Search NASA⌕ Search

SEARCH · Search NASA

Results for “successive training”

Search indexed NASA NTRS and DOE OSTI research on propulsion, heat transfer, battery materials and energy systems. Follow report and document links to the original sources.

Quote a phrase for an exact phrase match. Source license links do not imply unrestricted reuse.

At least 145 records · Page 8
All sourcesNASA NTRSDOE OSTI
DOE OSTI2025

Advancing the Frontiers of Deep Learning for Low-Dose 3D Cone-Beam CT Reconstruction

X-ray computed tomography (CT) is an important noninvasive medical imaging modality for studying the structural details of internal organs. Image reconstruction in CT is an inverse problem of recovering an object's internal structure from the absorption profile of X-ray beams (sinogram) measured using a detector. The classical variational approach for CT reconstruction minimizes an energy functional using an appropriate iterative algorithm. Motivated by the success of deep learning (DL), researchers have begun to leverage training data and enhanced computing capabilities in recent years to produce high-fidelity reconstructed images. Nonetheless, much of the academic research in DL algorithms for CT has focused primarily on the two-dimensional setting (with simplified forward operators and noise model) for proofs-of-concept, and a comprehensive benchmarking of various classical and data-driven CT reconstruction approaches has not beenundertaken. The key objective of our CT reconstruction grand challenge was to promote methodological advancements for both classical and DL-based approaches for clinical CT with a reasonably accurately simulated 3D CT forward operator and noise model. We have utilized the publicly available LIDC-IDRI dataset and simulated sinograms and FDK images corresponding to two dose levels (clinical- and low-dose, constituting two tracks of the challenge) starting from the normal-dose images as the ground truth. In this paper, we summarize the motivation, context, and results of our challenge, and highlight the future research directions in DL for clinical CT.

X-ray tomography↗
DOE OSTI2022

Estimating Cosmological Constraints from Galaxy Cluster Abundance using Simulation-Based Inference

Inferring the values and uncertainties of cosmological parameters in a cosmology model is of paramount importance for modern cosmic observations. In this paper, we use the simulation-based inference (SBI) approach to estimate cosmological constraints from a simplified galaxy cluster observation analysis. Using data generated from the Quijote simulation suite and analytical models, we train a machine learning algorithm to learn the probability function between cosmological parameters and the possible galaxy cluster observables. The posterior distribution of the cosmological parameters at a given observation is then obtained by sampling the predictions from the trained algorithm. Our results show that the SBI method can successfully recover the truth values of the cosmological parameters within the 2σ limit for this simplified galaxy cluster analysis, and acquires similar posterior constraints obtained with a likelihood-based Markov Chain Monte Carlo method, the current state-of the-art method used in similar cosmological studies.

46 INSTRUMENTATION RELATED TO NUCLEAR SCIENCE AND ↗
DOE OSTI2020

Autotuning of Double-Dot Devices In Situ with Machine Learning

The current practice of manually tuning quantum dots (QDs) for qubit operation is a relatively time-consuming procedure that is inherently impractical for scaling up and applications. In this work, we report on the in situ implementation of a recently proposed autotuning protocol that combines machine learning (ML) with an optimization routine to navigate the parameter space. In particular, we show that a ML algorithm trained using exclusively simulated data to quantitatively classify the state of a double-QD device can be used to replace human heuristics in the tuning of gate voltages in real devices. Here, we demonstrate active feedback of a functional double-dot device operated at millikelvin temperatures and discuss success rates as a function of the initial conditions and the device performance. Modifications to the training network, fitness function, and optimizer are discussed as a path toward further improvement in the success rate when starting both near and far detuned from the target double-dot range.

71 CLASSICAL AND QUANTUM MECHANICS, GENERAL PHYSIC↗
DOE OSTI2021

Lightweight LSTM for CAN Signal Decoding

This paper describes an approach to identify undecoded Controller Area Network (CAN) data from one vehicle, based on the data similarity to previously decoded CAN data from another vehicle. Modern vehicles communicate data and signals from on-board sensors and controllers through the CAN bus. Networked sensors contain information such as wheel speeds, fuel gauges, turn signals, and radar signals. In the effort to use this information and make cars safer through human-in-the-loop CPS, signals on the CAN bus such as wheel speed and radar can be used to support the driver. However, data from the CAN bus are encoded and in some cases compressed, and different car manufacturers use different encoding schemes to represent data on the CAN bus. With hundreds of messages and thousands of possible encoding schemes to consider, it is laborious to identify the unique bits and encoding schemes that represent signals on each vehicle. In this study, we propose a method for training a Long Short-Term Memory (LSTM) neural network on known radar signals from one vehicle manufacturer, a Toyota, and successfully apply the network to identify the encoding for radar signals on a different vehicle, a Honda. By augmenting the training dataset with varied encoding bit boundaries, a small and lightweight LSTM network can learn to recognize radar data across different encoding schemes. The results are an improvement on exhaustive-search algorithms and other methods previously used in the search for such signals.

