Search NASA⌕ Search

SEARCH · Search NASA

Results for “training efficiency”

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 235 records · Page 13
All sourcesNASA NTRSDOE OSTI
DOE OSTI2022

Accelerating Collective Communication in Data Parallel Training across Deep Learning Frameworks

This work develops new techniques within Horovod, a generic communication library supporting data parallel training across deep learning frameworks. In particular, we improve the Horovod control plane by implementing a new coordination scheme that takes advantage of the characteristics of the typical data parallel training paradigm, namely the repeated execution of collectives on the gradients of a fixed set of tensors. Using a caching strategy, we execute Horovod’s existing coordinator-worker logic only once during a typical training run, replacing it with a more efficient decentralized orchestration strategy using the cached data and a global intersection of a bitvector for the remaining training duration. Next, we introduce a feature for end users to explicitly group collective operations, enabling finer grained control over the communication buffer sizes. To evaluate our proposed strategies, we conduct experiments on a world-class supercomputer — Summit. We compare our proposals to Horovod’s original design and observe 2x performance improvement at a scale of 6000 GPUs; we also compare them against tf.distribute and torch.DDP and achieve 12% better and comparable performance, respectively, using up to 1536 GPUs; we compare our solution against BytePS in typical HPC settings and achieve about 20% better performance on a scale of 768 GPUs. Finally, we test our strategies on a scientific application (STEMDL) using up to 27,600 GPUs (the entire Summit) and show that we achieve a near-linear scaling of 0.93 with a sustained performance of 1.54 exaflops (with standard error +- 0.02) in FP16 precision.

Romero, Joshua↗
DOE OSTI2025

Triangle Method for Dense ReLU Layers [SWR-25-72]

This software is an implementation of the methods for initializing and training neural networks to be more efficient per parameter, described more fully below and in the related publication: In theory, depth should make a ReLU network EXPONENTIALLY more efficient by enabling it to produce an exponential number of piecewise linear sections in its output. This reasoning is largely based on the work of mathematicians that have hand-constructed networks that make good use of depth. In practice however, even very deep ReLU networks that have been randomly initialized will behave identically to their shallow counterparts - missing an entire exponential dimension of efficiency. The triangle method is a first attempt at realizing the exponential potential of deep networks. Instead of randomly setting weights, we force pairs of neurons in each layer learn to build triangles (i.e. functions from [0,1] -> [0,1] that look like triangles). This is a very efficient pattern for generating lots of linear pieces because composing two triangular functions doubles the number of pieces with each composition. The triangle method is more than just a different initialization, it is a new paradigm of training. Instead of making direct updates to the matrix weights, we do an extra step of backpropagation to collect the derivatives of the loss function with respect to the shapes of the triangles, training them to tilt left or right. This process essentially holds the networks hand throughout the loss landscape and forces it to always use depth effectively by producing triangular shapes internally. This can produce several orders of magnitude of improvement on convex one-dimensional regression problems. Much more theoretical work is needed to realize its full potential beyond this context, but the implementation in this repository will still work in arbitrary numbers of dimensions. The file Triangle_Method.py is a generalized form of the method that will build each neuron its own custom 1-d convex activation function (with exponential efficiency). Example usage on one dimensional problems can be found in Example_Usage.ipynb and an example of using this in a real neural network can be found in Example_VGG16_CIFAR10.ipynb.

Milkert, Max [National Renewable Energy Laboratory↗
DOE OSTI2025

A Universal Augmentation Framework for Long-Range Electrostatics in Machine Learning Interatomic Potentials

Most current machine learning interatomic potentials (MLIPs) rely on short-range approximations, without explicit treatment of long-range electrostatics. To address this, we recently developed the Latent Ewald Summation (LES) method, which infers electrostatic interactions, polarization, and Born effective charges (BECs), just by learning from energy and force training data. Here, in this study, we present LES as a standalone library, compatible with any short-range MLIP, and demonstrate its integration with methods such as MACE, NequIP, Allegro, CACE, CHGNet, and UMA. We benchmark LES-enhanced models on distinct systems, including bulk water, polar dipeptides, and gold dimer adsorption on defective substrates, and show that LES not only captures correct electrostatics but also improves accuracy. Additionally, we scale LES to large and chemically diverse data by training MACELES-OFF on the SPICE set containing molecules and clusters, making a universal MLIP with electrostatics for organic systems, including biomolecules. MACELES-OFF is more accurate than its short-range counterpart (MACE-OFF) trained on the same data set, predicts dipoles and BECs reliably, and has better descriptions of bulk liquids. By enabling efficient long-range electrostatics without directly training on electrical properties, LES paves the way for electrostatic foundation MLIPs.

