Search NASA⌕ Search

SEARCH · Search NASA

Results for “Machine Learning Algorithms”

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 649 records · Page 36
All sourcesNASA NTRSDOE OSTI
DOE OSTI2024

Design and optimization of a modular hydrogen-based integrated energy system to maximize revenue via nuclear-renewable sources

Here, this paper demonstrates a novel modular distributed framework that uses optimal energy-dispatching strategies to enable greater flexibility and profitability in nuclear-renewable integrated energy systems (NR-IES). Hydrogen is used as a commodity in this framework since its production can improve grid stability and system operational flexibility, decarbonize heavy industry, and create an additional revenue stream for electricity generators, particularly nuclear power plants with high operational expenses. The proposed solution addresses the challenges associated with merging multiple software and services from various domains by using functional mock-up units (FMU) to co-simulate diverse subsystems designed in various platforms. The tightly coupled integrated energy system (IES) is optimized to maximize revenue by utilizing the deep reinforcement learning (DRL) technique to make smart dispatching decisions based on variable electricity prices and the availability of renewable energy. Proximal policy optimization (PPO) algorithm is used in training and testing the DRL agent. Over a period of 120 days, the proposed hydrogen-based IES framework showed about 10% revenue boost compared to a non-hydrogen generating baseline IES while also providing an easily-adoptable framework which can help to improve the flexibility of future generation nuclear power plants.

29 ENERGY PLANNING, POLICY, AND ECONOMY↗
DOE OSTI2025

sOPTICS: a modified density-based algorithm for identifying galaxy groups/clusters and brightest cluster galaxies

A direct approach to studying the galaxy–halo connection is to analyse groups and clusters of galaxies that trace the underlying dark matter haloes, emphasizing the importance of identifying galaxy clusters and their associated brightest cluster galaxies (BCGs). In this work, we test and propose a robust density-based clustering algorithm that outperforms the traditional Friends-of-Friends (FoF) algorithm in the currently available galaxy group/cluster catalogues. Our new approach is a modified version of the Ordering Points To Identify the Clustering Structure (OPTICS) algorithm, which accounts for line-of-sight positional uncertainties due to redshift space distortions by incorporating a scaling factor, and is thereby referred to as sOPTICS. When tested on both a galaxy group catalogue based on semi-analytic galaxy formation simulations and observational data, our algorithm demonstrated robustness to outliers and relative insensitivity to hyperparameter choices. In total, we compared the results of eight clustering algorithms. The proposed density-based clustering method, sOPTICS, outperforms FoF in accurately identifying giant galaxy clusters and their associated BCGs in various environments with higher purity and recovery rate, also successfully recovering 115 BCGs out of 118 reliable BCGs from a large galaxy sample. Furthermore, when applied to an independent observational catalogue without extensive re-tuning, sOPTICS maintains high recovery efficiency, confirming its flexibility and effectiveness for large-scale astronomical surveys.

79 ASTRONOMY AND ASTROPHYSICS↗
DOE OSTI2025

Advancing Multiscale Simulation of Plasma-Surface Interfaces

We report the development of an atomistic-informed, surface-state-dependent predictive model for particle exchange in a carbon-tungsten plasma-surface interface. The predictive model uses machine learning (ML) techniques to learn the energy and angular distributions for particle exchange and rate functions for surface state evolution from molecular dynamics simulations of cumulative bombardment of tungsten by energetic carbon ions. Each predictive component is sensitive to the energy and trajectory of incident plasma species and the surface state. The surface state is represented by a set of surface state descriptors, which were derived from the atomistic surface state for each independent carbon bombardment event. These descriptors are representative of the composition and degree of amorphization of the outermost angstrom of surface material and were chosen to optimize predictive performance for particle exchange at the interface. The distributions for particle exchange (reflection/sputtering) are demonstrated to vary with each surface state descriptor, motivating the development of surface-state-dependent particle exchange models for plasma simulations. The performance of various ML methods was compared, including polynomial quantile regression, artificial neural networks, k-nearest neighbors, and random forest algorithms, with polynomial regression performing the best for interpolation and extrapolation of learned relationships. In addition to the particle exchange model, a neutral network was developed and used to identify data sufficiency throughout surface descriptor space, which will enable real-time feedback during future data production to ensure data is produced where it is most needed, and we provide commentary on improvements to the data production workflow for future endeavors.

