Assessing Emulator Design and Training for Modal Aerosol Microphysics Parameterizations in E3SMv2
Shady E. Ahmed, Hui Wan, Saad Qadeer, Panos Stinis, Kezhen Chong, Mohammad Taufiq Hassan Mozumder, Kai Zhang, Ann S. Almgren · 2026
Read Assessing Emulator Design and Training for Modal Aerosol Microphysics Parameterizations in E3SMv2 by Shady E. Ahmed, Hui Wan, Saad Qadeer, Panos Stinis, Kezhen Chong, Mohammad Taufiq Hassan Mozumder, Kai Zhang, Ann S. Almgren (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2604-21233.txt.
Introduction
Shady Ahmed and seven colleagues test how design and training choices affect neural stand-ins for aerosol microphysics in the E3SMv2 climate model. Focusing on cloud-free, four-mode aerosol processes, they vary network size and variable scaling while watching training closely. Tuning, scaling, and network size strongly drive accuracy, but a modest, well-scaled plain network reproduces key concentration changes with promising accuracy, offering guidance for other sky processes.
Catalog record
| CREATOR | Shady E. Ahmed, Hui Wan, Saad Qadeer, Panos Stinis, Kezhen Chong, Mohammad Taufiq Hassan Mozumder, Kai Zhang, Ann S. Almgren |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 39836 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2604.21233
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Assessing Emulator Design and Training for Modal Aerosol Microphysics Parameterizations in E3SMv2 Shady E. Ahmed, Hui Wan, Saad Qadeer, Panos Stinis, Kezhen Chong, Mohammad Taufiq Hassan Mozumder, Kai Zhang, Ann S. Almgren Abstract Toward the goal of using Scientific Machine Learning (SciML) emulators to improve the numerical representation of aerosol processes in global atmospheric models, we explore the emulation of aerosol microphysics processes under cloud-free conditions in the 4-mode Modal Aerosol Module (MAM4) within the Energy Exascale Earth System Model version 2 (E3SMv2). To develop an in-depth understanding of the challenges and opportunities in applying SciML to aerosol processes, we begin with a simple feedforward neural network architecture that has been used in earlier studies, but we systematically examine key emulator design choices, including architecture complexity and variable normalization, while closely monitoring training convergence behavior. Our results show that optimization convergence, scaling strategy, and network complexity strongly influence emulation accuracy. When effective scaling is applied and convergence is achieved, the relatively simple architecture, used together with a moderate network size, can reproduce key features of the microphysics-induced aerosol concentration changes with promising accuracy. These findings provide practical clues for the next stages of emulator development; they also provide general insights that are likely applicable to the emulation of other aerosol processes, as well as other atmospheric physics involving multi-scale variability. Keywords Aerosol microphysics \cdot Feedforward neural network \cdot modal aerosol module \cdot E3SM 1 Introduction With the rapid advancement and growing adoption of artificial intelligence (AI) and Scientific Machine Learning (SciML), the Earth system science community has increasingly sought to incorporate these methods into modeling, prediction, and analysis workflows. Recent progress (see, e.g., summary in [ 1 ] ) has included emulation of global atmospheric and ocean models [ 2 , 3 , 4 , 5 ] , development of AI- and SciML-based cloud and turbulence parameterizations [ 6 , 7 , 8 , 9 , 10 ] , and the release of curated community datasets [ 11 , 12 ] . Within this broader movement, the Energy Exascale Earth System Model (E3SM) community has identified several promising avenues for integrating AI and SciML in Earth system model (ESM) development and application, including full-model emulation, process emulation, parameter estimation (calibration), workflow automation, and model initialization or spin-up [ 13 ] . Some efforts have been directed toward applying AI and SciML to improve the representation of aerosol life cycles and their interactions with clouds and radiation [ 14 , 15 , 16 , 17 ] , while many avenues remain to be explored. In the realm of traditional numerical models based on partial differential equations (PDEs) and ordinary differential equations (ODEs), various global atmospheric models and chemical transport models have incorporated parameterizations of aerosol microphysics to simulate changes in the size and composition of aerosol particles caused by processes such as nucleation, condensation–evaporation, coagulation, and aging [ 18 , 19 , 20 , 21 ] . Due to the complexity of these processes as well as the