Kim, Dongjin [University of California, Berkeley, ↗
DOE OSTI2024

Kansas City, Missouri, Streetlight Electric Vehicle Charging: Strategies and challenges for site selection of streetlight electric vehicle infrastructure in Kansas City, Missouri (Final Report)

Public streetlight charging, whether on streets in central business districts or residential areas, provides easy charging access for apartment residents and homeowners alike. While most electric vehicle (EV) drivers charge at home, they do so in garages or on driveways they own. For renters and residents of multifamily housing (MFH), however, this may not be an option. EVs have a lower cost of ownership compared to conventional vehicles, and a used EV may be an affordable option for a lower-income household. But without easy access to charging, even a low-cost used EV may not be an option for a prospective buyer. An affordable curbside charging network has the potential to expand EV adoption into neighborhoods that have to date seen minimal interest and uptake of the technology and associated charging infrastructure. Streetlight charging networks can provide an economical, scalable, and effective approach to providing equitable and convenient charging. Metropolitan Energy Center (MEC) is dedicated to the mission of creating resource efficiency, environmental health, and economic vitality in the Kansas City region and beyond. Since 1983, MEC has provided resources, outreach, and training to make alternative fuels and energy efficiency commonplace. MEC led a streetlight charging pilot project that installed limited EV charging infrastructure on the streetlight system in Kansas City, Missouri, to demonstrate and test the benefits of curbside charging for EVs at existing on-street parking locations. The project aimed to cost-effectively expand the charging network in Kansas City to support residential charging and provide infrastructure in one or more charging deserts throughout the city. This pilot evaluates the impact and overall success of streetlight charging based on community feedback, utilization of charging infrastructure, technical feasibility, and cost. The project has pursued a data- and community-driven site selection process designed to identify sites with high demand and high opportunity for EV charging. This project was funded by the U.S. Department of Energy (DOE) and awarded to MEC through a competitive proposal process. The novelty and complexity of this project required an organization that could facilitate collaboration across levels of government, community members, and industry partners. For the past 25 years, through Kansas City Regional Clean Cities, MEC has worked with numerous public and private fleets on a variety of projects to improve the environmental performance and efficiency of the regional vehicle fleet. To advance affordable, efficient, and clean transportation efforts, DOE Clean Cities and Communities coalitions create local networks of public and private sector stakeholders and engage communities. Rooted within their local communities, the coalitions serve as experts and ambassadors, bringing to bear the collective knowledge, experience, and practical know-how of the entire network from within DOE, its national laboratories, and diverse stakeholders in the field. MEC and its project partners made in-kind contributions to leverage federal dollars for the benefit of the Kansas City community. Findings from this project will help determine the best applications for streetlight charging technologies to maximize funding impact and serve community needs. The team evaluated locations based on expected charging demand, technical feasibility, safety considerations, and enhanced charging network siting needs. Throughout the project, the team gathered feedback and evaluated ways to make public charging for EVs available to all community members. The insights will help Kansas City and other communities streamline future efforts to support EV drivers through public charging in the city right-of-way. Furthermore, this project will inform citywide guidance for future installations. MEC is committed to a transparent and publicly accessible approach that encourages the collaborative evaluation of streetlight charging. The project has engaged the community to proactively identify and evaluate the benefits and impacts of streetlight charging. It was a priority for the project to ensure the benefits of this pilot are distributed equitably to all members of the Kansas City community and that new charging opportunities and associated resources are available in diverse neighborhoods across the city. The charging infrastructure supports an affordable curbside charging network that will enable more drivers to choose EVs and provide easy charging access for all community members interested in driving an EV. The community feedback received through this project informed future resources and opportunities to make EVs more accessible to all members of the Kansas City community. MEC worked with several community partners on this project, including Missouri University of Science and Technology (MST), Pennsylvania State University (Penn State), the National Renewable Energy Laboratory (NREL); the city of Kansas City, Missouri; Evergy; Black and McDonald (B&M); LilyPad EV; EVNoire; and Westside Housing Organization (WHO). Project partners contributed to the cost match required for DOE grants through capital expenditures, personnel, and other in-kind contributions. Detailed descriptions of project team organizations can be found in Appendix A. Project Partners. Analysts at NREL and MST/PennState developed site maps based on demand and equity considerations. MEC conducted outreach to community members to garner input on project design and site selection, and received approval from the Missouri Public Service Commission (PSC) for Evergy’s EV charging station ownership. MEC worked with all partners to gather additional siting criteria and developed a site selection evaluation checklist, and partners conducted site visits to proposed installation sites. Next, B&M, Evergy, and the city executed all site agreements, conducted site-specific engineering design, acquired associated permits, and issued notices to proceed site by site or in small batches. Finally, from January to April 2023, the project team installed 23 EV charging stations built on Kansas City’s streetlight system in six council districts. Evergy will own, operate, and monitor the stations for 10 years, sharing charging data with MEC for at least 1 year.

