Cloud Processing Research Articles

Investigating the sign of stratocumulus adjustments to aerosols in the ICON global storm-resolving model

Abstract. Aerosols can cause brightening of stratocumulus clouds, thereby cooling the climate. Observations and models disagree on the magnitude of this cooling, partly because of the aerosol-induced liquid water path (LWP) adjustment, with climate models predicting an increase in the LWP and satellites observing a weak decrease in response to increasing aerosols. With higher-resolution global climate models, which allow the simulation of mesoscale circulations in which stratocumulus clouds are embedded, there is hope to start bridging this gap. In this study, we present boreal summertime simulations conducted with the ICOsahedral Non-hydrostatic (ICON) global storm-resolving model (GSRM). Compared to geostationary satellite data, ICON produces realistic cloud coverage in the stratocumulus regions; however, these clouds look cumuliform, and the sign of LWP adjustments disagrees with observations. We investigate this disagreement with a causal approach, which combines time series with knowledge of cloud processes, allowing us to diagnose the sources of observation–model discrepancies. The positive ICON LWP adjustment results from a superposition of processes, with an overestimated positive response due to (1) precipitation suppression, (2) a lack of wet scavenging, and (3) cloud deepening under a weak inversion, despite (4) small negative influences from cloud-top entrainment enhancement. We also find that precipitation suppression and entrainment enhancement occur at different intensities during the day and the night, implying that daytime satellite studies suffer from selection bias. This causal methodology can guide modelers on how to modify model parameterizations and setups to reconcile conflicting studies concerning the sign and magnitude of LWP adjustments across different data sources.

Improved keypoint localization network for tea bud based on YOLO framework

The precise localization of picking points is the primary challenge that should be addressed by the robotic tea bud harvester. The efficient identification and accurate positioning of picking points directly affects the efficiency of tea buds harvesting. Compared with the target detection and semantic segmentation algorithms, the keypoint detection algorithms do not require to incorporate target shape features for accurately localizing the picking point, which allows to significantly decrease the computational load, and thus facilitates the subsequent point cloud processing. This study introduces a tea bud keypoint detection network referred to as TBKNet. This model uses the concept of structural reparameterization to reconstruct the CSPDarknet53 backbone network, and constructs an asymptotic parallel feature fusion network by incorporating the CA attention mechanism into the asymptotic feature pyramid network. In addition, in order to accelerate the convergence of the model, a keypoint regression loss function, which considers the angle and Euclidean distance, is proposed. The experimental results show that TBNKet reaches the highest accuracy with an average precision of 87.1 % for keypoint detection. It has 5.7 %, 16.9 %, and 3.3 % higher mAP values compared with the DEKR, YOLOPose, and YOLOPv8-Pose models, respectively. This is efficient in identifying and localizing keypoints of tea buds, and it can provide guidance for similar intensive harvesting tasks.

Green Plant Landscape Design for Urban Air Quality Purification with Computer Image Processing in Cloud, Grid, and Cluster Computing

Green Plant Landscape Design for Urban Air Quality Purification with Computer Image Processing in Cloud, Grid, and Cluster Computing

Integrating vision and lidar based hyperlocal metadata for optimal capacity expansion planning in hillside road networks

The rapidly expanding population and scarcity of affordable housing in many urban areas, such as Los Angeles, have resulted in large inventories of structures in hilly landscapes. Such urbanization has also rendered road capacities inadequate in many instances, with hazards such as floods, fires, and earthquakes posing additional risks to such inventory. Planning construction activities such as widening, structural, and geotechnical retrofits to increase capacity and resiliency by cutting through slopes and incorporating landslide protection measures requires hyperlocal information for road segments in hillside areas. Performing traditional surveys for large areas of such road networks can be both time-consuming and cost prohibitive. At the same time, data acquired from satellites may not provide the required spatial resolution for planning, especially regarding road-adjacent features such as slopes. This paper proposes a novel framework to fuse hyperlocal data regarding road and adjacent slopes gathered from the street-level mobile mapper equipped with cameras and lidar into the network design problem (NDP) to develop an optimal road width expansion plan. The hyperlocal features are derived from a colorized point cloud map and deep learning techniques to estimate the relevant features in hillside landscapes. Specifically, a deep learning-based segmentation and detection algorithms combined with novel point cloud processing techniques are used to obtain a spatially dense profile of the road and the adjacent slope. By leveraging such hyperlocal information, we explore all the potential road capacity and improvement options for road capacity expansion. Furthermore, we demonstrate the advantages of our proposed method over conventional techniques, particularly working with budget constraints. In terms of novel contributions, incorporating hyperlocal information into NDP for hillside road capacity expansion is novel, as well as algorithms to automatically detect and characterize relevant roadside features from remotely sensed lidar point clouds.

