Ensemble Prediction Research Articles

Biomolecules rely on water and ions for stable folding, but these interactions are often transient, dynamic, or disordered and thus hidden from experiments and evaluation challenges that represent biomolecules as single, ordered structures. Here, we compare blindly predicted ensembles of water and ion structure to the cryo-EM densities observed around the Tetrahymena ribozyme at 2.2-2.3 Å resolution, collected through target R1260 in the CASP16 competition. Twenty-six groups participated in this solvation "cryo-ensemble" prediction challenge, submitting over 350 million atoms in total, offering the first opportunity to compare blind predictions of dynamic solvent shell ensembles to cryo-EM density. Predicted atomic ensembles were converted to density through local alignment and these densities were compared to the cryo-EM densities using Pearson correlation, Spearman correlation, mutual information, and precision-recall curves. These predictions show that an ensemble representation is able to capture information of transient or dynamic water and ions better than traditional atomic models, but there remains a large accuracy gap to the performance ceiling set by experimental uncertainty. Overall, molecular dynamics approaches best matched the cryo-EM density, with blind predictions from bussilab_plain_md, SoutheRNA, bussilab_replex, coogs2, and coogs3 outperforming the baseline molecular dynamics prediction. This study indicates that simulations of water and ions can be quantitatively evaluated with cryo-EM maps. We propose that further community-wide blind challenges can drive and evaluate progress in modeling water, ions, and other previously hidden components of biomolecular systems.

Abstract The remnants of Hurricane Ida demonstrated the destructive potential of extratropically transitioning storms in New York City (NYC) on 1–2 September 2021. Central Park set a new hourly rainfall record of 3.47 inches (88.14 mm) from 0100 to 0200 UTC on 2 September, far exceeding the city’s stormwater capacity. Historic flash flooding killed 13 residents and caused hundreds of millions of dollars in damage. Despite the impacts, the physical mechanisms responsible for this record-breaking rainfall in NYC remain underexplored. In this study, we use the NOAA National Severe Storms Laboratory’s Warn-on-Forecast System (WoFS), an ensemble forecast and data assimilation system, to examine the role of atmospheric processes across spatial scales in producing extreme hourly rainfall rates in NYC. We apply multiple analysis techniques, including ensemble sensitivity analysis and direct comparison of the wettest and driest ensemble members. We also implement a new front-detection algorithm to investigate how front location influences individual supercell trajectories. Our analyses support past findings that low-topped supercells embedded in Ida produced the highest rainfall rates. The number and trajectory of these storms were shaped by the synoptic circulation, which affected the steering flow, and by the strength of the low-level jet transporting tropical moisture and instability to the slow-moving warm front. Additionally, NYC rainfall rates were sensitive to the strength and position of the warm front, with variations shifting maximum precipitation too far northwest in some members. Our results highlight important features at the synoptic, meso-, and convective scales with the potential to improve the predictability of future extreme rainfall events.

Accurately forecasting the demand for air passengers is vital for the aviation industry to formulate appropriate management strategies. Decomposition ensemble learning has attracted much attention from researchers of this problem because it is an effective way to improve forecasting accuracy. In contrast to common ways of generating ensemble forecasts, such as artificial intelligence and linear addition, our study employs the Choquet fuzzy integral. The Choquet integral is effective regardless of the training sample size and it uses a nonadditive fuzzy measure to explain the influence of the inputs on air passenger demand. Data on monthly air passenger flows from major airports in Taiwan were used to assess the effectiveness of the proposed decomposition ensemble models using the Choquet fuzzy integral to generate ensemble forecasts. The results in terms of level and directional forecasting accuracy showed that the proposed models— especially those that integrated smoothing (LOESS) (STL) and radial basis function network with the Choquet integral—significantly outperformed single (non-ensemble) forecasting models and the benchmark models considered.