32 ENERGY CONSERVATION, CONSUMPTION, AND UTILIZATI↗
DOE OSTI2020

Deep Learning-Based Queue-Aware Eco-Approach and Departure System for Plug-In Hybrid Electric Buses at Signalized Intersections: A Simulation Study

Eco-Approach and Departure (EAD) has been considered as a promising eco-driving strategy for vehicles traveling in an urban environment, where information such as signal phase and timing (SPaT) and geometric intersection description is well utilized to guide vehicles passing through intersections in the most energy-efficient manner. Previous studies formulated the optimal trajectory planning problem as finding the shortest path on a graphical model. While this method is effective in terms of energy saving, its computation efficiency can be further enhanced by adopting machine learning techniques. In this paper, we propose an innovative deep learning-based queue-aware eco-approach and departure (DLQ-EAD) system for a plug-in hybrid electric bus (PHEB), which is able to provide an online optimal trajectory for the vehicle considering both the downstream traffic condition (i.e. traffic lights, queues) and the vehicle powertrain efficiency. Based on optimal solutions obtained from the graph-based trajectory planning algorithm (GTPA), a deep neural network (DNN) is developed to learn the optimal vehicle speed for the next time step given its current state. It is demonstrated that the trained DNN can provide energy-efficient trajectories with high computational efficiency and high flexibility adopting to dynamic changes in the surrounding environment. To address the impact of downstream traffic, a queue prediction model is further developed using data from radars and connected vehicles (CVs), as well as signal timing data from SPaT messages. A comprehensive simulation study in the microscopic traffic modeling software PTV VISSIM shows that the proposed DLQ-EAD can achieve 18.7%-24.0% energy efficiency improvements for a single PHEB on various traffic congestion levels. The proposed queue prediction model can be of practical significance even at low penetration rates of CVs. Specifically, additional energy savings of 2.0%-8.2% can be further achieved with 20% vehicles in the network.

Ye, Fei↗
DOE OSTI2025

Operation of helium sub-atmospheric multistage cryogenic centrifugal compressor trains: Part 1 – Steady state modeling and speed ratio selection

Helium cryogenic systems which can provide cooling below the normal boiling point of helium (approximately 4.2 K) are often required by superconducting radio-frequency niobium resonators utilized in modern high-energy particle accelerators. Achieving temperatures below 4.2 K generally involves operating a cryogenic vessel with liquid helium under sub-atmospheric conditions, thereby lowering the saturation pressure and corresponding saturation temperature. Over the last several decades, multi-stage cryogenic centrifugal compressor trains (CC’s) have been operated efficiently and reliably within large-scale cryogenic systems to continuously evacuate helium vapor generated by a device within the vessel, maintaining sub-atmospheric conditions in the vessel while pressurizing the return vapor to above atmospheric conditions. Traditionally, these CC systems have been operated using empirically derived control philosophies and insight gathered from previous operational experience. Recent efforts at the Facility for Rare Isotope Beams (FRIB) have been aimed at the development of a theoretical basis to characterize the operation of multi-stage cryogenic centrifugal compressor train and utilizing predictive model results to generate control parameters. The objective of this research was identifying operational points which adequately balance cryogenic system efficiency, stability, and overall ease of operation. Furthermore, this manuscript provides an overview of the predictive model development, characterization of the FRIB cryogenic centrifugal compressors and implementation of the predicted performance results during steady-state system operation.