70 PLASMA PHYSICS AND FUSION TECHNOLOGY↗
DOE OSTI2025

River Dissolved Oxygen Prediction Using Machine Learning Models and Wireless Sensor Measurements

Simultaneous flooding&heat and droughts&heat events can potentially destabilize hydro-meteorological conditions to deteriorate the water quality of Neches River. Machine learning (ML) models utilizing wireless sensor measurements have been applied to predict water quality and optimize various water management strategies. This study aims to develop ML models to predict dissolved oxygen (DO) prediction under various hydro-meteorological conditions and enhance water management decision-making. Wireless sensor measurements of DO, water temperature, sample depth, conductivity, turbidity, and pH, along with discharge from the United States Geological Survey stations, are collected for model inputs at the Pine Island Bayou C749 station (PIB-C749) and Neches River Saltwater Barrier (SWB). Multilayer perceptron neural networks, recurrent neural networks, long short-term memory (LSTM), and bidirectional LSTM (BiLSTM) with and without attention mechanism (AT) are tested to determine the best model, which is applied the rolling forecast method to predict 14-day DO. Traditional and recurrent transfer learning (TL and RTL) methods are adopted to overcome insufficient data at the SWB. The input feature importance analysis using the integrated gradients (IG) algorithm is applied to determine dominant inputs. The results show LSTM-based models are capable handling long sequential data. AT-BiLSTM and RTL-LSTM demonstrate the best performance at the PIB-C749 (RMSE=0.054) and the SWB (RMSE=0.028), respectively. TL and RTL methods significantly improve model performance at the SWB. DO, temperature, and pH show higher importance, consistent with hydrodynamics and water chemistry. Both best models are applied to predict 14-day DO and demonstrate reasonable performance for decision-making. Hydro-meteorological conditions of 2017 flood and 2012 drought events are simulated and reveal that possible hypoxia occurs after flooding due to increasing temperature and turbidity, and DO concentration decreases significantly under heat and drought conditions. In conclusion, LSTM-based models utilizing wireless sensor data can be a timely and effective approach to make appropriate decisions on water resource management.

54 ENVIRONMENTAL SCIENCES↗
DOE OSTI2024

How to Build a Quantum Supercomputer: Scaling from Hundreds to Millions of Qubits

In the span of four decades, quantum computation has evolved from an intellectual curiosity to a potentially realizable technology. Today, small-scale demonstrations have become possible for quantum algorithmic primitives on hundreds of physical qubits and proof-of-principle error-correction on a single logical qubit. Nevertheless, despite significant progress and excitement, the path toward a full-stack scalable technology is largely unknown. There are significant outstanding quantum hardware, fabrication, software architecture, and algorithmic challenges that are either unresolved or overlooked. These issues could seriously undermine the arrival of utility-scale quantum computers for the foreseeable future. Here, we provide a comprehensive review of these scaling challenges. We show how the road to scaling could be paved by adopting existing semiconductor technology to build much higher-quality qubits, employing system engineering approaches, and performing distributed quantum computation within heterogeneous high-performance computing infrastructures. These opportunities for research and development could unlock certain promising applications, in particular, efficient quantum simulation/learning of quantum data generated by natural or engineered quantum systems. To estimate the true cost of such promises, we provide a detailed resource and sensitivity analysis for classically hard quantum chemistry calculations on surface-code error-corrected quantum computers given current, target, and desired hardware specifications based on superconducting qubits, accounting for a realistic distribution of errors. Furthermore, we argue that, to tackle industry-scale classical optimization and machine learning problems in a cost-effective manner, heterogeneous quantum-probabilistic computing with custom-designed accelerators should be considered as a complementary path toward scalability.