Abstract. Data assimilation (DA) is an essential component of numerical weather and climate prediction. Efficient implementation of DA algorithms benefits both research and operational prediction. Currently, a variety of DA software programs are available. One of the notable DA libraries is the Parallel Data Assimilation Framework (PDAF) designed for ensemble data assimilation. The DA framework is widely used with complex high-dimensional climate models, and is applied for research on atmosphere, ocean, sea ice and marine ecosystem modelling, as well as operational ocean forecasting. Meanwhile, there are increasing demands for flexible and efficient DA implementations using Python due to the increasing amount of intermediate complexity models as well as machine learning based models coded in Python. To accommodate for such demands, we introduce a Python interface to PDAF, pyPDAF. pyPDAF allows for flexible DA system development while retaining the efficient implementation of the core DA algorithms in the Fortran-based PDAF. The ideal use-case of pyPDAF is a DA system where the model integration is independent from the DA program, which reads the model forecast ensemble, produces an analysis, and updates the restart files of the model, or a DA system where the model can be used in Python. With implementations of both PDAF and pyPDAF, this study demonstrates the use of pyPDAF and PDAF in a coupled data assimilation (CDA) setup in a coupled atmosphere-ocean model, the Modular Arbitrary-Order Ocean-Atmosphere Model (MAOOAM). This study demonstrates that pyPDAF allows for PDAF functionalities from Python where users can utilise Python functions to handle case-specific information from observations and numerical model. The study also shows that pyPDAF can be used with high-dimensional systems with little slow-down per analysis step of only up to 13 % for the localized ensemble Kalman filter LETKF in the example used in this study. The study also shows that, compared to PDAF, the overhead of pyPDAF is comparatively smaller when computationally intensive components dominate the DA system. This can be the case for systems with high-dimensional state vectors.

Modeling the conformational heterogeneity of protein-small molecule interactions is important for understanding natural systems and evaluating designed systems but remains an outstanding challenge. We reasoned that while residue-level descriptions of biomolecules are efficient for de novo structure prediction, for probing heterogeneity of interactions with small molecules in the folded state, an entirely atomic-level description could have advantages in speed and generality. We developed a graph neural network called PLACER (protein-ligand atomistic conformational ensemble resolver) trained to recapitulate correct atomic positions from partially corrupted input structures from the Cambridge Structural Database and the Protein Data Bank; the nodes of the graph are the atoms in the system. PLACER accurately generates structures of diverse organic small molecules given knowledge of their atom composition and bonding. When given a description of the larger protein context, it builds up structures of small molecules and protein side chains for protein-small molecule docking. Because PLACER is rapid and stochastic, ensembles of predictions can be readily generated to map conformational heterogeneity. In enzyme design efforts described here and elsewhere, we find that using PLACER to assess the accuracy and preorganization of the designed active sites results in higher success rates and higher activities; we obtain a preorganized retroaldolase with a kcat/KM of 11,000 M-1min-1, considerably higher than any pre-deep learning design for this reaction. We anticipate that PLACER will be widely useful for rapidly generating conformational ensembles of small molecule and small molecule-protein systems and for designing higher activity preorganized enzymes.