Compressor train control↗
DOE OSTI2022

Accelerated computation of lattice thermal conductivity using neural network interatomic potentials

We report with the development of the density functional theory (DFT) and ever-increasing computational capacity, an accurate prediction of lattice thermal conductivity based on the Boltzmann transport theory becomes computationally feasible, contributing to a fundamental understanding of thermal conductivity as well as a choice of the optimal materials for specific applications. However, steep costs in evaluating third-order force constants limit the theoretical investigation to crystals with high symmetry and few atoms in the unit cell. Currently, machine learning potentials are garnering attention as a computationally efficient high-fidelity model of DFT, and several studies have demonstrated that the lattice thermal conductivity could be computed accurately via the machine learning potentials. However, test materials were mostly crystals with high symmetries, and the applicability of machine learning potentials to a wide range of materials has yet to be demonstrated. Furthermore, establishing a standard training set that provides consistent accuracy and computational efficiencies across a wide range of materials would be useful. To address these issues, herein we compute lattice thermal conductivities at 300 K using neural network interatomic potentials. As test materials, we select 25 materials with diverse symmetries and a wide range of lattice thermal conductivities between 10 -1 and 10 3 Wm -1 K -1 . Among various choices of training sets, we find that molecular dynamics trajectories at 50–700 K consistently provide results at par with DFT for the test materials. In contrast to pure DFT approaches, the computational cost in the present approach is uniform over the test materials, yielding a speed gain of 2–10 folds. When a smaller reduced training set is used, the relative efficiency increases by up to ~50 folds without sacrificing accuracy significantly. The current work will broaden the application scope of machine learning potentials by establishing a robust framework for accurately computing lattice thermal conductivity with machine learning potentials.

36 MATERIALS SCIENCE↗
DOE OSTI2022

Industrial Assessment Center – integration of education and practice. Final progress report

The goal of the DOE’s Industrial Assessment Center (IAC) program is twofold: first, to help US manufacturing competitiveness by providing assessments and recommendations for small and medium-sized enterprises (SMEs) on energy efficiency, productivity, sustainability and competitiveness – including measuring the impacts of these recommendations on reducing greenhouse gas emissions; and second, to address a growing shortage of engineering professionals with applied energy and manufacturing-related skills by training a diverse cross-section of engineering students through hands-on involvement in these assessments. IUPUI has an IAC, which was established in 2011, in the Department of Mechanical and Energy Engineering, Purdue School of Engineering and Technology, IUPUI. The IAC has won the awards twice. We have built the center that has the expertise and infrastructure for quality energy assessments to manufactures and commercial building owners and aligned our current research and teaching activities to the DOE’ training mission with programs to train the next generation of industrial energy efficiency experts, with theory and hands-on experience in conducting energy assessment for small and median manufacturing enterprises. We have made recommendations that have the potential to save about $20M annually. We also initiated research projects to meet DOE priorities, specifically in the areas of smart manufacturing and cybersecurity. On the energy assessment side, the Center has provided energy assessments to 144 qualified companies, primarily located in Indiana. On the training side, the center has trained students in various programs, such as the department’s Bachelor of Science (BS) programs in Mechanical Engineering (ME) and BS in Energy Engineering (EEN), both are ABET accredited engineering programs, as well as a graduate level Energy Management and Assessment certificate program. At the present, the center has trained 95 students, 43 or them received the DOE issued certificates. Research projects were developed within the center to advance energy efficiency related technologies, which provided excellent opportunities for our trainees. The center has also been building a professional network with utilities, Manufacturing Extension Partnership (MEP), government agencies, and manufacturing companies, to increase our client base and promote collaborations. The center received the 2019 Center of Excellence of the Year Award and three students received the “Outstanding Achievement in Energy Engineering by an IAC Student” awards in the past five years.