Measurement report: Impact of cloud processes on secondary organic aerosols at a forested mountain site in southeastern China

Measurement report: Impact of cloud processes on secondary organic aerosols at a forested mountain site in southeastern China

Water Conservancy Hub Geometric Modeling Based on Point Clouds

For hydraulic engineering, accurate geometric modeling is crucial for design, analysis, and maintenance. In recent years, point cloud technology, with its high precision and detailed data capture capabilities, has gradually become the preferred method for geometric modeling in hydraulic engineering. However, traditional point cloud processing methods exhibit slow data processing speeds and poor mesh quality when reconstructing 3D models. Therefore, we have studied the meshing method for the 3D reconstruction of large-scale discrete point cloud data and proposed a meshing method based on the Delaunay criterion and the growth method to achieve rapid meshing of discrete point cloud data. To demonstrate the efficiency and convenience of the proposed method, point cloud reconstruction experiments are conducted on various buildings. We compare our approach with the Poisson, BPA, and ONet methods. Additionally, we apply it to the Jiangsu Gaogang Water Conservancy Hub in China. The results show that our proposed method is superior in both reconstruction quality and efficiency, providing a promising solution for the reconstruction of 3D building models from point clouds.

On flange-based 3D hand–eye calibration for soft robotic tactile welding

This paper investigates the direct application of standardized designs on the robot for conducting robot hand–eye calibration by employing 3D scanners with collaborative robots. The well-established geometric features of the robot flange are exploited by directly capturing its point cloud data. In particular, an iterative method is proposed to facilitate point cloud processing towards a refined calibration outcome. Several extensive experiments are conducted over a range of collaborative robots, including Universal Robots UR5 & UR10 e-series, Franka Emika, and AUBO i5 using an industrial-grade 3D scanner Photoneo Phoxi S & M and a commercial-grade 3D scanner Microsoft Azure Kinect DK. Experimental results show that translational and rotational errors converge efficiently to less than 0.28 mm and 0.25 degrees, respectively, achieving a hand–eye calibration accuracy as high as the camera’s resolution, probing the hardware limit. A welding seam tracking system is presented, combining the flange-based calibration method with soft tactile sensing. The experiment results show that the system enables the robot to adjust its motion in real-time, ensuring consistent weld quality and paving the way for more efficient and adaptable manufacturing processes.

Improved 3D Object Detection Based on PointPillars

Despite the recent advancements in 3D object detection, the conventional 3D point cloud object detection algorithms have been found to exhibit limited accuracy for the detection of small objects. To address the challenge of poor detection of small-scale objects, this paper adopts the PointPillars algorithm as the baseline model and proposes a two-stage 3D target detection approach. As a cutting-edge solution, point cloud processing is performed using Transformer models. Additionally, a redefined attention mechanism is introduced to further enhance the detection capabilities of the algorithm. In the first stage, the algorithm uses PointPillars as the baseline model. The central concept of this algorithm is to transform the point cloud space into equal-sized columns. During the feature extraction stage, when the features from all cylinders are transformed into pseudo-images, the proposed algorithm incorporates attention mechanisms adapted from the Squeeze-and-Excitation (SE) method to emphasize and suppress feature information. Furthermore, the 2D convolution of the traditional backbone network is replaced by dynamic convolution. Concurrently, the addition of the attention mechanism further improves the feature representation ability of the network. In the second phase, the candidate frames generated in the first phase are refined using a Transformer-based approach. The proposed algorithm applies channel weighting in the decoder to enhance channel information, leading to improved detection accuracy and reduced false detections. The encoder constructs the initial point features from the candidate frames for encoding. Meanwhile, the decoder applies channel weighting to enhance the channel information, thereby improving the detection accuracy and reducing false detections. In the KITTI dataset, the experimental results verify the effectiveness of this method in small objects detection. Experimental results show that the proposed method significantly improves the detection capability of small objects compared with the baseline PointPillars. In concrete terms, in the moderate difficulty detection category, cars, pedestrians, and cyclists average precision (AP) values increased by 5.30%, 8.1%, and 10.6%, respectively. Moreover, the proposed method surpasses existing mainstream approaches in the cyclist category.

