CondensNet: Enabling stable long-term climate simulations via hybrid deep learning models with adaptive physical constraints

TL;DR

This work tackles the instability of hybrid DL-GCM climate models by identifying water-vapor oversaturation during condensation as a key failure mode and introducing CondensNet, a physically constrained neural architecture. CondensNet combines BasicNet, which learns cloud-related tendencies, with ConCorrNet, which adaptively enforces saturation-adjustment constraints when relative humidity exceeds 100%, yielding the PCNN-GCM framework that couples CondensNet to CAM5.2 under AMIP forcing. The results show that PCNN-GCM achieves long-term stability comparable to SPCAM while delivering substantial computational speedups (roughly 100×–372×) over SPCAM, and provides improved accuracy for water-cycle and related thermodynamic variables relative to CAM5 and prior DL-GCMs. This approach advances hybrid climate modeling by demonstrating an effective, interpretable method to impose physical constraints in DL parametrizations, paving the way for lightweight, stable, and scalable long-term climate simulations across land and ocean domains.

Abstract

Accurate and efficient climate simulations are crucial for understanding Earth's evolving climate. However, current general circulation models (GCMs) face challenges in capturing unresolved physical processes, such as cloud and convection. A common solution is to adopt cloud resolving models, that provide more accurate results than the standard subgrid parametrisation schemes typically used in GCMs. However, cloud resolving models, also referred to as super paramtetrizations, remain computationally prohibitive. Hybrid modeling, which integrates deep learning with equation-based GCMs, offers a promising alternative but often struggles with long-term stability and accuracy issues. In this work, we find that water vapor oversaturation during condensation is a key factor compromising the stability of hybrid models. To address this, we introduce CondensNet, a novel neural network architecture that embeds a self-adaptive physical constraint to correct unphysical condensation processes. CondensNet effectively mitigates water vapor oversaturation, enhancing simulation stability while maintaining accuracy and improving computational efficiency compared to super parameterization schemes. We integrate CondensNet into a GCM to form PCNN-GCM (Physics-Constrained Neural Network GCM), a hybrid deep learning framework designed for long-term stable climate simulations in real-world conditions, including ocean and land. PCNN-GCM represents a significant milestone in hybrid climate modeling, as it shows a novel way to incorporate physical constraints adaptively, paving the way for accurate, lightweight, and stable long-term climate simulations.

Figures (11)

• Figure 1: Methodology of the CondensNet model. CondensNet is a physically-constrained DL parametrization coupled with a climate dynamics engine to support hybrid modeling. The network architecture mainly has two parts: BasicNet for learning the cloud representation and ConCorrNet for condensation physical constraint.
• Figure 2: Total energy (a) and total precipitable water (d) time evolution for different models, including stable (light green), biased (yellow), drifted (orange), and crashed (dark red) ResMLP DL parametrizations, part of NN-GCM, as well as SPCAM (dark green). Total energy (b) and total precipitable water (e) time evolution of Unstable ResMLP models, part of NN-GCM. Total energy (c) and total precipitable water (f) time evolution of stable CondensNet models, part of PCNN-GCM, using as baseline the same configuration of the unstable ResMLP models, to show the effects of CondensNet stabilization properties. Relative humidity for SPCAM reference (g), NN-GCM model failing after 5000 time steps (h), stable NN-GCM, and new PCNN-GCM featuring CondensNet (j).
• Figure 3: Precipitation (a--d) and vertical profile of specific humidity (h--k), for SPCAM, CAM5, NN-GCM, and PCNN-GCM, respectively, and corresponding differences with respect to SPCAM reference (e--g, for precipitation, and m--o, for specific humidity). The fields are annual means (1998-2002) computed as reported in Methods \ref{['sec:means']}, and their differences are computed as in Equation \ref{['eq:pattern-diff']}, reported in Methods \ref{['sec:errors']}. We also provide error metrics, namely RMSE, for all subfigures related to differences (i.e., e--g for precipitation and m--o for specific humidity).
• Figure 4: Subfigure (a) shows the execution time (ET) in seconds for one simulation time step across different models, namely SPCAM, CAM5, PCNN-GCM, and for different numbers of MPI processes. Subfigure (b) provides a detailed view of ET and of the simulated years per day (SYPD) for each model for different MPI processes. DL parametrization, CondensNet, inference can be run on both CPU or GPU; we report both results.
• Figure 5: The PCNN-GCM framework. Panels(a) and(b) show the conventional subgrid parametrization and super parametrization approaches, respectively; panel(c) highlights the DL parametrization concept, and panel(d) details the internal architecture of CondensNet; panel(e) is the host GCM. The host GCM plus CondensNet form the PCNN-GCM framework.