Search

Numerical weather prediction (NWP) models often underperform compared to simpler climatology-based precipitation forecasts in northern tropical Africa, even after statistical postprocessing. AI-based forecasting models show promise but have avoided precipitation due to its complexity. Synoptic-scale forcings like African easterly waves and other tropical waves (TWs) are important for predictability in tropical Africa, yet their value for predicting daily rainfall remains unexplored. This study uses two machine-learning models--gamma regression and a convolutional neural network (CNN)--trained on TW predictors from satellite-based GPM IMERG data to predict daily rainfall during the July-September monsoon season. Predictor variables are derived from the local amplitude and phase information of seven TW from the target and up-and-downstream neighboring grids at 1-degree spatial resolution. The ML models are combined with Easy Uncertainty Quantification (EasyUQ) to generate calibrated probabilistic forecasts and are compared with three benchmarks: Extended Probabilistic Climatology (EPC15), ECMWF operational ensemble forecast (ENS), and a probabilistic forecast from the ENS control member using EasyUQ (CTRL EasyUQ). The study finds that downstream predictor variables offer the highest predictability, with downstream tropical depression (TD)-type wave-based predictors being most important. Other waves like mixed-Rossby gravity (MRG), Kelvin, and inertio-gravity waves also contribute significantly but show regional preferences. ENS forecasts exhibit poor skill due to miscalibration. CTRL EasyUQ shows improvement over ENS and marginal enhancement over EPC15. Both gamma regression and CNN forecasts significantly outperform benchmarks in tropical Africa. This study highlights the potential of ML models trained on TW-based predictors to improve daily precipitation forecasts in tropical Africa.

Numerical weather prediction (NWP) models struggle to skillfully predict tropical precipitation occurrence and amount, calling for alternative approaches. For instance, it has been shown that fairly simple, purely data-driven logistic regression models for 24-hour precipitation occurrence outperform both climatological and NWP forecasts for the West African summer monsoon. More complex neural network based approaches, however, remain underdeveloped due to the non-Gaussian character of precipitation. In this study, we develop, apply and evaluate a novel two-stage approach, where we train a U-Net convolutional neural network (CNN) model on gridded rainfall data to obtain a deterministic forecast and then apply the recently developed, nonparametric Easy Uncertainty Quantification (EasyUQ) approach to convert it into a probabilistic forecast. We evaluate CNN+EasyUQ for one-day ahead 24-hour accumulated precipitation forecasts over northern tropical Africa for 2011--2019, with the Integrated Multi-satellitE Retrievals for GPM (IMERG) data serving as ground truth. In the most comprehensive assessment to date we compare CNN+EasyUQ to state-of-the-art physics-based and data-driven approaches such as a monthly probabilistic climatology, raw and postprocessed ensemble forecasts from the European Centre for Medium-Range Weather Forecasts (ECMWF), and traditional statistical approaches that use up to 25 predictor variables from IMERG and the ERA5 reanalysis.Generally, statistical approaches perform about en par with post-processed ECMWF ensemble forecasts. The CNN+EasyUQ approach, however, clearly outperforms all competitors for both occurrence and amount. Hybrid methods that merge CNN+EasyUQ and physics-based forecasts show slight further improvement. Thus, the CNN+EasyUQ approach can likely improve operational probabilistic forecasts of rainfall in the tropics, and potentially even beyond.

Scientific communities are increasingly adopting machine learning and deep learning models in their applications to accelerate scientific insights. High performance computing systems are pushing the frontiers of performance with a rich diversity of hardware resources and massive scale-out capabilities. There is a critical need to understand fair and effective benchmarking of machine learning applications that are representative of real-world scientific use cases. MLPerf is a community-driven standard to benchmark machine learning workloads, focusing on end-to-end performance metrics. In this paper, we introduce MLPerf HPC, a benchmark suite of large-scale scientific machine learning training applications driven by the MLCommons Association. We present the results from the first submission round, including a diverse set of some of the world's largest HPC systems. We develop a systematic framework for their joint analysis and compare them in terms of data staging, algorithmic convergence, and compute performance. As a result, we gain a quantitative understanding of optimizations on different subsystems such as staging and on-node loading of data, compute-unit utilization, and communication scheduling, enabling overall $>10 \times$ (end-to-end) performance improvements through system scaling. Notably, our analysis shows a scale-dependent interplay between the dataset size, a system's memory hierarchy, and training convergence that underlines the importance of near-compute storage. To overcome the data-parallel scalability challenge at large batch sizes, we discuss specific learning techniques and hybrid data-and-model parallelism that are effective on large systems. We conclude by characterizing each benchmark with respect to low-level memory, I/O, and network behavior to parameterize extended roofline performance models in future rounds.

