Search NASA⌕ Search

SEARCH · Search NASA

Results for “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 19 records
All sourcesNASA NTRSDOE OSTI
DOE OSTI2025

Co-training of multiple neural networks for simultaneous optimization and training of physics-informed neural networks for composite curing

This paper introduces a Physics-Informed Neural Network (PINN) technique that co-trains neural networks (NNs) that represent each function in a system of equations to simultaneously solve equations representing an out-of-autoclave (OOA) cure process while conducting optimization in adherence to process requirements. Specifically, this co-training approach benefits from using NNs to represent OOA inputs (air temperature profile) and outputs (part and tool temperature profiles and degree of cure). Production requirements can then be levied on the inputs, such as maximum air temperature and minimum cure cycle, and simultaneously on the outputs, such as degree of cure, maximum part temperature, and part temperature rate limits. The technique is validated with finite element (FE) simulations and physical experiments for curing a Toray T830H-6 K/3900-2D composite panel. Furthermore, this novel approach efficiently models and optimizes the OOA cure process.

Composite curing↗
DOE OSTI2025

Cyber-Informed Engineering (CIE) Workbook: End-of-Train (EoT) / Head-of-Train (HoT) Communications

This workbook presents a vulnerability (CVE-2025-1727 ) found in train applications and guides a digital risk assessment and mitigation analysis and application of Cyber-Informed Engineering principles to mitigate the potential consequences and ultimately the hazard through the engineering discipline because of exploiting this vulnerability. Workshop participants are encouraged to use the workbook to capture insights and lessons learned. The workbook guides the participant to: • Understand the HE communication vulnerability • Map digital threats to physical consequences • Use bowtie analysis to illustrate both “security” and “engineering” barriers • Apply CIE principles to ensure that even if communications are compromised, the physical engineered system still behaves safely. • Produce an actionable set of engineered and infosec controls for implementation

42 - ENGINEERING↗
DOE OSTI2025

2025 Peregrine in-situ monitoring and training dataset for laser powder bed fusion and binder jet printers

Peregrine, a software tool developed at Oak Ridge National Laboratory (ORNL), was used to collect and analyze in-situ monitoring (ISM) data from a Concept Laser M2 (Colibrium Additive) laser powder bed fusion (L-PBF) printer and an ExOne Innovent (Desktop Metal) binder jet printer. Data for four builds (print jobs) were saved to HDF5 (high performance data) files for release. Additionally, process anomalies were annotated by the authors across 37 image stacks (i.e., print layers) and are also provided as HDF5 files.

36 MATERIALS SCIENCE↗
DOE OSTI2025

NFPA Distributed Energy Resources Safety Training (DERST) For Emergency Responders

The National Fire Protection Association, with support from the Department of Energy, executed a multi-year initiative to develop, enhance, and disseminate Distributed Energy Resources Safety Training (DERST) tools for U.S. emergency responders. As Distributed Energy Resources (DER)—such as solar photovoltaics, battery energy storage systems (ESS), electric vehicles (EVs), and associated infrastructure—become increasingly prevalent, the NFPA identified a critical need for up-to-date standardized, accessible, and effective safety training tailored for the fire service and related public safety professionals. The project delivered a comprehensive suite of educational resources to improve responders’ abilities to safely manage DER-related incidents. This included: • Revised Modular Training Courses: Updated classroom-based DER safety courses, now modular and accessible nationwide through fire academies and the North American Fire Training Directors (NAFTD) network. • Live Burn Testing & Research: A full-scale controlled burn of a DER-equipped residential structure provided real-world data and insights, forming the basis for updated best practices. • A Gamified Simulation Tool – Firefighters Incident Response Simulation Tool (FIRST): A first-of-its-kind, multiplayer, scenario-based simulation using the Unreal Engine 5.0 to train responders in a realistic virtual, multi-DER incident environment. • Field Familiarization Software Tools & Prop Guide: Digital DER field familiarization evolutions software guide and a prop development manual to support field-based DER training exercises, enhancing responders' hands-on familiarity with DER infrastructure and collaboration on virtual incident responses. • National Dissemination Strategy: Strategic partnerships with NAFTD, Vector Solutions, and others enabled wide-scale distribution, with over 5,000 departments accessing resources and 1,100+ departments adopting the simulator in the first seven months. Also provided a web portal for easy access to all training and simulation programs developed under this grant for the U.S. responder community. Key findings from the project—particularly from the burn test—led to paradigm shifts in fire response tactics. For example, traditional approaches to garage fires may be hazardous if DERs are present, due to explosive off gassing and thermal runaway risks. The new training emphasizes scene assessment, stand-off approaches, thermal imaging verification, and careful post-incident cooling of DER components to prevent reignition. This initiative has had a significant national impact, raising awareness, enhancing preparedness, and supporting safer DER incident response practices. Significant engagement from the media, public safety organizations, and PBS coverage has further amplified the reach and adoption of NFPA’s DER safety training, tools, and simulations.