Mohseni, Masoud↗
DOE OSTI2025

FluxRETAP: a REaction TArget Prioritization genome-scale modeling technique for selecting genetic targets

MOTIVATION: Metabolic engineering is rapidly evolving as a result of new advances in synthetic biology tools and automation platforms that enable high throughput strain construction, as well as the development of machine learning tools (ML) for biology. However, selecting genetic engineering targets that effectively guide the metabolic engineering process is still challenging. ML can provide predictive power for synthetic biology, but current technical limitations prevent the independent use of ML approaches without previous biological knowledge. RESULTS: Here, we present FluxRETAP, a simple and computationally inexpensive method that leverages the prior mechanistic knowledge embedded in genome-scale models for suggesting targets for genetic overexpression, downregulation or deletion, with the final goal of increasing the production of a desired metabolite. This method can provide a list of desirable engineering targets that can be combined with current ML pipelines. FluxRETAP captured 100% of reaction targets experimentally verified to improve Escherichia coli isoprenol production, 50% of targets that experimentally improved taxadiene production in E. coli and ∼60% of genetic targets from a verified minimal constrained cut-set in Pseudomonas putida, while providing additional high priority targets that could be tested. Overall, FluxRETAP is an efficient algorithm for identifying a prioritized list of testable genetic and reaction targets. AVAILABILITY AND IMPLEMENTATION: FluxRETAP is implemented in python and released under the creative commons license. The implementation and code are freely available at: https://github.com/JBEI/FluxRETAP.

Czajka, Jeffrey J↗
DOE OSTI2024

Flux REaction TArget Prioritization (Flux RETAP) v1

Metabolic engineering is evolving rapidly as a result of new advances in synthetic biology and automation, as well as the irruption of machine learning (ML). ML has been shown to provide the predictive power synthetic biology lacked and needed, and to be able to effectively guide the metabolic engineering process. However, current technical limitations prevent the independent application of ML approaches to metabolic engineering without the use of previous biological knowledge in the form of a prioritized list of desirable engineering targets. Here, we present FluxRETAP, a simple and computationally inexpensive method that leverages the prior mechanistic knowledge embedded in genome-scale metabolic models (GSMs) for suggesting targets for genetic overexpression, downregulation or deletion, with the final goal of increasing metabolite production. FluxRETAP captured 100% of reaction targets experimentally verified to improve Escherichia coli isoprenol production in the literature accessible to us, 50% of targets that experimentally improved taxadiene production in E. coli and ~60% of genetic targets from a verified minimal constrained cut-set in Pseudomonas putida while providing additional high priority targets that could be tested. Overall, FluxRETAP is an efficient algorithm for identifying a prioritized list of testable genetic and reaction targets which can also be utilized in ML pipelines.

Czajka, Jeffrey [Battelle Memorial Institute, Paci↗
DOE OSTI2025

Nonlinear Ensemble Filtering with Diffusion Models: Application to the Surface Quasigeostrophic Dynamics

The intersection between classical data assimilation methods and novel machine learning techniques has attracted significant interest in recent years. Here, we explore another promising solution in which diffusion models are used to formulate a robust nonlinear ensemble filter for sequential data assimilation. Unlike standard machine learning methods, the proposed ensemble score filter (EnSF) is completely training free and can efficiently generate a set of analysis ensemble members. Here, in this study, we apply the EnSF to a surface quasigeostrophic model and compare its performance against the popular local ensemble transform Kalman filter (LETKF), which makes Gaussian assumptions in the analysis step. Numerical tests demonstrate that EnSF maintains stable performance in the absence of localization and for a variety of experimental settings. We find that while LETKF maintains optimal performance in the case of linear observations of the entire state and a perfect model, EnSF shows improvements over LETKF when nonlinear observations are assimilated and the system is subject to unexpected model errors. A spectral decomposition of the analysis results in this nonlinear observation regime shows that the largest improvements over LETKF occur at large scales (small wavenumbers), where LETKF lacks sufficient ensemble spread. Overall, this initial application of EnSF to a geophysical model of intermediate complexity motivates further development of the algorithm for more realistic problems.