Non-linear phase field models are increasingly used for the simulation of fracture propagation models. The numerical simulation of fracture networks of realistic size requires the efficient parallel solution of large coupled non-linear systems. Although in principle efficient iterative multi-level methods for these types of problems are available, they are not widely used in practice due to the complexity of their parallel implementation. Here, we present Utopia, which is an open-source C++ library for parallel non-linear multilevel solution strategies. Utopia provides the advantages of high-level programming interfaces while at the same time a framework to access low-level data-structures without breaking code encapsulation. Complex numerical procedures can be expressed with few lines of code, and evaluated by different implementations, libraries, or computing hardware. In this paper, we investigate the parallel performance of our implementation of the recursive multilevel trust-region (RMTR) method based on the Utopia library. RMTR is a globally convergent multilevel solution strategy designed to solve non-convex constrained minimization problems. In particular, we solve pressure-induced phase-field fracture propagation in large and complex fracture networks. Solving such problems is deemed challenging even for a few fractures, however, here we are considering networks of realistic size with up to 1000 fractures., Comment: CCF Trans. HPC (2021)

The predictability of eight southern European tropical-like cyclones, seven of which Medicanes, is studied evaluating ECMWF operational ensemble forecasts against operational analysis data. Forecast cyclone trajectories are compared to the cyclone trajectory in the analysis by means of a dynamic time warping technique, which allows to find a match in terms of their overall spatio-temporal similarity. Each storm is treated as an object and its forecasts are analysed using metrics that describe intensity, symmetry, compactness, and upper-level thermal structure. This object-based approach allows to focus on specific storm features, while tolerating their shifts in time and space to some extent. The compactness and symmetry of the storms are generally underpredicted, especially at long lead times. However, forecast accuracy tends to strongly improve at short lead times, indicating that the ECMWF ensemble forecast model can adequately reproduce Medicanes, albeit only few days in advance. In particular, late forecasts which have been initialised when the cyclone has already developed are distinctly more accurate than earlier forecasts in predicting its kinematic and thermal structure, confirming previous findings of high sensitivity of Medicane simulations to initial conditions. Findings reveal a markedly non-gradual evolution of ensemble forecasts with lead time. Specifically, a rapid increase in the probability of cyclone occurrence ("forecast jump") is seen in most cases, generally between 5 and 7 days lead time. Jumps are also found for ensemble median and/or spread for storm thermal structure forecasts. This behaviour is compatible with the existence of predictability barriers. On the other hand, storm position forecasts often exhibit a consistent spatial distribution of storm position uncertainty and bias between consecutive forecasts.

Tropical cyclones that evolve from a non-tropical origin may pose a special challenge for predictions, as they often emerge at the end of a multi-scale cascade of atmospheric processes. Climatological studies have shown that the 'tropical transition' (TT) pathway plays a prominent role in cyclogenesis, in particular over the North Atlantic Ocean. Here we use operational European Centre for Medium-Range Weather Forecasts ensemble predictions to investigate the TT of North Atlantic Hurricane Chris (2012), whose formation was preceded by the merger of two potential vorticity (PV) maxima, eventually resulting in the storm-inducing PV streamer. The principal goal is to elucidate the dynamic and thermodynamic processes governing the predictability of cyclogenesis and subsequent TT. Dynamic time warping is applied to identify ensemble tracks that are similar to the analysis track. This technique permits small temporal and spatial shifts in the development. The formation of the pre-Chris cyclone is predicted by those members that also predict the merging of the two PV maxima. The position of the storm relative to the PV streamer determines whether the pre-Chris cyclone follows the TT pathway. The transitioning storms are located inside a favorable region of high equivalent potential temperatures that result from a warm seclusion underneath the cyclonic roll-up of the PV streamer. A systematic investigation of consecutive ensemble forecasts indicates that forecast improvements are linked to specific events, such as the PV merging. The present case exemplifies how a novel combination of Eulerian and Lagrangian ensemble forecast analysis tool allows to infer physical causes of abrupt changes in predictability., Comment: 27 pages, 15 figures, supplementary material; submitted to Monthly Weather Review