14 SOLAR ENERGY↗
DOE OSTI2025

Fair Concurrent Training of Multiple Models in Federated Learning

Federated learning (FL) enables collaborative learning across multiple clients. In most FL work, all clients train a single learning task. However, the recent proliferation of FL applications may increasingly require multiple FL tasks to be trained simultaneously, sharing clients’ computing resources, which we call Multiple-Model Federated Learning (MMFL). Current MMFL algorithms use naïve average-based client-task allocation schemes that often lead to unfair performance when FL tasks have heterogeneous difficulty levels, as the more difficult tasks may need more client participation to train effectively. Furthermore, in the MMFL setting, we face a further challenge that some clients may prefer training specific tasks to others, and may not even be willing to train other tasks, e.g., due to high computational costs, which may exacerbate unfairness in training outcomes across tasks. We address both challenges by firstly designing FedFairMMFL, a difficulty-aware algorithm that dynamically allocates clients to tasks in each training round, based on the tasks’ current performance levels. We provide guarantees on the resulting task fairness and FedFairMMFL’s convergence rate. We then propose novel auction designs that incentivizes clients to train multiple tasks, so as to fairly distribute clients’ training efforts across the tasks, and extend our convergence guarantees to this setting. Here, we finally evaluate our algorithm with multiple sets of learning tasks on real world datasets, showing that our algorithm improves fairness by improving the final model accuracy and convergence speed of the worst performing tasks, while maintaining the average accuracy across tasks.

Federated learning↗
DOE OSTI2024

An Active Learning-Based Streaming Pipeline for Reduced Data Training of Structure Finding Models in Neutron Diffractometry

Structure determination workloads in neutron diffractometry are computationally expensive and routinely require several hours to many days to determine the structure of a material from its neutron diffraction patterns. The potential for machine learning models trained on simulated neutron scattering patterns to significantly speed up these tasks have been reported recently. However, the amount of simulated data needed to train these models grows exponentially with the number of structural parameters to be predicted and poses a significant computational challenge. To overcome this challenge, we introduce a novel batch-mode active learning (AL) policy that uses uncertainty sampling to simulate training data drawn from a probability distribution that prefers labelled examples about which the model is least certain. We confirm its efficacy in training the same models with ∼ 75% less training data while improving the accuracy. We then discuss the design of an efficient stream-based training workflow that uses this AL policy and present a performance study on two heterogeneous platforms to demonstrate that, compared with a conventional training workflow, the streaming workflow delivers ∼ 20% shorter training time without any loss of accuracy.

Wang, Tianle [Brookhaven National Laboratory (BNL)↗
DOE OSTI2025

Reliable Determination of Pulses and Pulse-Shape Instability in Ultrashort Laser Pulse Trains Using Polarization-Gating and Transient-Grating Frequency-Resolved Optical Gating Using the RANA Approach

Devices that measure the presence of instability in the pulse shapes in trains of ultrashort laser pulses do not exist, so this task necessarily falls to pulse-measurement devices, like Frequency-Resolved Optical Gating (FROG) and its variations, which have proven to be a highly reliable class of techniques for measuring stable trains of ultrashort laser pulses. Fortunately, multi-shot versions of FROG have also been shown to sensitively distinguish trains of stable from those of unstable pulse shapes by displaying readily visible systematic discrepancies between the measured and retrieved traces in the presence of unstable pulse trains. However, the effects of pulse-shape instability and algorithm stagnation can be indistinguishable, so a never-stagnating algorithm—even when instability is present—is required and is generally important. In previous work, we demonstrated that our recently introduced Retrieved-Amplitude N-grid Algorithmic (RANA) approach produces highly reliable (100%) pulse-retrieval in the second-harmonic-generation (SHG) version of FROG for thousands of sample trains of pulses with stable pulse shapes. Further, it does so even for trains of unstable pulse shapes and thus both reliably distinguishes between the two cases and provides a rough measure of the degree of instability as well as a reasonable estimate of most typical pulse parameters. Here, we perform the analogous study for the polarization-gating (PG) and transient-grating (TG) versions of FROG, which are often used for higher-energy pulse trains. We conclude that PG and TG FROG, coupled with the RANA approach, also provide reliable indicators of pulse-shape instability. In addition, for PG and TG FROG, the RANA approach provides an even better estimate of a typical pulse in an unstable pulse train than SHG FROG does, even in cases of significant pulse-shape instability.