42 ENGINEERING↗
DOE OSTI2025

Sequence length scaling in vision transformers for scientific images on frontier

Vision Transformers (ViTs) are pivotal for foundational models in scientific imagery, including Earth science applications, due to their capability to process large sequence lengths. While transformers for text have inspired scaling sequence lengths in ViTs, adapting these for ViTs introduces unique challenges. We develop distributed sequence parallelism for ViTs, enabling them to handle up to 1M tokens. Our approach, leveraging DeepSpeed-Ulysses and Long-Sequence-Segmentation with model sharding, is the first to apply sequence parallelism in ViT training, achieving a 94% batch scaling efficiency on 2,048 AMD-MI250X GPUs. Evaluating sequence parallelism in ViTs, particularly in models up to 10B parameters, highlighted substantial bottlenecks. We countered these with hybrid sequence, pipeline, and flash attention strategies, to scale beyond single GPU memory limits. Our method significantly enhances climate modeling accuracy by 20% in temperature predictions, marking the first training of a vision transformer model to convergence with a sequence length of 188K tokens, using full self-attention.

Tsaris, Aristeidis (aris) [ORNL] (ORCID:0000000277↗
DOE OSTI2023

Differentiable modelling to unify machine learning and physical models for geosciences

Process-based modelling offers interpretability and physical consistency in many domains of geosciences but struggles to leverage large datasets efficiently. Machine-learning methods, especially deep networks, have strong predictive skills yet are unable to answer specific scientific questions. Here, in this Perspective, we explore differentiable modelling as a pathway to dissolve the perceived barrier between process-based modelling and machine learning in the geosciences and demonstrate its potential with examples from hydrological modelling. ‘Differentiable’ refers to accurately and efficiently calculating gradients with respect to model variables or parameters, enabling the discovery of high-dimensional unknown relationships. Differentiable modelling involves connecting (flexible amounts of) prior physical knowledge to neural networks, pushing the boundary of physics-informed machine learning. It offers better interpretability, generalizability, and extrapolation capabilities than purely data-driven machine learning, achieving a similar level of accuracy while requiring less training data. Additionally, the performance and efficiency of differentiable models scale well with increasing data volumes. Under data-scarce scenarios, differentiable models have outperformed machine-learning models in producing short-term dynamics and decadal-scale trends owing to the imposed physical constraints. Differentiable modelling approaches are primed to enable geoscientists to ask questions, test hypotheses, and discover unrecognized physical relationships. Future work should address computational challenges, reduce uncertainty, and verify the physical significance of outputs.

58 GEOSCIENCES↗
DOE OSTI2025

Materials Graph Library (MatGL), an open-source graph deep learning library for materials science and chemistry

Graph deep learning models, which incorporate a natural inductive bias for atomic structures, are of immense interest in materials science and chemistry. Here, we introduce the Materials Graph Library (MatGL), an open-source graph deep learning library for materials science and chemistry. Built on top of the popular Deep Graph Library (DGL) and Python Materials Genomics (Pymatgen) packages, MatGL is designed to be an extensible “batteries-included” library for developing advanced model architectures for materials property predictions and interatomic potentials. At present, MatGL has efficient implementations for both invariant and equivariant graph deep learning models, including the Materials 3-body Graph Network (M3GNet), MatErials Graph Network (MEGNet), Crystal Hamiltonian Graph Network (CHGNet), TensorNet and SO3Net architectures. MatGL also provides several pre-trained foundation potentials (FPs) with coverage of the entire periodic table, and property prediction models for out-of-box usage, benchmarking and fine-tuning. Finally, MatGL integrates with PyTorch Lightning to enable efficient model training.

chemistry↗
DOE OSTI2024

Reinforcement learning applied to dilute combustion control for increased fuel efficiency

To reduce the modeling burden for control of spark-ignition engines, reinforcement learning (RL) has been applied to solve the dilute combustion limit problem. Q-learning was used to identify an optimal control policy to adjust the fuel injection quantity in each combustion cycle. A physics-based model was used to determine the relevant states of the system used for training the control policy in a data-efficient manner. The cost function was chosen such that high cycle-to-cycle variability (CCV) at the dilute limit was minimized while maintaining stoichiometric combustion as much as possible. Experimental results demonstrated a reduction of CCV after the training period with slightly lean combustion, contributing to a net increase in fuel conversion efficiency of 1.33%. To ensure stoichiometric combustion for three-way catalyst compatibility, a second feedback loop based on an exhaust oxygen sensor was incorporated into the fuel quantity controller using a slow proportional-integral (PI) controller. The closed-loop experiments showed that both feedback loops can cooperate effectively, maintaining stoichiometric combustion while reducing combustion CCV and increasing fuel conversion efficiency by 1.09%. Finally, a modified cost function was proposed to ensure stoichiometric combustion with a single controller. In addition, the learning period was shortened by half to evaluate the RL algorithm performance on limited training time. Experimental results showed that the modified cost function could achieve the desired CCV targets, however, the learning time was reduced by half and the fuel conversion efficiency increased only by 0.30%.