Ngo, Paul↗
DOE OSTI2022

Hybrid Quantum-Classical Neural Networks

Deep learning is one of the most successful and far-reaching strategies used in machine learning today. However, the scale and utility of neural networks is still greatly limited by the current hardware used to train them. These concerns have become increasingly pressing as conventional computers are soon expected to approach the physical limitations that will slow their performance improvements in the near future. For these reasons, scientists have begun to explore alternative computing platforms, like quantum computers, for training neural networks. In recent years, variational quantum circuits have emerged as one of the most successful approaches to quantum deep learning on noisy intermediate scale quantum devices. We propose a hybrid quantum-classical neural network architecture where each neuron is a variational quantum circuit. We empirically analyze the performance of this hybrid neural network on a series of binary classification data sets using a simulated IBM universal quantum computer and a state-of-the-art IBM universal quantum computer. On the simulated hardware, we observe that the hybrid neural network achieves around 10% higher classification accuracy and 20% better minimization of the cost function than an individual variational quantum circuit. On the quantum hardware, we observe that each model only performs well when the qubit and gate count is sufficiently small.

Arthur, Davis↗
DOE OSTI2020

Combining mechanistic and machine learning models for predictive engineering and optimization of tryptophan metabolism

Abstract Through advanced mechanistic modeling and the generation of large high-quality datasets, machine learning is becoming an integral part of understanding and engineering living systems. Here we show that mechanistic and machine learning models can be combined to enable accurate genotype-to-phenotype predictions. We use a genome-scale model to pinpoint engineering targets, efficient library construction of metabolic pathway designs, and high-throughput biosensor-enabled screening for training diverse machine learning algorithms. From a single data-generation cycle, this enables successful forward engineering of complex aromatic amino acid metabolism in yeast, with the best machine learning-guided design recommendations improving tryptophan titer and productivity by up to 74 and 43%, respectively, compared to the best designs used for algorithm training. Thus, this study highlights the power of combining mechanistic and machine learning models to effectively direct metabolic engineering efforts.

59 BASIC BIOLOGICAL SCIENCES↗
DOE OSTI2025

Model-based Hierarchical Reinforcement Learning for Improved Physical Security Design: A Prototype

Prior work in FY24 developed an adversarial AI agent aid in path analysis of physical protection systems. This agent, trained using a model-based reinforcement learning algorithm, was able to successfully learn the most vulnerable path in facilities. It was able to extend the current state of practice for physical protection design by exhibiting dynamic behavior based on current environmental conditions. Whereas PathTrace largely performs a static, graph-based analysis, the AI agent was able to make decisions based on relative position in the facility, current conditions (was the adversarial agnet discovered?), and proximity to secondary targets. The agent demonstrated some novel capabilities, but had limitations that need to be resolved before it can be used for production purposes. For example, the adversarial agent generalizes poorly and takes a relatively long time to train. Nonetheless, there is still considerable promise for developing the adversarial agent further in order to explore even richer, more dynamic behaviors (e.g., adversary motivations, environmental debris, and more). This work considers a complementary idea; development of a planning agent. The planning agent is envisioned as an auto-complete-like tool that can help accelerate security system design by human experts. The agent would respect existing barriers and sensors placed by a human expert while offering cost-effective suggestions (i.e., implicitly balancing effectiveness with cost) to improve the design. The goal is for this agent to be part of an expert’s toolbox, not to totally upend the current state-of-practice, or to displace human experts. The ultimate goal would be concurrent training of both the adversarial and planning agent together, to learn entirely through self-play. This would represent an entirely new way of performing system deign. We selected a hierarchical, model-based reinforcement learning algorithm to serve as the planning agent. This is an extension of concepts used in the prior FY24 adversarial agent work. There, we had a single agent acting an environment. Here, we have two different sub-agents (policies), working together, to form a complete agent. There is a manager policy, which can select abstract goals on slower time scales, and a worker, which performs primitive actions to reach goals selected by the manager. It is worth noting that this class of algorithm is challenging to work with. From our understanding, our work is one of the first successful uses of model-based reinforcement learning (MBRL) in nuclear energy1 , and likely the first hierarchical model-based reinforcement learning application in nuclear energy. Further, this work is one of the first known attempts to apply AI to perform a design tasks in nuclear energy. Consequently, there were significant implementation challenges and the bulk of the work was focused on successful implementation and algorithm design. The results presented here are very low technology readiness level as a consequence of the lack of related literature, but still represent a significant step forward in the pursuit of applied AI for design.

