated 6 December 2025

E3SM is an integrated, exascale-ready modeling suite that couples atmosphere, ocean, land, and sea ice components to advance climate and energy-relevant research.
It employs scalable HPC techniques, high-resolution grids, and advanced diagnostic tools to capture extreme events and complex system feedbacks.
Hybrid ML–physics approaches and surrogate-based calibration methods enable efficient parameter tuning and robust, high-fidelity Earth system simulations.

The Energy Exascale Earth System Model (E3SM) is an exascale-ready, fully coupled Earth system modeling suite spearheaded by the U.S. Department of Energy. E3SM is designed to simulate atmosphere, ocean, land, sea ice, and associated biogeochemical components at resolutions and with subgrid process representation sufficient to advance understanding of energy-relevant climate phenomena, extreme events, and associated feedbacks. Emphasizing scalable HPC implementations, modularity, and extensible diagnostics, E3SM has underpinned noteworthy advances in both physical simulation and workflow science, enabling both traditional physics-based and novel AI/ML-accelerated paradigms.

1. Model Structure, Components, and Time Integration

E3SM comprises atmosphere (EAM), land (ELM), ocean, sea-ice, and river-routing models, linked via a high-throughput coupler. The atmospheric model (EAMv1/v2) integrates mass, momentum, energy, and tracer equations on a finite-volume or spectral element grid with typical resolutions ranging from 1° to sub-degree, up to 72 vertical levels extending to 0.1 hPa (Wan et al., 2021). Physics parameterizations include CLUBB (for turbulence and shallow convection), Zhang–McFarlane deep convection, MG2 cloud microphysics, radiative transfer, and aerosol/chemistry schemes. Time integration uses isolated sequential splitting, each physics process updating the prognostic state in turn. All model components share hierarchical Fortran-derived data structures, history output logic, and checkpointing/concurrency mechanisms, supporting robust coupling and diagnostic extensibility (Wan et al., 2021).

2. High-Resolution Land Modeling at Continental Scale

E3SM’s land model (ELM) has been demonstrably ported to kilometer-scale domains, enabling physically comprehensive, data-driven, and fully-coupled simulations over continental extents (Wang et al., 19 Jan 2025). The km-scale ELM features:

Input data: Daymet 1 km meteorological forcing, dynamically downscaled and mosaicked with global 0.5° land/soil/vegetation inputs, and processed for domain-specific ingestion using the KiloCraft toolkit.
Physics: All “core-capability” ELM physics, including multilayer heat/hydrology, CNP cycling, Farquhar photosynthesis, and subgrid land unit aggregation, operate at per-gridcell independence—no lateral diffusion.
Infrastructure: Parallelization is via static MPI domain decomposition, exploiting over 100,000 CPU cores (IBM POWER9) with PnetCDF/NetCDF/ADIOS for highly efficient, distributed I/O.
Performance: Strong scaling yields 58% efficiency at 100,800 cores, and weak scaling remains ~80–100% over 21.6 million gridcells; I/O throughput reaches 8.4 GB/s. Hemispheric-scale output generation, including ~4,540 GB domain writes, is performed without workflow bottlenecks.

These experiments establish that exascale resources, combined with modular preprocessing, enable new classes of high-fidelity, event-resolving simulations, including topographically and land-use sensitive extreme event modeling (Wang et al., 19 Jan 2025).

3. Diagnostics, Pathway Analysis, and In-situ Extraction

The need for detailed process-level diagnostics, conditional budget analyses, and efficient pathway quantification has driven significant tool co-development with E3SM:

CondiDiag1.0 integrates into EAM to provide model-agnostic, generic diagnostics for conditional sampling, process-level increment archiving, and vertical integration, all via lightweight Fortran modules and run-time namelist configuration. Arbitrary conditional metrics (e.g., RHI>125%, U10 thresholds) can be specified for both field and increment output at any physics checkpoint, with minimal code changes and I/O burden (Wan et al., 2021).
CLDERA-Tools augments E3SM via minimally invasive Fortran hooks, enabling registration of arbitrary quantities of interest and automated, numerically stable online statistics (mean, variance, covariance, Granger causality) via Kokkos/C reductions. The toolkit constructs time-dependent directed acyclic graphs (DAGs) of variable relationships (e.g., pollutant → aerosol → temperature) using sliding-window Pearson correlations or causal metrics, drastically reducing raw I/O requirements by 90–99% and overhead to <5%, even at 4,096 ranks (Steyer et al., 7 Aug 2024).

These tools underpin new approaches to “pathway analysis”—mapping the chain of state interactions in response to external forcing (e.g., volcanic eruptions), fully in-situ and at exascale, circumventing postprocessing and traditional data movement bottlenecks.

4. Parameter Estimation and Auto-Calibration

Model calibration is a central challenge in high-dimensional ESMs. E3SM has pioneered end-to-end auto-calibration workflows based on surrogate modeling and gradient-based optimization (Yarger et al., 2023):

Principal component analysis (PCA) reduces the dimensionality of stacked field outputs (≈47,400 in the paper) to 16 leading empirical orthonormal functions (EOFs), capturing ≈87% of variance.
Surrogate modeling leverages polynomial chaos expansions (PCE) on the PCA-reduced basis, trained on 250-ensemble “perturbed parameter” runs. Elastic-net penalization ensures tractable, cross-validated surrogates.
MAP optimization—incorporating field-specific inverse-Gamma priors—yields parameter choices substantially different from hand-tuned defaults, with up to 10% RMSE improvements in key fields and an average reduction of 2.7% across 45 outputs.
The approach requires less than an hour of ensemble-ML compute, reducing weeks-long calibration to a reproducible, scalable, and uncertainty-quantified pipeline.