The Chalmers Cloud Ice Climatology: retrieval implementation and validation

Abstract. Ice clouds are a crucial component of the Earth's weather system, and their representation remains a principal challenge for current weather and climate models. Several past and future satellite missions were explicitly designed to provide observations offering new insights into cloud processes, but these specialized cloud sensors are limited in their spatial and temporal coverage. Geostationary satellites have been observing clouds for several decades and can ideally complement the sparse measurements from specialized cloud sensors. However, the geostationary observations that are continuously and globally available over the full observation record are restricted to a small number of wavelengths, which limits the information they can provide on clouds. The Chalmers Cloud Ice Climatology (CCIC) is a novel cloud-property dataset that aims to provide an improved climate record of ice hydrometeor concentrations by applying state-of-the-art machine-learning techniques to retrieve ice cloud properties from globally gridded, single-channel geostationary observations that are readily available from 1980 onwards. CCIC offers a novel perspective on the record of geostationary IR observations by providing spatially and temporally continuous retrievals of the vertically integrated and vertically resolved concentrations of frozen hydrometeors, typically referred to as ice water path (IWP) and ice water content (IWC). In addition to that, CCIC provides 2D and 3D cloud masks and a 3D cloud classification. A fully convolutional quantile regression neural network constitutes the core of the CCIC retrieval, providing probabilistic estimates of IWP and IWC. The network is trained against CloudSat retrievals using 3.5 years of global collocations. Assessed on a held-out test dataset, the CCIC-provided IWP and IWC estimates achieve correlations exceeding 0.7 and 0.6, respectively, and biases better than −5 % and −2 % demonstrating considerable skill in estimating both IWP and IWC. In addition, CCIC is extensively validated against both in situ and remote sensing measurements from two flight campaign series and a ground-based radar. The results of this independent validation confirm the ability of CCIC to retrieve IWP and IWC. CCIC thus ideally complements temporally and spatially more limited measurements from dedicated cloud sensors by providing spatially and temporally continuous estimates of ice cloud properties. The CCIC network and its associated software are made accessible to the scientific community.

A user-friendly point cloud processing pipeline for interfacing PCL with YARP

A user-friendly point cloud processing pipeline for interfacing PCL with YARP

An oxidation flow reactor for simulating and accelerating secondary aerosol formation in aerosol liquid water and cloud droplets

Abstract. Liquid water in cloud droplets and aqueous aerosols serves as an important reaction medium for the formation of secondary aerosol through aqueous-phase reactions (aqSA). Large uncertainties remain in estimates of the production and chemical evolution of aqSA in the dilute solutions found in cloud droplets and the concentrated solutions found in aerosol liquid water, which is partly due to the lack of available measurement tools and techniques. A new oxidation flow reactor (OFR), the Accelerated Production and Processing of Aerosols (APPA) reactor, was developed to measure secondary aerosol formed through gas- and aqueous-phase reactions, both for laboratory gas mixtures containing one or more precursors and for ambient air. For simulating in-cloud processes, ∼ 3.3 µm diameter droplets formed on monodisperse seed particles are introduced into the top of the reactor, and the relative humidity (RH) inside it is controlled to 100 %. Similar measurements made with the RH in the reactor < 100 % provide contrasts for aerosol formation with no liquid water and with varying amounts of aerosol liquid water. The reactor was characterized through a series of experiments and used to form secondary aerosol from known concentrations of an organic precursor and from ambient air. The residence time distributions of both gases and particles are narrow relative to other OFRs and lack the tails at long residence time expected with laminar flow. Initial cloud processing experiments focused on the well-studied oxidation of dissolved SO2 by O3, with the observed growth of seed particles resulting from the added sulfuric acid agreeing well with estimates based on the relevant set of aqueous-phase reactions. The OH exposure (OHexp) for low RH, high RH, and in-cloud conditions was determined experimentally from the loss of SO2 and benzene and simulated from the KinSim chemical kinetics solver with inputs of the measured 254 nm UV intensity profile through the reactor and loss of O3 due to photolysis. The aerosol yield for toluene at high OHexp ranged from 21.4 % at low RH with dry seed particles present in the reactor to 78.1 % with cloud droplets present. Measurement of the composition of the secondary aerosol formed from ambient air using an aerosol mass spectrometer showed that the oxygen-to-carbon ratio (O : C) of the organic component increased with increasing RH (and liquid water content).