Artificial intelligence↗
DOE OSTI2026

Leveraging machine learning to enhance aerosol classification using Single-Particle Mass Spectrometry

Advancing automated classification of atmospheric aerosols from Single-Particle Mass Spectrometry (SPMS) data remains challenging due to overlapping ion signatures, compositional diversity, and limited labeled data. This study evaluates supervised and semi-supervised learning frameworks to enhance aerosol identification by jointly leveraging labeled and unlabeled spectra. Four models were compared: a supervised Support Vector Machine (SVM), a self-training SVM, a stacked autoencoder classifier, and a stacked autoencoder trained using a temporal-ensembling Mean Teacher approach. All models achieved high and stable accuracies (90.0 %–91.1 %), surpassing previous results on the same dataset (87 %) and matching the performance of state-of-the-art deep learning methods. Despite small global metric differences (≤ 1 %), semi-supervised variants yielded up to 5 %–10 % improvements for compositionally rare particle types – such as soot (0.77 % of spectra, F1-score: 0.93–0.97) and hazelnut pollen (0.98 % of spectra, F1-score: 0.97–1.00) – equating to roughly ∼ 187 additional correctly classified spectra. These gains are scientifically significant, as such rare particles exert disproportionate influence on radiative absorption and ice nucleation processes; their improved detection reduces modeled uncertainties in aerosol absorption optical depth and mixed-phase cloud ice nucleation rates. The models' residual misclassifications (≈ 9 %) largely arise from true spectral overlap among chemically adjacent species (e.g., Na- vs. K-feldspar, coated vs. uncoated feldspars), reflecting physical compositional continuity rather than algorithmic error. Collectively, these findings demonstrate that leveraging unlabeled data to learn robust spectral representations and refine classification enhances both fidelity and interpretability, bridging data-driven analysis with aerosol–climate process understanding.

54 ENVIRONMENTAL SCIENCES↗
DOE OSTI2026

Graph theory inspired anomaly detection at the LHC

Designing model-independent anomaly detection algorithms for analyzing LHC data remains a central challenge in the search for new physics, due to the high dimensionality of collider events. In this work, we develop a graph autoencoder as an unsupervised, model-agnostic tool for anomaly detection, using the LHC Olympics dataset as a benchmark. By representing jet constituents as a graph, we introduce a method to systematically control the information available to the model through sparse graph constructions that serve as physically motivated inductive biases. Specifically, (1) we construct graph autoencoders based on locally rigid Laman graphs and globally rigid unique graphs, and (2) we explore the clustering of jet constituents into subjets to interpolate between high- and low-level input representations. We obtain the best performance, measured in terms of the Significance Improvement Characteristic curve for an intermediate level of subjet clustering and certain sparse unique graph constructions. We further investigate the role of graph connectivity in jet classification tasks. Our results demonstrate the potential of leveraging graph-theoretic insights to refine and increase the interpretability of machine learning tools for collider experiments.

Automation↗
DOE OSTI2025

Replacing non-biomedical concepts improves embedding of biomedical concepts

Embeddings are semantically meaningful representations of words in a vector space, commonly used to enhance downstream machine learning applications. Traditional biomedical embedding techniques often replace all synonymous words representing biological or medical concepts with a unique token, ensuring consistent representation and improving embedding quality. However, the potential impact of replacing non-biomedical concept synonyms has received less attention. Embedding approaches often employ concept replacement to replace concepts that span multiple words, such as non-small-cell lung carcinoma, with a single concept identifier (e.g., D002289). Also, all synonyms of each concept are merged into the same identifier. Here, we additionally leveraged WordNet to identify and replace sets of non-biomedical synonyms with their most common representatives. This combined approach aimed to reduce embedding noise from non-biomedical terms while preserving the integrity of biomedical concept representations. We applied this method to 1,055 biomedical concept sets representing molecular signatures or medical categories and assessed the mean pairwise distance of embeddings with and without non-biomedical synonym replacement. A smaller mean pairwise distance was interpreted as greater intra-cluster coherence and higher embedding quality. Embeddings were generated using the Word2Vec algorithm applied to a corpus of 10 million PubMed abstracts. Our results demonstrate that the addition of non-biomedical synonym replacement reduced the mean intra-cluster distance by an average of 8%, suggesting that this complementary approach enhances embedding quality. Future work will assess its applicability to other embedding techniques and downstream tasks. Python code implementing this method is provided under an open-source license.