47 OTHER INSTRUMENTATION↗
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↗
DOE OSTI2026

Custom-trained Machine-learning Interatomic Potentials: ZnCl2 Aqueous Solution

This dataset was generated using an iterative active-learning strategy implemented in the ArcaNN software package (https://github.com/arcann-chem/arcann_training) to train machine-learning interatomic potentials for aqueous ZnCl2 solutions. Each active-learning cycle consisted of three stages: training, exploration, and labeling. The initial training set combined configurations generated in this work from enhanced-sampling ab initio molecular dynamics simulations with configurations from a previously reported neural-network-potential study of aqueous ZnCl2. The enhanced-sampling ab initio molecular dynamics simulations involved Zn–Cl separation and the chloride coordination number around Zn²? as collective variables. These configurations served as the seed dataset. Subsequent active-learning cycles expanded the training set by identifying and labeling configurations that were poorly represented by the current models, thereby improving coverage of ion-association states and changes in local coordination and charge-state environments relevant to the solution free-energy landscape. For all selected configurations, single-point calculations of the total energies and atomic forces were performed within density functional theory using the CP2K Quickstep module. Reference calculations employed the revPBE-D3 and r2SCAN exchange-correlation functionals. Motivated by recent work on aqueous Zn²?, the main revPBE calculations omitted D3 dispersion contributions involving Zn²?, while retaining the D3 correction for water and chloride. For comparison, fully dispersion-corrected revPBE-D3 reference calculations were also performed, with D3 applied to all species, including Zn²?. Valence electrons were treated explicitly, while core electrons were represented using norm-conserving Goedecker–Teter–Hutter pseudopotentials. The wave functions were expanded using the mixed Gaussian-and-plane-wave scheme with TZV2P-MOLOPT basis sets for all elements and a 600 Ry auxiliary plane-wave cutoff for the electron density. Self-consistent-field convergence was accelerated using the orbital-transformation and Direct Inversion in the Iterative Subspace algorithms, with a convergence threshold of 10?6. All single-point calculations were performed in periodic orthorhombic cells. The CELL_REF keyword in CP2K was used to define a fixed reference cell with a box length of 25 Å. This treatment ensured a consistent reference for configurations extracted from NpT trajectories with fluctuating cell dimensions. The resulting DFT energies and atomic forces constitute the ground-truth labels used to train the MLIPs. The resulting MLIP was trained for aqueous ZnCl2 solutions spanning concentrations from 0 to 30 molal and a broad pH range, from strongly acidic to strongly basic conditions. Representative examples of configurations included in the MLIP training dataset are provided below. These include 1) Representative configurations from the dataset labeled at the revPBE-D3 level, with D3 dispersion interactions involving Zn2+ excluded (revPBE-wo-D3). 2) Representative configurations from the dataset labeled at the fully dispersion-corrected revPBE-D3 level, with D3 interactions applied to all species, including Zn2+ (revPBE-D3). 3) Representative configurations from the dataset labeled at the r2SCAN level of theory (r2SCAN).

Dinpajooh, Mohammadhasan [Pacific Northwest Nation↗
DOE OSTI2025

Training Evaluation: A Case Study of Measuring Impact and ROI

The purpose of this case study was to evaluate the Compliance Officer/Coordinator (CO/CC) training program using Phillips’ approach to training evaluation. INL Learning Performance Consultants (LPCs) applied the ROI Methodology to measure the effectiveness of the CO/CC training program. INL CO/CCs participated in focus groups to provide feedback on their learning, enablers and barriers to applying what they learned, and opportunities to improve the CO/CC learning experience. With collaboration from the CO/CC training program owner, the LPCs demonstrated the business impact and return on investment (ROI) of the CO/CC training program. Included are training program recommendations and lessons learned using the ROI Methodology.