sharp contrast between their characteristic scales and the typical grid spacing in global models, it has been challenging to develop computational methods that are both accurate and computationally efficient. The modal method, which uses assumed classes of functions to represent the statistical distribution of particle size and composition (see, e.g., [ 22 , 23 , 24 ] ), offers a good trade-off between accuracy and cost, although current implementations still contain ample room for improvement [ 25 , 26 ] . ML-trained emulators are worth exploring for the purpose of addressing this challenge by providing fast and accurate emulators of the underlying physics. Several recent studies have taken initial steps toward this goal. Harder et al. 2021 introduced one of the first ML emulators for aerosol microphysics, targeting the M7 parameterization suite [ 23 ] in the ECHAM-HAM model [ 28 , 20 ] , where they also explored the role of physical constraints [ 29 ] . More recently, Bai and Rouson 2025 developed a deep operator network for emulating the aerosol microphysics parameterization suite in E3SM version 2 (E3SMv2). Collectively, these studies reported a notably wide range for the coefficient of determination ( R^{2} )—from roughly 0.48 to 0.99 [ 27 , 29 , 30 ] —revealing substantial variability in emulator performance. So far, the published studies lack a systematic analysis of why such variability arises and how specific architectural choices, training strategies, or properties of the problem being emulated contribute to these differences. In other words, the community has produced promising results but does not yet have a clear baseline or design-space understanding for aerosol microphysics emulation. Such a baseline and understanding will provide a necessary foundation for future work, especially since the physical processes of interest present unique challenges for ML emulation, including, e.g., the wide dynamic ranges in aerosol properties and the physical constraints that must be preserved. This study aims at addressing this gap in understanding by performing a controlled, detailed investigation of the design decisions that affect the accuracy and reliability of aerosol microphysics emulators. To disentangle emulator-dependent effects from complexities resulting from different modes of variability or boundary conditions in the Earth system, we start by restricting our analysis to one representative condition, namely boreal winter, and we defer it to future studies to investigate the topic of generalizability with respect to external factors such as seasonality, diurnal cycle, and aerosol source scenarios. This enables a clean evaluation of how model architecture, input/output configuration, and training setup affect emulator performance. In the first step, we focus on assessing what a standard feedforward neural network (FNN) can achieve in this controlled setting, so as to establish a clear, easy-to-obtain baseline for future developments. We aim at emulating the clear-air (i.e., cloud-free) aerosol microphysics suite in E3SMv2 exercised in the default 4-mode configuration of the aerosol module, in simulations conducted at the current workhorse resolution of about 165 km horizontal grid spacing, so that our findings can directly inform E3SM-focused AI/ML development. Nevertheless, various strategies presented in this paper, including the assessment of training convergence and the use of variable transformations to handle physical quantities spanning wide ranges of scales, are generally applicable to the emulation of other parameterizations and PDE/ODE models. While potential long-term plans include exploring AI/ML emulators to improve upon the current PDE-based representation of aerosol life cycles in E3SM by emulating, for example, measurements data and simulation output from much more detailed, physics-based, process-resolved numerical models, we begin by emulating the aerosol microphysics parameterization suite currently used by E3SMv2 so as to focus on emulator accuracy. This framing allows us to treat the problem as a controlled benchmark: the underlying parameterization is fixed, and the primary question becomes how accurately and robustly an ML emulator can reproduce its behavior. Emulating a parameterization suite currently in use also has immediate practical value, as microphysics can be a computationally expensive component of atmospheric aerosol models, and this was a key factor that motivated some of the earlier studies, e.g., [ 27 , 29 ] . The remainder of the manuscript is organized as follows. Section 2 introduces the aerosol microphysics module in E3SMv2, the input-output mapping to be emulated, and the neural network architecture used