This workflow demonstrates that gradient-enabled surrogates and modern Bayesian inference can systematically outperform manual parameter tuning in high-dimensional model spaces (Yarger et al., 2023).

5. Hybrid ML–Physics Modeling and Surrogates

E3SM serves as an active testbed for ML-augmented physical modeling at both process and system scales:

PHASE (“Physics-Integrated, Heterogeneity-Aware Surrogates”) demonstrates knowledge-guided grouping, data-type–aware encoders (LSTM, CNN, FC), cross-group transformer fusion, and task-specific prediction heads, with soft/hard physical constraints. Applied to ELM BGC spin-up (normally 1,200 simulated years), PHASE achieves over 60× acceleration, generating physically consistent, numerically stable restart files for equilibrium carbon pools. Generalization to higher resolution (0.5° grids) is realized through rapid few-shot adaptation (5–10% samples), confirming the capture of resolution-invariant regularity rather than gridpoint memorization (Gao et al., 27 Sep 2025).
Bias correction in E3SM’s atmosphere model leverages operator-learning surrogates: U-Net, Inception U-Net (IUNet), and M_M (multiscale upsampling) architectures trained on nudged ERA5 data. Cadence-limited corrections—applied every 3 h—enable direct ML operator injection into the EAM tendency budget. M_M yields the most consistent multi-variable, multi-level bias reductions; all ML-augmented runs remain stable and computationally practical on multi-year horizons. Overheads are manageable, especially compared to the cost of brute-force high-resolution simulations (Bora et al., 2 Dec 2025).
DeepONet architectures (with convolutional autoencoder-decoders) have been demonstrated to replace relaxation-based nudging in E3SMv2. This replacement yields 10× faster module evaluation, eliminates the need for external data I/O, and achieves comparable or improved fidelity in high-dimensional bias correction tasks (Bora et al., 2023).

These efforts highlight that hybrid approaches, whether via physics-constrained surrogates or operator-based real-time bias correction, can deliver exascale performance and stability while enabling entirely new adaptive, uncertainty-quantified, or event-driven workflows.

6. Evaluation, Coupling, and Configuration-Specific Capabilities

Comparative assessments of E3SM configurations demonstrate that mean state fidelity and parameterization tuning remain critical for process coupling and mode fidelity:

Evaluation of tropical Atlantic–Pacific multi-decadal teleconnections across 27 CMIP6 models and two E3SM configurations (standard E3SMv2 and E3SM-MMF/super-parameterized) shows that only the standard, carefully tuned E3SMv2 accurately reproduces observed bidirectional teleconnection pathways. Multi-scale frameworks with explicit embedded CRMs do not, by themselves, overcome biases—such as reversed Walker circulation or SST patterns—unless mean-state errors and cloud/radiation parameterizations are also explicitly addressed (Xia et al., 4 Apr 2025).
Statistical skill metrics (Total Performance Metric, TPM) combining pattern correlations in tropical and extratropical domains allow robust, domain-specific diagnosis of configuration weaknesses. Recommendations include targeted cloud parameter retuning, CRM feedback refinement, and improved oceanic wave speeds to address coupled-basin teleconnection fidelity (Xia et al., 4 Apr 2025).
The CIME framework enables seamless configuration of E3SM model components, allowing rapid switching between fully coupled (atm-land-ocean-ice) and offline or data-driven workflows. This supports robust scaling tests, flexible ensemble design, and cross-component diagnostic integration (Wang et al., 19 Jan 2025).

A plausible implication is that the continual interplay between model tuning, diagnostic innovation, and computational scalability is essential for achieving both process representation and predictive accuracy in exascale ESM applications.

7. Exascale Performance, Portability, and Future Directions

E3SM’s modular integration, focus on data-movement minimization, and extensible ML/AI interfaces make it a candidate platform for exascale-fluid, hybrid physical–statistical Earth system science:

In-situ statistics and pathway analysis (e.g., with CLDERA-Tools) show <5% overhead while saving multiple terabytes in daily I/O (Steyer et al., 7 Aug 2024).
GPU and future architecture readiness is supported via Kokkos, OpenACC, and pytorch-fortran bridges, ensuring physics and ML components can co-reside and scale on emerging hardware.
ML operator modularity (TorchScript, external ABI) and cadence-limited updates restrict computational footprint, support parameterization integrity, and anticipate GPU-native exascale ESMs.
Surrogate-enabled ensemble prediction, parameter sensitivity, and uncertainty quantification become tractable for thousands of runs.
Model-agnostic diagnostic frameworks (CondiDiag1.0, CLDERA-Tools) ensure portability across major ESMs evolving from the spectral element/CESM lineages.

Overall, E3SM embodies a comprehensive, extensible, and computationally viable architecture for high-resolution, physically and statistically robust Earth system simulation at the exascale frontier.