99 GENERAL AND MISCELLANEOUS↗
DOE OSTI2026

Temporal multi-omic analysis uncovers sex-biased molecular programs underlying skeletal muscle adaptation to endurance training

Background. Exercise training is known to benefit health and reduce disease risk. While adaptations in skeletal muscles are fundamental to many of the health benefits of exercise training, the common and sex-specific molecular regulators that mediate these adaptations remain to be fully elucidated. Methods. To this end, we leveraged skeletal muscle multi-omics data generated by the Molecular Transducers of Physical Activity Consortium (MoTrPAC), where 6 month-old male and female rats endurance trained for 1, 2, 4, or 8 weeks. Our objective was to identify shared and sex-specific multi-omic molecular responses to endurance training in skeletal muscle, and relate them to phenotypic adaptations. Results. We identified largely sexually-conserved transcriptomic and proteomic enrichments in the gastrocnemius, which correlated with skeletal muscle responses from a published exercise study in humans. We uncovered sex-consistent post-translational modifications, including decreased oxidation of MYH2 and deacetylation of the ß-oxidation enzyme HADHA. Pathway enrichment analyses revealed sex-specific remodeling across the acetylome, redox proteome, and phosphoproteome; females decreased mitochondrial protein oxidation and increased mitochondrial cristae proteins, indicative of enhanced redox buffering and mitochondrial efficiency. Despite observed decreases in the oxidation of key mitochondrial proteins, females displayed increases in the oxidation of proteins involved in glucose catabolism relative to males after 8 weeks of training, suggestive of sex-biased subcellular reactive oxygen species generation. Conclusions. This work shows a large portion of the adaptive response to endurance training in skeletal muscle is shared between females and males, while there are distinct and nuanced sex-specific adaptations that are evident, particularly at the level of post-translational regulation.

Many, Gina M.↗
DOE OSTI2025

Similarity Metric for Data Optimization and Efficient Training of Reactive Machine Learning Force Fields for Hydrocarbon Radiolysis

Radiolysis is a common approach to sterilize polymers, chemically modify them for upcycling, and accelerate their decomposition for recycling purposes. Reactive molecular dynamics (MD) simulations provide a powerful tool to generate atomic-level trajectories of the reactive processes and quantify radiolytic chemical degradation pathways. For this, machine learning (ML) surrogate models for reactive force fields with quantum mechanical accuracy are now widely used, which require ML training data sets that can provide information on atomic environments for target chemical systems. However, radiolysis chemistry can be highly complex and diverse, which poses significant challenges for generating training data to parametrize ML models. In this regard, we developed a method for optimizing the training data set using a cosine similarity metric to help guide training set selection for radiolysis of polyethylene, a model hydrocarbon polymer, as well as to enhance the transferability of our reactive ML force field (MLFF) to a variety of molecular and polymeric systems. Our approach performs atom-by-atom comparisons between local atomic environments to pinpoint important data points associated with rare and localized events, such as radiolysis damage within structures. We apply this approach to train the Chebyshev Interaction Model for Efficient Simulation (ChIMES) MLFF model, which expresses the atomic interaction potentials in terms of linear combinations of many-body Chebyshev polynomials. We first show that our method can reduce our training set size by ∼70% while improving overall accuracy compared to more standard MD model fitting approaches. We then validate our optimum model against diverse hydrocarbon simulation data, including simple alkanes and systems with unsaturated carbon bonds, over a wide range of thermodynamic conditions. Finally, we use our ChIMES model to perform MD simulations of radiolytic damage with large-scale systems that help avoid system size effects. Overall, our approach yields an MD force field that retains most of the accuracy of the underlying quantum method while yielding many orders of improvement in computational efficiency. In conclusion, our efforts will have impact on future hydrocarbon polymer radiolysis studies, where the chemical details of the polymer–radiation interactions can have a strong effect on the resulting products observed in experiments.