Background: The scientific literature on artificial intelligence-enabled electrocardiography (AI-ECG) has defined a robust performance of AI models in detecting and predicting several structural heart disorders (SHDs) using ECGs. However, as a diagnostic test, the real-world clinical utility of AI-ECG reliability requires the consistency of its results when repeated under similar conditions. Aim: To evaluate the reliability of AI-ECG models for different ECGs for the same person, across different diagnostic labels, and using varied modeling approaches. Methods: We used ECG images (2000-2024) from 5 hospitals and an outpatient network within a large, integrated US health system. For each individual, we identified multiple ECGs recorded within a 30-day period. We evaluated 7 models: 6 convolutional neural networks (CNNs) trained to detect individual SHDs, including LV systolic dysfunction, left valve diseases and severe LVH; an ensemble XGBoost integrating individual CNNs as a composite screen for multiple SHDs. We used concordance correlation coefficient (CCC), Spearman correlation, Cohen’s kappa, and percent agreement in binary screen status to test model reliability. We evaluated factors associated with different AI-ECG outputs (Δ probability> 0.5) and assessed stability across ECG layouts (digital, printed, photo). Results: Across sites, we identified 1,118,263 ECG pairs, with a median 1 (1-3) days between ECGs. The ensemble XGBoost had the higher test-retest correlation (CCC: 0.89-0.92) and agreement (kappa: 0.75-0.82) between pairs compared with CNNs (CCC: 0.78-0.88; kappa: 0.57-0.72). After adjusting for demographics, ECG pairs that included one or both inpatient ECG were significantly more likely to yield unstable predictions (ORs: 1.60 [1.50-1.70] and 1.91 [1.78-2.05], respectively) compared with pairs with both ECGs obtained in outpatient settings. Among outpatient pairs across sites, the XGBoost model had a CCC of 0.89-0.94, a Spearman correlation of 0.90-0.94, and a kappa of 0.78-0.84, with concordance rates of 89-92%. Notably, ensemble model predictions were also stable across different ECG layouts. Conclusion: An ensemble AI-ECG model integrating multiple CNN predictions had higher reliability compared with models for individual disorders. Discordance was more common in inpatient ECGs, suggesting instability in high-acuity settings. Reliable ensemble AI-ECG model outputs support readiness for clinical implementation for SHD screening.

Abstract. Prediction and mitigation of extreme weather events are important scientific and societal challenges. Recently, Miyoshi and Sun (2022) proposed a control simulation experiment framework that assesses the controllability of chaotic systems under observational uncertainty, and within this framework, Sun et al. (2023) developed a method to prevent extreme events in the Lorenz 96 model. However, since their method is primarily designed to apply control inputs to all grid variables, the success rate decreases to approximately 60 % when applied to a single site, at least in a specific setting. Herein, we propose an approach that mitigates extreme events by updating local interventions based on multi-scenario ensemble forecasts. Our method achieves a high success rate, reaching 94 % even when applying interventions at one site per step, albeit with a moderate increase in the intervention cost. Furthermore, the success rate increases to 99.4 % for interventions at two sites. Unlike control-theoretic approaches adopting a top–down strategy, which determine inputs by optimizing cost functions, our bottom–up approach mitigates extreme events by effectively utilizing limited intervention options.

Abstract. This paper describes a new version of the Real-time Air Quality Modeling System (RAQMS) which uses National Unified Operational Prediction Capability (NUOPC) coupling to combine the RAQMS chemical mechanism with the Global Ensemble Forecasting System with Aerosols (GEFS-Aerosols), the Goddard Chemistry Aerosol Radiation and Transport model (GOCART) aerosol mechanism, and NOAA's Unified Forecast System (UFS) version 9.1 Finite Volume Cubed Sphere (FV3) dynamical core. We also present an application of TROPOMI CO column data assimilation in UFS-RAQMS with the NOAA Grid Point Statistical Interpolation (GSI) three-dimensional variational (3D-Var) analysis system to constrain UFS-RAQMS CO. We validate UFS-RAQMS control and TROPOMI CO data assimilation CO analyses for the period 15 July–30 September 2019 against independent satellite, ground-based, and airborne observations. We show that the largest impacts of the TROPOMI CO data assimilation are in the lower troposphere over Siberia and Indonesia. We find that UFS-RAQMS biomass burning signatures in CO column are not consistent with those in aerosol optical depth (AOD) near the Siberian and Indonesian biomass burning source regions within our control experiment. Assimilation of TROPOMI CO improves the representation of the biomass burning AOD/CO relationship in UFS-RAQMS by increasing the CO column, which suggests that the biomass burning CO emissions from the Blended Global Biomass Burning Emissions Product (GBBEPx) used in UFS-RAQMS are too low for boreal wildfires.

