Abstract

Machine learning models have been employed to perform either physics-free data-driven or hybrid dynamical downscaling of climate data. Most of these implementations operate over relatively small downscaling factors because of the challenge of recovering fine-scale information from coarse data. This limits their compatibility with many global climate model outputs, often available between ∼50–100 km resolution, to scales of interest such as cloud resolving or urban scales. This study systematically examines the capability of a type of super-resolving convolutional neural network (SR-CNNs) to downscale surface wind speed data over land surface from different coarse resolutions (25 km, 48 km, and 100 km resolution) to 3 km. For each downscaling factor, we consider three CNN configurations that generate super-resolved predictions of fine-scale wind speed, which take between 1 to 3 input fields: coarse wind speed, fine-scale topography, and diurnal cycle. In addition to fine-scale wind speeds, probability density function parameters are generated, through which sample wind speeds can be generated accounting for the intrinsic stochasticity of wind speed. For generalizability assessment, CNN models are tested on regions with different topography and climate that are unseen during training. The evaluation of super-resolved predictions focuses on subgrid-scale variability and the recovery of extremes. Models with coarse wind and fine topography as inputs exhibit the best performance compared with other model configurations, operating across the same downscaling factor. Our diurnal cycle encoding results in lower out-of-sample generalizability compared with other input configurations.

Abstract

Hyperparameter optimization (HPO) is an important step in machine learning (ML) model development, but common practices are archaic—primarily relying on manual or grid searches. This is partly because adopting advanced HPO algorithms introduces added complexity to the workflow, leading to longer computation times. This poses a notable challenge to ML applications, as suboptimal hyperparameter selections curtail the potential of ML model performance, ultimately obstructing the full exploitation of ML techniques. In this article, we present a two-step HPO method as a strategic solution to curbing computational demands and wait times, gleaned from practical experiences in applied ML parameterization work. The initial phase involves a preliminary evaluation of hyperparameters on a small subset of the training dataset, followed by a re-evaluation of the top-performing candidate models post-retraining with the entire training dataset. This two-step HPO method is universally applicable across HPO search algorithms, and we argue it has attractive efficiency gains.

As a case study, we present our recent application of the two-step HPO method to the development of neural network emulators for aerosol activation. Although our primary use case is a data-rich limit with many millions of samples, we also find that using up to 0.0025% of the data—a few thousand samples—in the initial step is sufficient to find optimal hyperparameter configurations from much more extensive sampling, achieving up to 135× speed-up. The benefits of this method materialize through an assessment of hyperparameters and model performance, revealing the minimal model complexity required to achieve the best performance. The assortment of top-performing models harvested from the HPO process allows us to choose a high-performing model with a low inference cost for efficient use in global climate models (GCMs).

Abstract

Vertical profiles of temperature and dewpoint are useful in predicting deep convection that leads to severe weather which threatens property and lives. Currently, forecasters rely on observations from radiosonde launches and numerical weather prediction (NWP) models. Radiosonde observations are, however, temporally and spatially sparse, and NWP models contain inherent errors that influence short-term predictions of high impact events. This work explores using machine learning (ML) to postprocess NWP model forecasts, combining them with satellite data to improve vertical profiles of temperature and dewpoint. We focus on different ML architectures, loss functions, and input features to optimize predictions. Because we are predicting vertical profiles at 256 levels in the atmosphere, this work provides a unique perspective at using ML for 1-D tasks. Compared to baseline profiles from the Rapid Refresh (RAP), ML predictions offer the largest improvement for dewpoint, particularly in the mid- and upper-atmosphere. Temperature improvements are modest, but CAPE values are improved by up to 40%. Feature importance analyses indicate that the ML models are primarily improving incoming RAP biases. While additional model and satellite data offer some improvement to the predictions, architecture choice is more important than feature selection in fine-tuning the results. Our proposed deep residual UNet performs the best by leveraging spatial context from the input RAP profiles; however, the results are remarkably robust across model architecture. Further, uncertainty estimates for every level are well-calibrated and can provide useful information to forecasters.