42 ENGINEERING↗
DOE OSTI2020

Rapid Identification of X-ray Diffraction Patterns Based on Very Limited Data by Interpretable Convolutional Neural Networks

Large volumes of data from material characterizations call for rapid and automatic data analysis to accelerate materials discovery. Herein, we report a convolutional neural network (CNN) that was trained based on theoretical data and very limited experimental data for fast identification of experimental X-ray diffraction (XRD) patterns of metal–organic frameworks (MOFs). To augment the data for training the model, noise was extracted from experimental data and shuffled; then it was merged with the main peaks that were extracted from theoretical spectra to synthesize new spectra. For the first time, one-to-one material identification was achieved. Theoretical MOFs patterns (1012) were augmented to a whole data set of 72 864 samples. It was then randomly shuffled and split into training (58 292 samples) and validation (14 572 samples) data sets at a ratio of 4:1. For the task of discriminating, the optimized model showed the highest identification accuracy of 96.7% for the top 5 ranking on a test data set of 30 hold-out samples. Neighborhood component analysis (NCA) on the experimental XRD samples shows that the samples from the same material are clustered in groups in the NCA map. Analysis on the class activation maps of the last CNN layer further discloses the mechanism by which the CNN model successfully identifies individual MOFs from the XRD patterns. Furthermore, this CNN model trained by the data augmentation technique would not only open numerous potential applications for identifying XRD patterns for different materials, but also pave avenues to autonomously analyze data by other characterization tools such as FTIR, Raman, and NMR spectroscopies.

36 MATERIALS SCIENCE↗
DOE OSTI2022

Detecting Large Explosions With Machine Learning Models Trained on Synthetic Infrasound Data

Explosions produce low-frequency acoustic (infrasound) waves capable of propagating globally, but the spatio-temporal variability of the atmosphere makes detecting events difficult. Machine learning (ML) is well-suited to identify the subtle and nonlinear patterns in explosion infrasound signals, but a previous lack of ground-truth data inhibited training of generalized models. We introduce a physics-based method that propagates infrasound sources through realistic atmospheres to create 28,000 synthetic events, which are used to train ML classifiers. A simple artificial neural network and modern temporal convolutional network discriminate synthetic events from background noise with >90% accuracy and, more importantly, successfully identify the majority of real-world explosion signals recorded during the Humming Road Runner experiment. ML models trained entirely on physics-based synthetics advance explosion detection capabilities and make ML more viable to related fields lacking training data.

45 MILITARY TECHNOLOGY, WEAPONRY, AND NATIONAL DEF↗
DOE OSTI2025

Towards automated and real-time multi-object detection of anguilliform fishes from sonar data using YOLOv8 deep learning algorithm

Eels (Anguilla spp.), including American eels (Anguilla rostrata), European eels (Anguilla anguilla), and Japanese eels (Anguilla japonica), are species of critical management and regulatory concern due to their vulnerability to various stressors during downstream migrations. Accurate and efficient detection of migrating eels can improve our understanding of fish behaviors and fish-hydraulic structure interactions, thereby facilitating the design, operation, and optimization of more effective downstream passage facilities from both biological and economic perspectives. However, a real-time, automated framework for detecting migrating eels in real-world applications is currently lacking. Leveraging imaging sonar as a reliable technology for fish passage monitoring, field data are acquired using imaging sonar and then converted to single sonar frames/images for subsequent analysis. In this study, a framework based on the You Only Look Once Version 8 (YOLOv8)-based convolutional neural network is proposed for multi-object detection of eels and non-eel fish using the sonar images after image subtraction and additional wavelet denoising. The results from both training and testing phases demonstrate that the framework's ability can successfully detect both eels and non-eel fish in preprocessed sonar images, achieving F1-scores and mAP@0.50 exceeding 0.84. Additionally, the incorporation of wavelet denoising during preprocessing slightly improve detection performance. Furthermore, the transferability of this framework from eel to lamprey detection is demonstrated to be feasible given the similar morphological characteristics of these two species. Overall, the proposed framework achieves accurate and efficient detection of migrating eels, providing reliable and real-time information that can help conserve vulnerable eel and eel-like populations.