Hydrocarbons↗
DOE OSTI2026

Prediction and Experimental Verification of Electrolyte Solvation Structure from an OMol25-Trained Interatomic Potential

A molecular-level understanding of electrolyte solvation structure and ion–ion correlations is critical to developing next-generation battery chemistries. Atomistic simulation capabilities with sufficient accuracy, speed, and transferability to deliver reliable structural insights while avoiding arduous system-specific reparameterization are thus highly desirable. Machine learning interatomic potentials (MLIPs) trained on large, chemically diverse data sets are revolutionizing computational chemistry, enabling molecular dynamics simulations of battery electrolytes with near-DFT accuracy over 10,000× faster than DFT. While previous MLIP training data sets with suitable elemental coverage for electrolytes have been based on inorganic materials, the Open Molecules 2025 (OMol25) data set provides large-scale molecular DFT MLIP training data with broad elemental coverage and specifically samples tens of millions of electrolyte configurations. Here, we integrate computational modeling with experimental validation to systematically assess the ability of large-scale MLIPs pretrained on materials data or on OMol25 to accurately resolve nanoscale structural organization and ion-solvation characteristics in Na-ion battery electrolytes across diverse physicochemical conditions and compositional regimes. We find that the OMol25-trained Universal Model of Atoms (UMA-OMol) predicts experimentally measured densities and X-ray structure factors in substantially better agreement compared to state-of-the-art models trained only on inorganic materials data. Using UMA-OMol, we further analyze systematic trends in solvation structure as a function of cation identity, anion chemistry, salt concentration, and solvent topology. We observe that increasing system temperature amplifies the heterogeneity within the solvation environment, perturbing cation–solvent interactions and promoting the formation of contact ion pairs (CIPs). Moreover, subtle variations in the solvent topology of glyme-based electrolytes cause pronounced changes in ion correlations and solvation structure. The experimental agreement and microscopic insights shown here position OMol25-trained MLIPs as a practical route to predictive, high-throughput electrolyte simulations beyond the limits of classical force fields and direct DFT molecular dynamics, serving as a powerful tool for accelerating the design of next-generation Na-ion battery electrolytes and beyond.

MLIPs↗
DOE OSTI2025

Stable Simulation of the Community Atmosphere Model Using Machine‐Learning Physical Parameterization Trained With Experience Replay

In recent years, machine learning (ML) models have been used to improve physical parameterizations of general circulation models (GCMs). A significant challenge of integrating ML models into GCMs is the online instability when they are coupled for long‐term simulation. We present a new strategy that demonstrates robust online stability when the physical parameterization package of an atmospheric GCM is replaced by a deep ML model. The method uses experience replay with a multistep training scheme of the ML model in which the model's own output at the previous time step is used in the training. Predicted physics tendencies in the replay buffer with the most recent errors in the training iterations are reused, making the ML model learn from its own errors. The training method reduces the gap between the offline and online environments of the ML model. The method is used to train the ML model as the physical parameterization of the Community Atmosphere Model (CAM5) with training data from the Multi‐scale Modeling Framework high resolution simulations. Three 6‐year online simulations of the CAM5 are carried out by using the ML physics package. The simulated spatial distributions of precipitation, surface temperature and zonally averaged atmospheric fields demonstrate overall better accuracy than that of the standard CAM5 and benchmark model even without the use of additional physical constraints or tuning. This work is the first to demonstrate a solution to address the online instability problem in climate modeling with ML physics by using experience replay.

54 ENVIRONMENTAL SCIENCES↗
DOE OSTI2026

Pre-training Vision Models for the Classification of Alerts from Wide-field Time-domain Surveys

Modern wide-field time-domain surveys facilitate the study of transient, variable and moving phenomena by conducting image differencing and relaying alerts to their communities. Machine learning tools have been used on data from these surveys and their precursors for more than a decade, and convolutional neural networks (CNNs), which make predictions directly from input images, saw particularly broad adoption through the 2010s. Since then, continually rapid advances in computer vision have transformed the standard practices around using such models. It is now commonplace to use standardized architectures pre-trained on large corpora of everyday images (e.g., ImageNet). In contrast, time-domain astronomy studies still typically design custom CNN architectures and train them from scratch. Here, we explore the effects of adopting various pre-training regimens and standardized model architectures on the performance of alert classification. We find that the resulting models match or outperform a custom, specialized CNN like what is typically used for filtering alerts. Moreover, our results show that pre-training on galaxy images from Galaxy Zoo tends to yield better performance than pre-training on ImageNet or training from scratch. We observe that the design of standardized architectures are much better optimized than the custom CNN baseline, requiring significantly less time and memory for inference despite having more trainable parameters. On the eve of the Legacy Survey of Space and Time and other image-differencing surveys, these findings advocate for a paradigm shift in the creation of vision models for alerts, demonstrating that greater performance and efficiency, in time and in data, can be achieved by adopting the latest practices from the computer vision field.