algorithms↗
DOE OSTI2025

Artificial intelligence driven laser parameter search: Inverse design of photonic surfaces using greedy surrogate-based optimization

Photonic surfaces designed with specific optical characteristics are becoming increasingly crucial for novel energy harvesting and storage systems. The design of these surfaces can be achieved by texturing materials using lasers. The optimal adjustment of laser fabrication parameters to achieve target surface optical properties is an open challenge. Thus, we develop a surrogate-based optimization approach. Our framework employs the Random Forest algorithm to model the forward relationship between the laser fabrication parameters and the resulting optical characteristics. During the optimization process, we use a greedy, prediction-based exploration strategy that iteratively selects batches of laser parameters to be used in experimentation by minimizing the predicted discrepancy between the surrogate model’s outputs and the user-defined target optical characteristics. This strategy allows for efficient identification of optimal fabrication parameters without the need to model the error landscape directly. We demonstrate the efficiency and effectiveness of our approach on two synthetic benchmarks and two specific experimental applications of photonic surface inverse design targets. By calculating the average performance of our algorithm compared to other state of the art optimization methods, we show that our algorithm performs, on average, twice as well across all benchmarks. Additionally, a warm starting inverse design technique for changed target optical characteristics enhances the performance of the introduced approach.

97 MATHEMATICS AND COMPUTING↗
DOE OSTI2025

TorchBraid: High-Performance Layer-Parallel Training of Deep Neural Networks with MPI and GPU Acceleration

TorchBraid is a high-performance implementation of layer-parallel training for deep neural networks (DNNs) supporting MPI-based parallelism and GPU acceleration. Layer-parallel training has been developed to overcome the serialization inherent in forward and backward propagation of DNNs that limits utilization of computational resources in the strong scaling limit. To achieve this, TorchBraid integrates the PyTorch neural network framework with the state-of-the-art XBraid time-parallel library. Furthermore, this article presents the use and performance of TorchBraid, in addition to solutions for overcoming the algorithmic challenges inherent in combining automatic differentiation with layer-parallel. Results are presented with and without GPU acceleration for the Tiny ImageNet and MNIST image classification data sets, as well as recurrent neural networks. Overall, TorchBraid enables fast training of DNNs, both in a strong and weak scaling context. In addition to the TorchBraid software, several new advances in applying layer-parallel algorithms are detailed. Integration of layer-parallel with data-parallel algorithms is presented for the first time, showing the computational advantages of the combination. Standard deep learning techniques, like batch-normalization, are developed for layer-parallel training. Finally, a new approach combining layer-parallel with spatial coarsening in order to accelerate training for 3D image classification shows roughly a 10× speedup over serial execution.

Layer-parallel↗
DOE OSTI2025

Accelerating iterative ptychography with an integrated neural network

Electron ptychography is a powerful and versatile tool for high-resolution and dose-efficient imaging. Iterative reconstruction algorithms are powerful but also computationally expensive due to their relative complexity and the many hyperparameters that must be optimised. Gradient descent-based iterative ptychography is a popular method, but it may converge slowly when reconstructing low spatial frequencies. Here, in this work, we present a method for accelerating a gradient descent-based iterative reconstruction algorithm by training a neural network (NN) that is applied in the reconstruction loop. The NN works in Fourier space and selectively boosts low spatial frequencies, thus enabling faster convergence in a manner similar to accelerated gradient descent algorithms. We discuss the difficulties that arise when incorporating a NN into an iterative reconstruction algorithm and show how they can be overcome with iterative training. We apply our method to simulated and experimental data of gold nanoparticles on amorphous carbon and show that we can significantly speed up ptychographic reconstruction of the nanoparticles.