This study presents the first systematic comparison of the (thermo-)dynamics associated with all major tropical wave types causing rainfall modulation over northern tropical Africa: Madden Julian Oscillation (MJO), Equatorial Rossby waves (ERs), mixed Rossby-gravity waves (MRGs), Kelvin waves, tropical disturbances (TDs, including African Easterly Waves), and eastward inertio-gravity waves (EIGs). Reanalysis and radiosonde data were analyzed for the period 1981--2013 based on space-time filtering of outgoing longwave radiation. The identified circulation patterns are largely consistent with theory. The slow modes, MJO and ER, mainly impact precipitable water, whereas the faster Kelvin waves, MRGs, and TDs primarily modulate moisture convergence. Monsoonal inflow intensifies during wet phases of the MJO, ERs, and MRGs, associated with a northward shift of the intertropical discontinuity for MJO and ERs. During passages of vertically tilted imbalanced wave modes, such as MJO, Kelvin waves, and TDs, and partly MRGs, increased vertical wind shear and improved conditions for up- and downdrafts facilitate the organization of convection. The balanced ERs are not tilted and rainfall is triggered by large-scale moistening and stratiform lifting. The MJO and ERs interact with intraseasonal variations of the Indian monsoon and extratropical Rossby wave trains. The latter causes a trough over the Atlas Mountains associated with a tropical plume and rainfall over the Sahara. Positive North Atlantic and Arctic Oscillation signals precede tropical plumes in case of the MJO. The results unveil which dynamical processes need to be modeled realistically to represent the coupling between tropical waves and rainfall in northern tropical Africa., Comment: 33 pages, 11 figures, supplementary material; submitted to Journal of Climate

Low-latitude rainfall variability on the daily to intraseasonal timescale is often related to tropical waves, including convectively coupled equatorial waves, the Madden-Julian Oscillation (MJO), and tropical disturbances. Despite the importance of rainfall variability for vulnerable societies in tropical Africa, the relative influence of tropical waves for this region is largely unknown. This article presents the first systematic comparison of the impact of six wave types on precipitation over northern tropical Africa during the transition and full monsoon seasons, using two satellite products and a dense rain gauge network. Composites of rainfall anomalies in the different datasets show} comparable modulation intensities in the West Sahel and at the Guinea Coast, varying from less than 2 to above 7 mm/d depending on the wave type. African Easterly Waves (AEWs) and Kelvin waves dominate the 3-hourly to daily timescale and explain 10-30% locally. On longer timescales (7-20d), only the MJO and equatorial Rossby (ER) waves remain as modulating factors and explain about up to one third of rainfall variability. Eastward inertio-gravity waves and mixed Rossby-gravity (MRG) waves are comparatively unimportant. An analysis of wave superposition shows that low-frequency waves (MJO, ER) in their wet phase amplify the activity of high-frequency waves (TD, MRG) and suppress them in the dry phase. The results stress that more attention should be paid to tropical waves when forecasting rainfall over northern tropical Africa., Comment: 34 pages, 12 figures, supplementary material; submitted to Journal of Climate

Accumulated precipitation forecasts are of high socioeconomic importance for agriculturally dominated societies in northern tropical Africa. In this study, we analyze the performance of nine operational global ensemble prediction systems (EPSs) relative to climatology-based forecasts for 1 to 5-day accumulated precipitation based on the monsoon seasons 2007-2014 for three regions within northern tropical Africa. To assess the full potential of raw ensemble forecasts across spatial scales, we apply state-of-the-art statistical postprocessing methods in form of Bayesian Model Averaging (BMA) and Ensemble Model Output Statistics (EMOS), and verify against station and spatially aggregated, satellite-based gridded observations. Raw ensemble forecasts are uncalibrated, unreliable, and underperform relative to climatology, independently of region, accumulation time, monsoon season, and ensemble. Differences between raw ensemble and climatological forecasts are large, and partly stem from poor prediction for low precipitation amounts. BMA and EMOS postprocessed forecasts are calibrated, reliable, and strongly improve on the raw ensembles, but - somewhat disappointingly - typically do not outperform climatology. Most EPSs exhibit slight improvements over the period 2007-2014, but overall have little added value compared to climatology. We suspect that the parametrization of convection is a potential cause for the sobering lack of ensemble forecast skill in a region dominated by mesoscale convective systems.