Pathways of ice multiplication in nimbostratus clouds during the Indian summer monsoon

The present study illustrates the microphysical parameters of nimbostratus clouds during the Indian summer monsoon using airborne and radar observations conducted as part of the Cloud Aerosol Interaction and Precipitation Enhancement Experiment (CAIPEEX), and simulations with the Weather Research and Forecasting (WRF) model. The study also analyzes the impact of secondary ice production (emphasizing the microphysical pathways of Hallett-Mossop (HM) as represented in the models) on cloud processes using model sensitivity simulations. The effect of possible pathways of riming during the HM process is the focus of the sensitivity simulations. Aircraft observations showed higher ice particle concentrations in the Hallett–Mossop zone (−3 °C to - 8 °C) along with the existence of smaller and larger cloud droplets, rimed needles, columns, and snow particles. Observations strongly suggested the active presence of the HM process in this cloud. The observed mean values of microphysical parameters including liquid water content, ice number concentrations, and maximum vertical velocity, agreed well with model simulations. The number concentration and water content of ice crystals and snow decreased in the HM zone during the HM inactive case. A reduction in rainfall is also observed in the absence of HM. HM increases convection as it causes a sudden increase in the number concentration of small ice particles, which causes rainwater to freeze quickly and water vapour to be consumed by diffusional growth. Latent heating is produced by both effects, which energizes the convection.

Vertical slice equal sampling and transformer network for point cloud ground filtering

ABSTRACT Ground filtering is a critical yet challenging aspect of Digital Elevation Model (DEM) generation from LiDAR point clouds. Existing deep learning approaches for point cloud processing encounter difficulties in addressing the conflict between the demands of large-scale object filtering tasks and GPU memory. In response to this challenge, this study presents a novel ground filtering method utilizing deep learning. The proposed method incorporates Vertical Slicing Equally Sampling (VSES) to locally sample the original point cloud, effectively organizing the unordered sequence of points and reducing their number while maintaining the representation of the terrain. Subsequently, the resampled point cloud is fed into the TransGF network for classification processing. TransGF is a network model that leverages Transformer as its backbone. Initially, TransGF employs a linear layer to project the input data from a low-dimensional space of (n, 3) to a high-dimensional space of (n, 32). It then extracts features using a module comprising six consecutive Transformer blocks. Finally, TransGF outputs the probability of each point in the samples belonging to either ground or non-ground points. This method is evaluated using the publicly accessible OpenGF ground filtering dataset, showcasing its superior performance in testing scenarios. Noteworthy advantages include structural conciseness, ease of training, and a promising potential for practical applications. The results affirm the effectiveness of TransGF in overcoming challenges associated with ground filtering in LiDAR point cloud processing, positioning it as a valuable contribution to the advancement of DEM generation techniques.

CloudRoots-Amazon22: Integrating Clouds with Photosynthesis by Crossing Scales

Abstract How are rain forest photosynthesis and turbulent fluxes influenced by clouds? To what extent are clouds affected by local processes driven by rain forest energy, water, and carbon fluxes? These interrelated questions were the main drivers of the intensive field experiment CloudRoots-Amazon22 which took place at the Amazon Tall Tower Observatory (ATTO)/Campina supersites in the Amazon rain forest during the dry season, in August 2022. CloudRoots-Amazon22 collected observational data to derive cause–effect relationships between processes occurring at the leaf level up to canopy scales in relation to the diurnal evolution of the clear-to-cloudy transition. First, we studied the impact of cloud and canopy radiation perturbations on the subdiurnal variability of stomatal conductance. Stoma opening is larger in the morning, modulated by the cloud optical thickness. Second, we combined 1-Hz frequency measurements of the stable isotopologues of carbon dioxide and water vapor with measurements of turbulence to determine carbon dioxide and water vapor sources and sinks within the canopy. Using scintillometer observations, we inferred 1-min sensible heat flux that responded within minutes to the cloud passages. Third, collocated profiles of state variables and greenhouse gases enabled us to determine the role of clouds in vertical transport. We then inferred, using canopy and upper-atmospheric observations and a parameterization, the cloud cover and cloud mass flux to establish causality between canopy and cloud processes. This shows the need for a comprehensive observational set to improve weather and climate model representations. Our findings contribute to advance our knowledge of the coupling between cloudy boundary layers and primary carbon productivity of the Amazon rain forest.