Accurate crop yield prediction remains a cornerstone of sustainable agriculture and food security, particularly in regions vulnerable to climate fluctuations such as Sub-Saharan Africa. This study develops an adaptive hybrid ensemble learning model that integrates climatic and soil parameters to improve crop yield prediction accuracy. The proposed framework combines Decision Tree Regressor and Ridge Regression as base learners, while Linear Regression serves as a meta-model to optimize ensemble predictions. A dataset spanning 1990–2020 was analyzed and preprocessed using normalization and feature selection techniques based on agronomic significance. Model optimization was performed using GridSearchCV to fine-tune hyperparameters. Experimental results revealed that the stacking ensemble achieved superior performance, with an RMSE of 0.1318, MAE of 0.0804, and R² of 0.9766, outperforming individual models. The findings underscore the effectiveness of hybrid ensemble methods in modeling nonlinear agricultural systems and demonstrate the potential of machine learning to support data-driven agricultural decision-making. Future work will explore dynamic adaptation to real-time environmental data and regional transferability across diverse agricultural ecosystems.

Multimedia learning effectiveness varies widely across cultural contexts and individual learner characteristics, yet existing educational technologies lack computational frameworks that predict and optimize these interactions. This study introduces the Multimedia Integration Impact Assessment Model (MIIAM), a machine learning framework integrating cognitive style detection, cultural background inference, multimedia complexity optimization, and ensemble prediction into a unified architecture. MIIAM was validated with 493 software engineering students from Zimbabwe and South Africa through the analysis of 4.1 million learning interactions. The framework applied Random Forests for automated cognitive style classification, hierarchical clustering for cultural inference, and a complexity optimization engine for content analysis, while predictive performance was enhanced by an ensemble of Random Forests, XGBoost, and Neural Networks. The results demonstrated that MIIAM achieved 87% prediction accuracy, representing a 14% improvement over demographic-only baselines (p < 0.001). Cross-cultural validation confirmed strong generalization, with only a 2% accuracy drop compared to 11–15% for traditional models, while fairness analysis indicated substantially reduced bias (Statistical Parity Difference = 0.08). Real-time testing confirmed deployment feasibility with an average 156 ms processing time. MIIAM also optimized multimedia content, improving knowledge retention by 15%, reducing cognitive overload by 28%, and increasing completion rates by 22%. These findings establish MIIAM as a robust, culturally responsive framework for adaptive multimedia learning environments.

Abstract As a component in seamless weather–climate prediction, subseasonal forecasts provide essential guidance for decision-making across multiple sectors. Utilizing hindcasts from six models in Subseasonal to Seasonal Prediction project database, we investigated deterministic and probabilistic multimodel ensemble (MME) forecasts of summer precipitation and heavy rainfall events in the middle–lower reaches of Yangtze River at lead times of 1–4 weeks. Evaluations indicate that MME forecast skill improves as the ensemble size increases. For a fixed ensemble size, the balanced full-model ensemble always outperforms single-model ensemble, while remaining comparable or even slightly inferior to the selectively constructed ensemble comprising only the best-performing models. This highlights the importance of strategic model weighting in MME construction. Two weighted MMEs—respectively, calibrated using ensemble model output statistics based on the censored and shifted gamma (CSG) distribution and the generalized extreme value (GEV) distribution—exhibit significantly enhanced skill compared to the equal-weighted MME, particularly in the first week. To address forecast skill degradation in week 2, we finally proposed a novel conditional MME approach and developed ENSO-conditioned weighted MMEs [ENSO-conditioned CSG (c-CSG) and ENSO-conditioned GEV (c-GEV)] through incorporating the observed preceding winter ENSO condition. The 3-yr independent forecast test and case study of a heavy rainfall event demonstrate that the ENSO-conditioned MMEs outperform the conventional MMEs, highlighting their potential to enhance subseasonal precipitation forecast capabilities. Significance Statement Skillful subseasonal forecasts of monsoon rainfall offer critical support for decision-making in agriculture, water resource management, and disaster preparedness. This study investigates deterministic and probabilistic multimodel ensemble (MME) forecasts of summer rainfall and heavy rainfall events in the middle–lower reaches of the Yangtze River at lead times of 1–4 weeks. We highlight the importance of strategic model weighting in MME construction and propose a novel ENSO-conditioned weighted MME approach, demonstrating its potential to improve subseasonal precipitation forecast.