Deep learning↗
DOE OSTI2020

Automation of Active Space Selection for Multireference Methods via Machine Learning on Chemical Bond Dissociation

Predicting and understanding the chemical bond is one of the major challenges of computational quantum chemistry. Kohn–Sham density functional theory (KS-DFT) is the most common method, but approximate density functionals may not be able to describe systems where multiple electronic configurations are equally important. Multiconfigurational wave functions, on the other hand, can provide a detailed understanding of the electronic structures and chemical bonds of such systems. In the complete active space self-consistent field (CASSCF) method, one performs a full configuration interaction calculation in an active space consisting of active electrons and active orbitals. However, CASSCF and its variants require the selection of these active spaces. This choice is not black box; it requires significant experience and testing by the user, and thus active space methods are not considered particularly user-friendly and are employed only by a minority of quantum chemists. Our goal is to popularize these methods by making it easier to make good active space choices. Here, we present a machine learning protocol that performs an automated selection of active spaces for chemical bond dissociation calculations of main group diatomic molecules. The protocol shows high prediction performance for a given target system as long as a properly correlated system is chosen for training. Good active spaces are correctly predicted with a considerably better success rate than random guess (larger than 80% precision for most systems studied). Our automated machine learning protocol shows that a “black-box” mode is possible for facilitating and accelerating the large-scale calculations on multireference systems where single-reference methods such as KS-DFT cannot be applied.

37 INORGANIC, ORGANIC, PHYSICAL, AND ANALYTICAL CH↗
DOE OSTI2021

Identifying Genomic Islands with Deep Neural Networks

Background Horizontal gene transfer is the main source of adaptability for bacteria, through which genes are obtained from different sources including bacteria, archaea, viruses, and eukaryotes. This process promotes the rapid spread of genetic information across lineages, typically in the form of clusters of genes referred to as genomic islands (GIs). Different types of GIs exist, and are often classified by the content of their cargo genes or their means of integration and mobility. While various computational methods have been devised to detect different types of GIs, no single method is capable of detecting all types. Results We propose a method, which we call Shutter Island, that uses a deep learning model (Inception V3, widely used in computer vision) to detect genomic islands. The intrinsic value of deep learning methods lies in their ability to generalize. Via a technique called transfer learning, the model is pre-trained on a large generic dataset and then re-trained on images that we generate to represent genomic fragments. We demonstrate that this image-based approach generalizes better than the existing tools. Conclusions We used a deep neural network and an image-based approach to detect the most out of the correct GI predictions made by other tools, in addition to making novel GI predictions. The fact that the deep neural network was re-trained on only a limited number of GI datasets and then successfully generalized indicates that this approach could be applied to other problems in the field where data is still lacking or hard to curate.

Computer Vision↗
DOE OSTI2023

What Machine Learning Can and Cannot Do for Inertial Confinement Fusion

Machine learning methodologies have played remarkable roles in solving complex systems with large data, well-defined input–output pairs, and clearly definable goals and metrics. The methodologies are effective in image analysis, classification, and systems without long chains of logic. Recently, machine-learning methodologies have been widely applied to inertial confinement fusion (ICF) capsules and the design optimization of OMEGA (Omega Laser Facility) capsule implosion and NIF (National Ignition Facility) ignition capsules, leading to significant progress. As machine learning is being increasingly applied, concerns arise regarding its capabilities and limitations in the context of ICF. ICF is a complicated physical system that relies on physics knowledge and human judgment to guide machine learning. Additionally, the experimental database for ICF ignition is not large enough to provide credible training data. Most researchers in the field of ICF use simulations, or a mix of simulations and experimental results, instead of real data to train machine learning models and related tools. They then use the trained learning model to predict future events. This methodology can be successful, subject to a careful choice of data and simulations. However, because of the extreme sensitivity of the neutron yield to the input implosion parameters, physics-guided machine learning for ICF is extremely important and necessary, especially when the database is small, the uncertain-domain knowledge is large, and the physical capabilities of the learning models are still being developed. In this work, we identify problems in ICF that are suitable for machine learning and circumstances where machine learning is less likely to be successful. This study investigates the applications of machine learning and highlights fundamental research challenges and directions associated with machine learning in ICF.