Training Warm‐Rain Bulk Microphysics Schemes Using Super‐Droplet Simulations

AbstractCloud microphysics is a critical aspect of the Earth's climate system, which involves processes at the nano‐ and micrometer scales of droplets and ice particles. In climate modeling, cloud microphysics is commonly represented by bulk models, which contain simplified process rates that require calibration. This study presents a framework for calibrating warm‐rain bulk schemes using high‐fidelity super‐droplet simulations that provide a more accurate and physically based representation of cloud and precipitation processes. The calibration framework employs ensemble Kalman methods including Ensemble Kalman Inversion and Unscented Kalman Inversion to calibrate bulk microphysics schemes with probabilistic super‐droplet simulations. We demonstrate the framework's effectiveness by calibrating a single‐moment bulk scheme, resulting in a reduction of data‐model mismatch by more than 75% compared to the model with initial parameters. Thus, this study demonstrates a powerful tool for enhancing the accuracy of bulk microphysics schemes in atmospheric models and improving climate modeling.

Lidar–radar synergistic method to retrieve ice, supercooled water and mixed-phase cloud properties

Abstract. Mixed-phase clouds are not well represented in climate and weather forecasting models, due to a lack of the key processes controlling their life cycle. Developing methods to study these clouds is therefore essential, despite the complexity of mixed-phase cloud processes and the difficulty of observing two cloud phases simultaneously. We propose in this paper a new method to retrieve the microphysical properties of mixed-phase clouds, ice clouds and supercooled water clouds using airborne or satellite radar and lidar measurements, called VarPy-mix. This new approach extends an existing variational method developed for ice cloud retrieval using lidar, radar and passive radiometers. We assume that the lidar attenuated backscatter β at 532 nm is more sensitive to particle concentration and is consequently mainly sensitive to the presence of supercooled water. In addition, radar reflectivity Z at 95 GHz is sensitive to the size of hydrometeors and hence more sensitive to the presence of ice particles. Consequently, in the mixed phase the supercooled droplets are retrieved with the lidar signal and the ice particles with the radar signal, meaning that the retrievals rely strongly on a priori and error values. This method retrieves simultaneously the visible extinction for ice αice and liquid αliq particles, the ice and liquid water contents IWC and LWC, the effective radius of ice re,ice and liquid re,liq particles, and the ice and liquid number concentrations Nice and Nliq. Moreover, total extinction αtot, total water content (TWC) and total number concentration Ntot can also be estimated. As the retrieval of ice and liquid is different, it is necessary to correctly identify each phase of the cloud. To this end, a cloud-phase classification is used as input to the algorithm and has been adapted for mixed-phase retrieval. The data used in this study are from DARDAR-MASK v2.23 products, based on the Cloud-Aerosol Lidar with Orthogonal Polarization (CALIOP) and Cloud Profiling Radar (CPR) observations from the CALIPSO and CloudSat satellites, respectively, belonging to the A-Train constellation launched in 2006. Airborne in situ measurements performed on 7 April 2007 during the Arctic Study of Tropospheric Aerosol, Clouds and Radiation (ASTAR) campaign and collected under the track of CloudSat–CALIPSO are compared with the retrievals of the new algorithm to validate its performance. Visible extinctions, water contents, effective radii and number concentrations derived from in situ measurements and the retrievals showed similar trends and are globally in good agreement. The mean percent error between the retrievals and in situ measurements is 39 % for αliq, 398 % for αice, 49 % for LWC and 75 % for IWC. It is also important to note that temporal and spatial collocations are not perfect, with a maximum spatial shift of 1.68 km and a maximum temporal shift of about 10 min between the two platforms. In addition, the sensitivity of remote sensing and that of in situ measurements is not the same, and in situ measurement uncertainties are between 25 % and 60 %.