Abstract Data preprocessing based on standardized anomaly (SA) has been shown to enhance spatial coherence and reduce computational costs by enabling statistical methods to postprocess forecasts from multiple stations simultaneously, rather than fitting the forecast system separately at each station, as is done with ensemble model output statistics. However, no studies have focused on SA combined with forest-based methods despite its distribution-free and highly flexible use for ensemble postprocessing. We propose Standardized Anomaly Quantile Regression Forest (SAQRF) for 10-m wind speed forecasts. The SAQRF employs a two-step approach involving SA preprocessing and spatial QRF modeling to capture underlying nonlinear features. QRF incorporates raw scale and static geographic predictors in a comparable setting, termed QRFGE. Similar settings are applied to the generalized random forest (GRF) named SAGRF and GRFGE. We compare four forest-based methods with competing statistical techniques, Standardized Anomaly Model Output Statistics (SAMOS) and Standardized Anomaly Boosting (SABST) approaches, using ECMWF’s high-resolution and ensemble forecasts from 2019 to 2020 in Hebei, China. The results show that forest-based methods achieve an average 5% improvement compared to SABST in the stationwise continuous ranked probability skill score.

Abstract Winter temperature in Northeast China significantly affects agriculture, energy use, and daily life, making accurate predictions is vital for social and economic planning. Prediction skills of the dynamical model for winter temperature prediction are different between warm and cold years, which has not been fully considered in previous multimodel ensemble predictions. Extreme gradient boosting (XGBoost), one of the popular deep learning methods, can use the gradient boosting to effectively handle large datasets and make it adept at capturing nonlinear temperature dynamics. This study utilizes three dynamical model predictions, along with on-site observation, and develops an XGBoost-Based Cold/Warm Year Ensemble Model (XEM). Validation results indicate that the XEM model shows different performances in winter temperature prediction among dynamical models. XEM can effectively improve accuracy and predictive skill of the winter temperature prediction comparing with the original model in Northeast China, with the average increase of anomaly correlation coefficient (ACC) by 0.49 and anomaly consistency score (PCS) by 21.1%. The XEM demonstrated remarkable enhancements in the independent sample validation in 2021 and 2022. The ACC achieved its maximum improvement of 0.85, while the PCS saw its highest enhancement at 45.9%, indicating a substantial leap in predictive performance. Additionally, the Shapley additive explanation (SHAP) value and gain analysis reveals that XEM can properly prioritize meteorological factors like 200-hPa geopotential height anomalies (H200) and 2-m air temperature (T2m) in its ensemble prediction, which provides robust insights on why XEM can effectively improve the winter predictions. Significance Statement This study introduces a pioneering extreme gradient boosting (XGBoost)-Based Cold/Warm Year Ensembled Model (XEM) model for predicting winter temperatures in Northeast China. By modeling cold and warm years separately and identifying key meteorological factors [such as H200 and zonal wind at 200 hPa (u200)], the model enhances prediction accuracy. It classifies the prediction year’s temperature attribute and ensembles results through weighted averaging. The approach achieves significant improvements in independent validation for 2021 and 2022, offering a robust and adaptable method for temperature prediction that could inform climate variability strategies in other regions.