79 ASTRONOMY AND ASTROPHYSICS↗
DOE OSTI2026

Tula: Optimizing Time, Cost, and Generalization in Distributed Large-Batch Training

Distributed training increases the number of batches processed per iteration either by scaling-out (adding more nodes) or scaling-up (increasing the batch-size). However, the largest configuration does not necessarily yield the best performance. Horizontal scaling introduces additional communication overhead, while vertical scaling is constrained by computation cost and device memory limits. Thus, simply increasing the batch-size leads to diminishing returns: training time and cost decrease initially but eventually plateaus, creating a knee-point in the time/cost vs. batch-size pareto curve. The optimal batch-size therefore depends on the underlying model, data and available compute resources. Large batches also suffer from worse model quality due to the well-known “generalization gap”. In this paper, we present Tula, an online service that automatically optimizes time, cost, and convergence quality for large-batch training of convolutional models. It combines parallel-systems modeling with statistical performance prediction to identify the optimal batchsize. Tula predicts training time and cost within 7.5−14% error across multiple models, and achieves up to 20× overall speedup and improves test accuracy by ≈9% on average over standard large-batch training on various vision tasks, thus successfully mitigating the generalization gap and accelerating training at the same time.

Tyagi, Sahil [ORNL] (ORCID:0009000783144745)↗
DOE OSTI2025

A Parallel Alternative for Energy-Efficient Neural Network Training and Inferencing

Energy efficiency of training and inferencing with large neural network models is a critical challenge facing the future of sustainable large-scale machine learning workloads. This paper introduces an alternative strategy, called phantom parallelism, to minimize the net energy consumption of traditional tensor (model) parallelism, the most energy-inefficient component of large neural network training. The approach is presented in the context of feed-forward network architectures as a preliminary, but comprehensive, proof-of-principle study of the proposed methodology. We derive new forward and backward propagation operators for phantom parallelism, implement them as custom autograd operations within an end-to-end phantom parallel training pipeline and compare its parallel performance and energy-efficiency against those of conventional tensor parallel training pipelines. Formal analyses that predict lower bandwidth and FLOP counts are presented with supporting empirical results on up to 256 GPUs that corroborate these gains. Experiments are shown to deliver ∼50% reduction in the energy consumed to train FFNs using the proposed phantom parallel approach when compared with conventional tensor parallel methods. Additionally, the proposed approach is shown to train smaller phantom models to the same model loss on smaller GPU counts as larger tensor parallel models on larger GPU counts offering the possibility for even greater energy savings.

Seal, Sudip [ORNL] (ORCID:0000000332330656)↗
DOE OSTI2025

Efficient Training of Deep Neural Operator Networks via Randomized Sampling

Neural operators (NOs) employ deep neural networks to learn the mappings between infinitedimensional function spaces. Deep operator network (DeepONet), a popular NO architecture, has demonstrated success in the real-time prediction of complex dynamics across various scientific and engineering applications. In this work, we introduce a random sampling technique to be adopted during the training of DeepONet, aimed at improving the generalization ability of the model, while significantly reducing the computational time. The proposed approach targets the trunk network of the DeepONet model that outputs the basis functions corresponding to the spatiotemporal locations of the bounded domain on which the physical system is defined. While constructing the loss function, DeepONet training traditionally considers a uniform grid of spatiotemporal points at which all the output functions are evaluated for each iteration. This approach leads to a larger batch size, resulting in poor generalization and increased memory demands, due to the limitations of the stochastic gradient descent (SGD) optimizer. The proposed random sampling over the inputs of the trunk net mitigates these challenges, improving generalization and reducing the memory requirements during training, resulting in significant computational gains. We validate our hypothesis through three benchmark examples, demonstrating substantial reductions in training time while achieving comparable or lower overall test errors relative to the traditional training approach. Our results indicate that incorporating randomization in the trunk network inputs during training enhances the efficiency and robustness of DeepONet, offering a promising avenue for improving the framework’s performance in modeling complex physical systems.

Karumuri, Sharmila [Department of Civil & Systems ↗