Investigating Temporal Characteristics of Polarimetric and Electrical Signatures in Three Severe Storms: Insights from the VORTEX-Southeast Field Campaign

Abstract The dual-polarization radar characteristics of severe storms are commonly used as indicators to estimate the size and intensity of deep convective updrafts. In this study, we track rapid fluctuations in updraft intensity and size by objectively identifying polarimetric fingerprints such as ZDR and KDP columns, which serve as proxies for mixed-phase updraft strength. We quantify the volume of ZDR and KDP columns to evaluate their utility in diagnosing temporal variability in lightning flash characteristics. Specifically, we analyze three severe storms that developed in environments with low-to-moderate instability and strong 0–6-km wind shear in northern Alabama during the 2016–17 VORTEX-Southeast field campaign. In these three cases (a tornadic supercell embedded in stratiform precipitation, a nontornadic supercell, and a supercell embedded within a quasi-linear convective system), we find that the volume of the KDP columns exhibits a stronger correlation with the total flash rate. The higher covariability of the KDP column volume with the total flash rate suggests that the overall electrification and precipitation microphysics were dominated by cold cloud processes. The lower covariability with the ZDR column volume indicates the presence of nonsteady updrafts or a less prominent role of warm rain processes in graupel growth and subsequent electrification. Furthermore, we observe that the majority of cloud-to-ground (CG) lightning strikes a carried negative charge to the ground. In contrast to findings from a tornadic supercell over the Great Plains, lightning flash initiations in the Alabama storms primarily occurred outside the footprint of the ZDR and KDP column objects. Significance Statement This study quantifies the correlation between mixed-phase updraft intensity and total lightning flash rate in three severe storms in northern Alabama. In the absence of direct updraft velocity measurements, we use polarimetric signatures, such as ZDR and KDP columns, as proxies for updraft strength. Our analysis of polarimetric radar and lightning mapping array data reveals that the lightning flash rate is more highly correlated with the KDP column volume than with the ZDR column volume in all three storms examined. This contrasts with previous findings in storms over the central Great Plains, where the ZDR column volume showed higher covariability with flash rate. Interestingly, lightning initiation in the Alabama storms mainly occurred outside the ZDR and KDP column areas, contrary to previous findings.

Guest editorial: Deep learning‐based point cloud processing, compression and analysis

Guest editorial: Deep learning‐based point cloud processing, compression and analysis

Meteorology Modulates the Impact of GCM Horizontal Resolution on Underestimation of Midlatitude Ocean Wind Speeds

AbstractWe utilize ocean 10‐m wind speed (U10m) from the microwave Multi‐sensor Advanced Climatology data set to examine the coupling between convective cloud and precipitation processes, synoptic state, and U10m and to evaluate the representation of U10m in global climate models (GCMs). We find that midlatitude U10m is underestimated by GCMs relative to observations. We examine two potential mechanisms to explain this model behavior: cold pool formation in cold air outbreaks (CAOs) associated with downdrafts that enhance U10m and sea surface temperature (SST) gradients affecting U10m through thermally forced surface winds at regional scales. When the effects of the CAO index (M) and SST gradients on U10m are accounted for, a relationship between GCM horizontal resolution and U10m appears. The strongest correlation between resolution and U10m is over the western boundary currents characterized by frequent CAOs atop strong SST gradients which drives the strongest surface fluxes on Earth.

Reduced‐Order Modeling for Linearized Representations of Microphysical Process Rates

AbstractRepresenting cloud microphysical processes in large scale atmospheric models is challenging because many processes depend on the details of the droplet size distribution (DSD, the spectrum of droplets with different sizes in a cloud). While full or partial statistical moments of droplet size distributions are the typical variables used in bulk models, prognostic moments are limited in their ability to represent microphysical processes across the range of conditions experienced in the atmosphere. Microphysical parameterizations employing prognostic moments are known to suffer from structural uncertainty in their representations of inherently higher dimensional cloud processes, which limit model fidelity and lead to forecasting errors. Here we investigate how data‐driven reduced‐order modeling can be used to learn predictors for microphysical process rates in bulk microphysics schemes in an unsupervised manner from higher dimensional bin distributions. Using simulations characteristic of marine stratiform clouds, we simultaneously learn lower dimensional representations of droplet size distributions and predict the evolution of the microphysical state of the system. Droplet collision‐coalescence, the main process for generating warm rain, is estimated to have an intrinsic dimension of three. This intrinsic dimension provides a lower limit on the number of degrees of freedom needed to accurately represent collision‐coalescence in models. We demonstrate how deep learning based reduced‐order modeling can be used to discover intrinsic coordinates describing the microphysical state of the system, where process rates such as collision‐coalescence are globally linearized. These implicitly learned representations of the DSD retain more information about the DSD than typical moment‐based representations.