The complexity and dimensionality of making deterministic photovoltaic power forecasts from ensemble numerical weather prediction

Advancing long-range daily precipitation ensemble forecasts with deep learning for Australia

CRPS-Net: A novel framework for AI-assisted meteo-marine ensemble forecasting

Abstract There has been an increasing interest in recent years in the use of machine learning (ML) algorithms such as random forests (RFs) in the context of severe weather prediction. However, there is a need to better understand the impacts of large-scale flow patterns on RF model performance to further improve their use of multiscale information. This work leverages real-time forecasts from convection-allowing Finite-Volume Cubed-Sphere Dynamical Core (FV3)-based ensemble forecasts produced by the University of Oklahoma Multiscale Data Assimilation and Predictability Laboratory during the Hazardous Weather Testbed (HWT) Spring Forecasting Experiments. Results show that by using an RF trained on all forecast cases, forecast cases that have relatively high importance for certain predictors have different discernible flow patterns compared to low-importance forecast cases for the same predictors through composite differences in different environmental variables. Maximum updraft helicity storm attribute predictors were associated with strong synoptic ascent, u 500 had compact shortwaves within northwesterly flow, and MUCAPE was associated with forcing displaced north of the region of severe weather. RF models trained on forecast cases with similar domain-averaged CAPE/shear were statistically more skillful than the baseline model trained irrespective of CAPE/shear patterns in forecasting the occurrence of severe weather. However, selecting training forecast cases based on spatial patterns or principal components of CAPE/shear using EOF analysis did not further improve the RF ability to forecast severe weather compared to the baseline model. The benefits of training on forecast cases based on domain-averaged CAPE/shear were maintained, and some of the benefits of training based on spatial patterns of CAPE/shear were maintained as the sample size increased.

Climate change will boost the invasion of the harmful cyanobacterium Raphidiopsis raciborskii in South America.

Ensemble forecasting is essential for quantifying forecast uncertainty and providing probabilistic weather predictions. However, the substantial computational demands of current global ensemble prediction systems based on conventional models limit ensemble sizes, hindering the representation of diverse weather scenarios. Recent advances in machine learning (ML) have greatly reduced computational costs and improved deterministic forecasting. Nonetheless, applying ML to ensemble forecasting poses challenges in addressing uncertainties in initial conditions and models, which are the major sources of forecasting errors. To address these challenges, we introduce FuXi-ENS, an advanced ML model that generates 6-hourly global ensemble weather forecasts up to 15 days ahead at a spatial resolution of 0.25°. Using a variational autoencoder framework, FuXi-ENS optimizes a loss function that combines the continuous ranked probability score (CRPS) with the Kullback-Leibler divergence, enabling flow-dependent perturbations. Comprehensive evaluations demonstrate that FuXi-ENS outperforms the ECMWF ensemble in key forecast metrics such as CRPS and Brier score.

Crude oil price forecasting is an important task in energy management and storage. In this regard, deep learning has been applied in the literature to generate accurate forecasts. The main purpose of this study is to design an ensemble prediction system based on various deep learning systems. Specifically, in the first stage of our proposed ensemble system, convolutional neural networks (CNNs), long short-term memory networks (LSTMs), bidirectional LSTM (BiLSTM), gated recurrent units (GRUs), bidirectional GRU (BiGRU), and deep feedforward neural networks (DFFNNs) are used as individual predictive systems to predict crude oil prices. Their respective parameters are fine-tuned by Bayesian optimization (BO). In the second stage, forecasts from the previous stage are all weighted by using the sequential least squares programming (SLSQP) algorithm. The standard tree-based ensemble models, namely, extreme gradient boosting (XGBoost) and random forest (RT), are implemented as baseline models. The main findings can be summarized as follows. First, the proposed ensemble system outperforms the individual CNN, LSTM, BiLSTM, GRU, BiGRU, and DFFNN. Second, it outperforms the standard XGBoost and RT models. Governments and policymakers can use these models to design more effective energy policies and better manage supply in fluctuating markets. For investors, improved predictions of price trends present opportunities for strategic investments, reducing risk while maximizing returns in the energy market.