70 PLASMA PHYSICS AND FUSION TECHNOLOGY↗
DOE OSTI2020

Development of Remote Cutting Tools for use at ATR - 20438

As a result of years of operations of the Advanced Test Reactor (ATR) at the Idaho National Lab (INL), the ATR canal has become congested due to the storage of materials including waste consisting of highly irradiated remote handled (RH) metals. In order to support its continuing nuclear mission the ATR requires canal space to support fuel storage and future core internal change out (CIC) components. INL staff determined that the canal must be cleaned out and waste removed to efficiently utilize the valuable canal floor space. It was determined that a methodology for processing, packaging, characterizing, and removing current and future irradiated hardware from the ATR canal is needed. Orano Federal Services (OFS) and its subcontractor Babcock Services Inc. (BSI) were selected to provide remote tooling in order to safely and efficiently size reduce the material currently stored in the ATR canal. INL developed detailed requirements that established the design and performance requirements for the tooling. Orano TN and BSI had recently successfully completed similar tasks in clearing highly activated materials from spent fuel pools and canals at commercial nuclear plants. That experience directly translated into a proven approach that could be used at the ATR thereby saving time and money and lowering risks. Based on BSI's demonstrated proficiency in the design and deployment of remote size reduction specialty tooling at comparable commercial facilities, Orano Federal Services was confident that a solution to INL's requirements could be delivered. The tooling had to be designed to be compatible with the operational requirements of the ATR canal as well as the functionality needed by the facility operators that will use the equipment. Some of the major design requirements and functions of the equipment included remote operations, capability to shear highly irradiated sections of aluminum and steel tubing, pipe, and components ranging in size from 0.6 cm to 16.5 cm diameter into 1.5 m -1.8 m lengths, utilizing buoyancy compensation to assist with underwater operations, and use commercially available existing technology and components as much as practical. The design team evaluated and selected commercially available equipment that could be modified to meet the design and performance specifications. In addition, a waste sizing table (WST) was designed and fabricated to further ensure the size reduction equipment would perform as expected and give the facility operators a stable and usable remote, underwater work platform. After successful development of the tooling, a mock up test and operator training was conducted to ensure the tooling performed as expected and met the requirements of INL. (authors)

12 MANAGEMENT OF RADIOACTIVE AND NON-RADIOACTIVE W↗
DOE OSTI2022

Impact of Timing in the Design Process on Students’ Application of Design for Additive Manufacturing Heuristics

The goal of this work is to study the way student designers use design for additive manufacturing (DfAM) rules or heuristics. It can be challenging for novice designers to create successful designs for additive manufacturing (AM), due to its recent surge in popularity and lack of formal education or training. A study was carried out to investigate the way novices apply DfAM heuristics when they receive them at different points in the design process. A design problem was presented to students, and three different groups of student participants were given a lecture on DfAM heuristics at three different points in the design process: before the initial design, between the initial design and redesign, and after the redesign. The novelty and quality of each of the resulting designs were evaluated. Results indicate that although the DfAM heuristics lecture had no impact on the overall quality of the designs generated, participants who were given the heuristics lecture after the initial design session produced designs that were better-suited for 3D printing in the second phase of the design activity. However, receiving this additional information appears to prevent students from creatively iterating upon their initial designs, as participants who received heuristic information between the design sessions experienced a decrease in novelty between the two sessions. Additionally, receiving the heuristics lecture increased all students’ perceptions of their ability to perform DfAM-related tasks. Furthermore, these results validate the practicality of design heuristics in lecture form as AM training tools while also emphasizing the importance of iteration in the design process.

computer-aided design↗
DOE OSTI2021

Advanced Energy Design Guide for K-12 School Buildings (AEDG) Training: Cooperative Research and Development Final Report, CRADA Number CRD-18-00761

U.S. Department of Energy (DOE), Office of Energy Efficiency and Renewable Energy (EERE), Building Technologies Office (BTO), Commercial Buildings Integration (CBI) Program supports the adoption of zero energy ready buildings design practices through targeted Science, Technology, Engineering, and Mathematics (STEM) workforce development. The National Renewable Energy Laboratory (NREL) identified the opportunity to provided targeted training to those professionals that have the greatest potential to impact adoption and successful design outcomes of zero energy ready buildings projects. In January 2018, ASHRAE released Achieving Zero Energy: Advanced Energy Design Guide for K-12 School Buildings (AEDG). The AEDG was developed in collaboration with NREL, American Institute of Architects (AIA), American Society of Heating, Refrigeration and Air-conditioning Engineers (ASHRAE), U.S. Green Building Council (USGBC), and Illuminating Engineers Society (IES). The purpose of this CRADA was to create a partnership between NREL, AIA, ASHRAE, USGBC, and IES to create and host on-demand, web-based training based on the AEDG.