33 ADVANCED PROPULSION SYSTEMS↗
DOE OSTI2022

High-Accuracy Semiempirical Quantum Models Based on a Minimal Training Set

A great need exists for computationally efficient quantum simulation approaches that can achieve an accuracy similar to high-level theories at a fraction of the computational cost. In this regard, we have leveraged a machine-learned interaction potential based on Chebyshev polynomials to improve density functional tight binding (DFTB) models for organic materials. The benefit of our approach is two-fold: (1) many-body interactions can be corrected for in a systematic and rapidly tunable process, and (2) high-level quantum accuracy for a broad range of compounds can be achieved with ~0.3% of data required for one advanced deep learning potential. Our model exhibits both transferability and extensibility through comparison to quantum chemical results for organic clusters, solid carbon phases, and molecular crystal phase stability rankings. Overall, our efforts thus allow for high-throughput physical and chemical predictions with up to coupled-cluster accuracy for systems that are computationally intractable with standard approaches.

36 MATERIALS SCIENCE↗
DOE OSTI2022

A Machine Learning Method for the Forensics Attribution of Separated Plutonium

Plutonium (Pu) source attribution would be a powerful tool to support nuclear nonproliferation efforts. This capability to find the source of a Pu sample would act as a deterrent to smuggling efforts, and also help regulatory agencies verify declared nuclear activities. Work at Texas A&M University yielded a nuclear forensics methodology, which is capable of determining separated Pu’s reactor of origin, fuel burnup, and the time since irradiation (TSI)—three parameters of interest. The methodology used a set of ten intra-element isotopic ratios found in separated Pu, which was compared to a library of isotopic ratio values produced using neutronics simulations for reactors of interest. By calculating the probability that unknown Pu sample’s isotopic ratio set matched a set in the library, the methodology could predict the three parameters of interest of the sample. One shortcoming of this methodology was an inability to correctly attribute spoofed Pu, where Pu sourced from two different reactors or two different fuel burnup levels are mixed. A new methodology to rectify this vulnerability using machine learning (ML) technique is developed, instead of the maximum likelihood calculation previously used and the results are satisfactory. The ML approach leverages the existing simulated data for training the algorithm, but use them efficiently by only using intra-element isotope ratios that contribute to the attribution one of the three parameters at a time. Previously, all isotope ratios were used to attribute all three parameters together. The new methodology attributes the Pu parameters in three steps, one for each parameter, rather than resolving all of the three parameters simultaneously like the previous maximum likelihood approach. First, a support vector machine classifier with a set of seven isotopic ratios finds the reactor of origin and a set of regression models trained using gaussian process predicts the burnup with a different set of seven isotopes. Finally, TSI is calculated analytically using decay equations. Thus far, the new methodology is capable of attributing pure Pu samples and has been validated using experimental data. The next step will to be augment the classifier training data set with spoofed Pu data.

73 NUCLEAR PHYSICS AND RADIATION PHYSICS↗
DOE OSTI2025

CAFE AU LAIT: Compute-Aware Federated Augmented Low-Rank AI Training

Federated finetuning is crucial for unlocking the knowledge embedded in pretrained Large Language Models (LLMs) when data are geographically distributed across clients. Unlike finetuning with data from a single institution, federated finetuning allows collaboration across multiple institutions, enabling the utilization of diverse and decentralized datasets while preserving data privacy. Given the high computing costs of LLM training and the emphasis on energy efficiency in Federated Learning (FL), Low-Rank Adaptation (LoRA) has emerged as a widely adopted algorithm due to its significantly reduced number of trainable parameters. However, this assumes that all data silos have the necessary computing resources to compute local updates of LLMs. Nevertheless, in practice, the computing resources across clients are highly heterogeneous: while some may have access to hundreds of GPUs, others might have limited or no GPU access. Recently, federated finetuning using synthetic data has been proposed, allowing clients to participate in a collaborative training run without training LLMs locally. However, our experimental results reveal a performance gap between models trained using synthetic data and those trained using local updates. Motivated by the observed heterogeneity in computing resources and the performance gap, we propose a novel two-stage algorithm that leverages the storage and computing capabilities of a strong server. In the first stage, under the coordination of the strong server, clients with limited computing resources collaborate to generate synthetic data, which is transferred to and stored on the strong server. In the second stage, the strong server uses this synthetic data on behalf of the resource-constrained clients to perform federated LoRA finetuning alongside clients with sufficient computing resources. This approach ensures that all clients can participate in the finetuning process. Experimental results demonstrate that incorporating local updates from even a small fraction of clients improves performance compared to using synthetic data for all clients. Furthermore, we incorporate the Gaussian mechanism in both stages to guarantee client-level differential privacy.