Abstract

Visible and infrared radiance products of geostationary orbiting platforms provide virtually continuous observations of Earth. In contrast, low-Earth orbiters observe passive microwave (PMW) radiances at any location much less frequently. Prior literature demonstrates the ability of a machine learning (ML) approach to build a link between these two complementary radiance spectra by predicting PMW observations using infrared and visible products collected from geostationary instruments, which could potentially deliver a highly desirable synthetic PMW product with nearly continuous spatiotemporal coverage. However, current ML models lack the ability to provide a measure of uncertainty of such a product, significantly limiting its applications. In this work, Bayesian deep learning is employed to generate synthetic Global Precipitation Measurement (GPM) Microwave Imager (GMI) data from Advanced Baseline Imager (ABI) observations with attached uncertainties over the ocean. The study first uses deterministic residual networks (ResNets) to generate synthetic GMI brightness temperatures with as little mean absolute error as 1.72 K at the ABI spatiotemporal resolution. Then, for the same task, we use three Bayesian ResNet models to produce a comparable amount of error while providing previously unavailable predictive variance (i.e., uncertainty) for each synthetic data point. We find that the Flipout configuration provides the most robust calibration between uncertainty and error across GMI frequencies, and then demonstrate how this additional information is useful for discarding high-error synthetic data points prior to use by downstream applications.

Abstract

Ships inside the Arctic basin require high-resolution (1–5 km), near-term (days to semimonthly) forecasts for guidance on scales of interest to their operations where forecast model predictions are insufficient due to their coarse spatial and temporal resolutions. Deep learning techniques offer the capability of rapid assimilation and analysis of multiple sources of information for improved forecasting. Data from the National Oceanographic and Atmospheric Administration’s Global Forecast System, Multi-scale Ultra-high Resolution Sea Surface Temperature (MEaSUREs), and the National Snow and Ice Data Center’s Multisensor Analyzed Sea ice Extent (MASIE) were used to develop the sea ice extent deep learning forecast model, over the freeze-up periods of 2016, 2018, 2019, and 2020 in the Beaufort Sea. Sea ice extent forecasts were produced for 1–7 days in the future. The approach was novel for sea ice extent forecasting in using forecast data as model input to aid in the prediction of sea ice extent. Model accuracy was assessed against a persistence model. While the average accuracy of the persistence model dropped from 97% to 90% for forecast days 1–7, the deep learning model accuracy dropped only to 93%. A k-fold (fourfold) cross-validation study found that on all except the first day, the deep learning model, which includes a U-Net architecture with an 18-layer residual neural network (Resnet-18) backbone, does better than the persistence model. Skill scores improve the farther out in time to 0.27. The model demonstrated success in predicting changes in ice extent of significance for navigation in the Amundsen Gulf. Extensions to other Arctic seas, seasons, and sea ice parameters are under development.

Abstract

Weather forecasting centers currently rely on statistical postprocessing methods to minimize forecast error. This improves skill but can lead to predictions that violate physical principles or disregard dependencies between variables, which can be problematic for downstream applications and for the trustworthiness of postprocessing models, especially when they are based on new machine learning approaches. Building on recent advances in physics-informed machine learning, we propose to achieve physical consistency in deep learning-based postprocessing models by integrating meteorological expertise in the form of analytic equations. Applied to the post-processing of surface weather in Switzerland, we find that constraining a neural network to enforce thermodynamic state equations yields physically-consistent predictions of temperature and humidity without compromising performance. Our approach is especially advantageous when data is scarce, and our findings suggest that incorporating domain expertise into postprocessing models allows the optimization of weather forecast information while satisfying application-specific requirements.

Abstract

Skillful weather prediction on subseasonal to seasonal time scales is crucial for many socioeconomic ventures. But forecasting, especially extremes, on these time scales is very challenging because the information from initial conditions is gradually lost. Therefore, data-driven methods are discussed as an alternative to numerical weather prediction models. Here, quantile regression forests (QRFs) and random forest classifiers (RFCs) are used for probabilistic forecasting of central European mean wintertime 2-m temperatures and cold wave days at lead times of 14, 21, and 28 days. ERA5 reanalysis meteorological predictors are used as input data for the machine learning models. For the winters of 2000/01–2019/20, the predictions are compared with a climatological ensemble obtained from E-OBS observational data. The evaluation is performed as full distribution predictions for continuous values using the continuous ranked probability skill score and as binary categorical forecasts using the Brier skill score. We find skill at lead times up to 28 days in the 20-winter mean and for individual winters. Case studies show that all used machine learning models are able to learn patterns in the data beyond climatology. A more detailed analysis using Shapley additive explanations suggests that both random forest (RF)-based models are able to learn physically known relationships in the data. This underlines that RF-based data-driven models can be a suitable tool for forecasting central European mean wintertime 2-m temperatures and the occurrence of cold wave days.