32 ENERGY CONSERVATION, CONSUMPTION, AND UTILIZATI↗
DOE OSTI2021

Where’s Swimmy?: Mining unique color features buried in galaxies by deep anomaly detection using Subaru Hyper Suprime-Cam data

Abstract We present the Swimmy (Subaru WIde-field Machine-learning anoMalY) survey program, a deep-learning-based search for unique sources using multicolored (grizy) imaging data from the Hyper Suprime-Cam Subaru Strategic Program (HSC-SSP). This program aims to detect unexpected, novel, and rare populations and phenomena, by utilizing the deep imaging data acquired from the wide-field coverage of the HSC-SSP. This article, as the first paper in the Swimmy series, describes an anomaly detection technique to select unique populations as “outliers” from the data-set. The model was tested with known extreme emission-line galaxies (XELGs) and quasars, which consequently confirmed that the proposed method successfully selected $\sim\!\! 60\%$–$70\%$ of the quasars and $60\%$ of the XELGs without labeled training data. In reference to the spectral information of local galaxies at z = 0.05–0.2 obtained from the Sloan Digital Sky Survey, we investigated the physical properties of the selected anomalies and compared them based on the significance of their outlier values. The results revealed that XELGs constitute notable fractions of the most anomalous galaxies, and certain galaxies manifest unique morphological features. In summary, deep anomaly detection is an effective tool that can search rare objects, and, ultimately, unknown unknowns with large data-sets. Further development of the proposed model and selection process can promote the practical applications required to achieve specific scientific goals.

Astronomy & Astrophysics↗
DOE OSTI2025

Analyzing and Exploring Training Recipes for Large-Scale Transformer-Based Weather Prediction

Abstract The rapid rise of deep learning (DL) in numerical weather prediction (NWP) has led to a proliferation of models which forecast atmospheric variables with comparable or superior skill than traditional physics-based NWP. However, among these leading DL models, there is a wide variance in both the training settings and architecture used. Further, the lack of thorough ablation studies makes it hard to discern which components are most critical to success. In this work, we show that it is possible to attain high forecast skill even with relatively off-the-shelf architectures, simple training procedures, and moderate compute budgets. Specifically, we train a minimally modified Swin Transformer V2 (SwinV2) on ERA5 data and find that it attains superior skill in terms of mean-square errors of deterministic forecasts when compared against the European Centre for Medium-Range Weather Forecasts’ Integrated Forecasting System (IFS). Almost all DL–NWP systems share a core set of hyperparameters and design decisions. To aid and expedite future DL–NWP research, we present an in-depth, systematic exploration of different loss functions, model sizes and depths, patch sizes, and multistep training objectives. We also examine the model performance with metrics beyond the typical accuracy (ACC) and RMSE and investigate how the performance scales with model size. Through our open-source code, scoring pipelines, and models, we share our findings on key aspects of the training pipeline. These ablations reduce the necessity for expensive hyperparameter tuning and lower the barrier to entry for future DL–NWP research. Significance Statement This study investigates the potential of using large-scale transformer-based models for weather prediction, showing that it is possible to achieve high forecast accuracy with simpler, off-the-shelf architectures. By training a minimally modified SwinV2 transformer on ERA5 data, we show that the model achieves competitive forecast skill in terms of mean-square error for key variables, outperforming the European Centre for Medium-Range Weather Forecasts’ Integrated Forecasting System (IFS) at all lead times. Our findings suggest that effective training strategies, such as multistep fine-tuning and channel-weighted losses, significantly enhance the model’s performance. However, we also highlight that these improvements come with trade-offs in other areas, such as ensemble spread and high-frequency spatial detail. This work highlights the promise of deep learning in improving weather forecasts, which could lead to better preparedness and response to weather events, ultimately benefiting society by providing more reliable weather predictions.

Willard, Jared D. [Lawrence Berkeley National Labo↗