Wang, Jiayi [ORNL]↗
DOE OSTI2025

Biologically-informed excitatory and inhibitory ratio for robust spiking neural network training

Spiking neural networks drawing inspiration from biological constraints of the brain promise an energy-efficient paradigm for artificial intelligence. However, challenges exist in identifying guiding principles to train these networks in a robust fashion. In addition, training becomes an even more difficult problem when incorporating biological constraints of excitatory and inhibitory connections. In this work, we identify several key factors, such as low initial firing rates and diverse inhibitory spiking patterns, that determine the overall ability to train in the context of spiking networks with various ratios of excitatory to inhibitory neurons. The results indicate networks with biologically-realistic excitatory:inhibitory ratios can reliably train at low activity levels and in noisy environments. Additionally, the Van Rossum distance, a measure of spike train synchrony, provides insight into the importance of inhibitory neurons to increase network robustness to noise. This work supports further biologically-informed large-scale networks and energy efficient hardware implementations.

bio-inspired computing↗
DOE OSTI2025

Artificial Intelligence for Data Center Operations (AIOps): Cooperative Research and Development (Final Report)

High performance computing data centers will increasingly need to rely on automation to keep pace with exascale growth in compute capability and to manage and optimize the data center environment and facility resources. Artificial intelligence and machine learning approaches provide the means to improve HPC data center operational efficiency, by learning historical trends and training models to operate on real-time data collected from both IT and facilities sources. NREL has developed methods of real-time collection, aggregation and streaming of these data in the ESIF HPC Data Center and has collected a significant dataset of relevant metrics across computer systems, racks, environmental, building and utility sources for research into various predictive analytics problems. HPE's Advanced Technology Group (ATG) is doing comprehensive research into exascale monitoring and management for High Performance Computing (HPC) systems (hereinafter HPE's Data Monitoring/ Management Technology). NREL and HPE will collaborate to add Artificial Intelligence (AI) to NREL's real-time data collection/ aggregation/ streaming system and HPE's Data Monitoring/ Management System, with the goal of improving the operational efficiency of NREL's Energy Systems Integration Facility (ESIF) HPC Data Center through data analytics on both historical and real-time data from IT systems and facilities operations. This collaboration will consist of efforts in Data Management, Data Analytics, and AI/ML Optimization for both manual and autonomous intervention in data center operations. This will be a multi-year, multi-staged effort with a goal towards building capabilities for an Advanced Smart Facility, and demonstration of these techniques in the NREL ESIF HPC Data Center.

97 MATHEMATICS AND COMPUTING↗
DOE OSTI2021

Multi-Objective Hyperparameter Optimization for Spiking Neural Network Neuroevolution

Neuroevolution has had significant success over recent years, but there has been relatively little work applying neuroevolution approaches to spiking neural networks (SNNs). SNNs are a type of neural network that includes temporal processing component, are not easily trained using other methods, and can be deployed into energy-efficient neuromorphic hardware. In this work, we investigate two evolutionary approaches for training SNNs. We explore the impact of the hyperparameters of the evolutionary approaches, including tournament size, population size, and representation type, on the performance of the algorithms. We present a multi-objective Bayesian-based hyperparameter optimization approach to tune the hyperparameters to produce the most accurate and smallest SNNs. We show that the hyperparameters can significantly affect the performance of these algorithms. We also perform sensitivity analysis and demonstrate that every hyperparameter value has the potential to perform well, assuming other hyperparameter values are set correctly.

Parsa, Maryam↗