Abstract

Emulating numerical weather prediction (NWP) model outputs is important to compute large datasets of weather fields in an efficient way. The purpose of the present paper is to investigate the ability of generative adversarial networks (GANs) to emulate distributions of multivariate outputs (10-m wind and 2-m temperature) of a kilometer-scale NWP model. For that purpose, a residual GAN architecture, regularized with spectral normalization, is trained against a kilometer-scale dataset from the AROME Ensemble Prediction System (AROME-EPS). A wide range of metrics is used for quality assessment, including pixelwise and multiscale Earth-mover distances, spectral analysis, and correlation length scales. The use of wavelet-based scattering coefficients as meaningful metrics is also presented. The GAN generates samples with good distribution recovery and good skill in average spectrum reconstruction. Important local weather patterns are reproduced with a high level of detail, while the joint generation of multivariate samples matches the underlying AROME-EPS distribution. The different metrics introduced describe the GAN’s behavior in a complementary manner, highlighting the need to go beyond spectral analysis in generation quality assessment. An ablation study then shows that removing variables from the generation process is globally beneficial, pointing at the GAN limitations to leverage cross-variable correlations. The role of absolute positional bias in the training process is also characterized, explaining both accelerated learning and quality-diversity trade-off in the multivariate emulation. These results open perspectives about the use of GAN to enrich NWP ensemble approaches, provided that the aforementioned positional bias is properly controlled.

Abstract

Exploring the climate impacts of various anthropogenic emissions scenarios is key to making informed decisions for climate change mitigation and adaptation. State-of-the-art Earth system models can provide detailed insight into these impacts but have a large associated computational cost on a per-scenario basis. This large computational burden has driven recent interest in developing cheap machine learning models for the task of climate model emulation. In this paper, we explore the efficacy of randomly wired neural networks for this task. We describe how they can be constructed and compare them with their standard feedforward counterparts using the ClimateBench dataset. Specifically, we replace the serially connected dense layers in multilayer perceptrons, convolutional neural networks, and convolutional long short-term memory networks with randomly wired dense layers and assess the impact on model performance for models with 1 million and 10 million parameters. We find that models with less-complex architectures see the greatest performance improvement with the addition of random wiring (up to 30.4% for multilayer perceptrons). Furthermore, of 24 different model architecture, parameter count, and prediction task combinations, only one had a statistically significant performance deficit in randomly wired networks relative to their standard counterparts, with 14 cases showing statistically significant improvement. We also find no significant difference in prediction speed between networks with standard feedforward dense layers and those with randomly wired layers. These findings indicate that randomly wired neural networks may be suitable direct replacements for traditional dense layers in many standard models.

Abstract

Physics-based simulations of Arctic sea ice are highly complex, involving transport between different phases, length scales, and time scales. Resultantly, numerical simulations of sea ice dynamics have a high computational cost and model uncertainty. We employ data-driven machine learning (ML) to make predictions of sea ice motion. The ML models are built to predict present-day sea ice velocity given present-day wind velocity and previous-day sea ice concentration and velocity. Models are trained using reanalysis winds and satellite-derived sea ice properties. We compare the predictions of three different models: persistence (PS), linear regression (LR), and a convolutional neural network (CNN). We quantify the spatiotemporal variability of the correlation between observations and the statistical model predictions. Additionally, we analyze model performance in comparison to variability in properties related to ice motion (wind velocity, ice velocity, ice concentration, distance from coast, bathymetric depth) to understand the processes related to decreases in model performance. Results indicate that a CNN makes skillful predictions of daily sea ice velocity with a correlation up to 0.81 between predicted and observed sea ice velocity, while the LR and PS implementations exhibit correlations of 0.78 and 0.69, respectively. The correlation varies spatially and seasonally: lower values occur in shallow coastal regions and during times of minimum sea ice extent. LR parameter analysis indicates that wind velocity plays the largest role in predicting sea ice velocity on 1-day time scales, particularly in the central Arctic. Regions where wind velocity has the largest LR parameter are regions where the CNN has higher predictive skill than the LR.