4DSTEM↗
DOE OSTI2025

A resolution independent neural operator

The Deep operator network (DeepONet) is a powerful yet simple neural operator architecture that utilizes two deep neural networks to learn mappings between infinite-dimensional function spaces. This architecture is highly flexible, allowing the evaluation of the solution field at any location within the desired domain. However, it imposes a strict constraint on the input space, requiring all input functions to be discretized at the same locations; this limits its practical applications. Here, in this work, we introduce a general framework for operator learning from input–output data with arbitrary number and locations of sensors. This begins by introducing a resolution-independent DeepONet (RI-DeepONet), enabling it to handle input functions that are arbitrarily, but sufficiently finely, discretized. To this end, we propose two dictionary learning algorithms to adaptively learn a set of appropriate continuous basis functions, parameterized as implicit neural representations (INRs), from correlated signals defined on arbitrary point cloud data. These basis functions are then used to project arbitrary input function data as a point cloud onto an embedding space (i.e., a vector space of finite dimensions) with dimensionality equal to the dictionary size, which can be directly used by DeepONet without any architectural changes. In particular, we utilize sinusoidal representation networks (SIRENs) as trainable INR basis functions. The introduced dictionary learning algorithms are then used in a similar way to learn an appropriate dictionary of basis functions for the output function data, which defines a new neural operator architecture referred to as the R esolution I ndependent N eural O perator (RINO). In the RINO, the operator learning task simplifies to learning a mapping from the coefficients of input basis functions to the coefficients of output basis functions. We demonstrate the robustness and applicability of RINO in handling arbitrarily (but sufficiently richly) sampled input and output functions during both training and inference through several numerical examples.

Deep operator network (DeepONet)↗
DOE OSTI2026

Integrated System Planning: Emerging Software Requirements in the Power Industry

Power system planning software remains fragmented across organizational boundaries, with specialized tools for capacity expansion, production cost modeling, power flow, and dynamic analysis operating on incompatible data models and assumptions. This article argues that the fragmentation is not merely a technical problem but a predictable consequence of Conway's law: software architectures mirror the departmental structures within which they are developed. Regulatory milestones like Federal Energy Regulatory Commission (FERC) Order 888 formalized these divisions, but the roots trace back to the distinct engineering disciplines-mechanical, chemical, and electrical-that staffed generation and transmission planning departments in vertically integrated utilities. As the industry moves toward integrated system planning (ISP) that coordinates generation, transmission, and distribution investment decisions, the software ecosystem must evolve accordingly. We identify five categories of software requirements to enable this transition: coherent data inputs decoupled from individual applications, unified and extensible data schemas, modular component representations that support multiple abstraction levels, lifecycle management of planning datasets, and well-defined application programming interface (API) contracts that separate data exchange from algorithmic control. We examine how these requirements interact with three common workflow patterns-serial gate clearing, sequential multiapplication, and convergence oriented-and discuss the interface design principles each demands. We then outline a vision for platform-based planning architectures where specialized analytical services compose through standardized interfaces and where artificial intelligence (AI)/machine learning (ML) tools augment decision support within a disciplined software infrastructure. The practices proposed here offer a path from today's siloed tool collections toward collaborative planning ecosystems capable of handling the complexity of modern power system transformation.

24 POWER TRANSMISSION AND DISTRIBUTION↗
DOE OSTI2026

ZENN: A thermodynamics-inspired computational framework for heterogeneous data–driven modeling

Traditional entropy-based methods—such as cross-entropy loss in classification problems—have long been essential tools for representing the information uncertainty and physical disorder in data and for developing artificial intelligence algorithms. However, the rapid growth of data across various domains has introduced new challenges, particularly the integration of heterogeneous datasets with intrinsic disparities. To address this, we introduce a zentropy-enhanced neural network (ZENN), extending zentropy theory into the data science domain via intrinsic entropy, enabling more effective learning from heterogeneous data sources. ZENN simultaneously learns both energy and intrinsic entropy components, capturing the underlying structure of multisource data. To support this, we redesign the neural network architecture to better reflect the intrinsic properties and variability inherent in diverse datasets. We demonstrate the effectiveness of ZENN on classification tasks and energy landscape reconstructions, showing its superior generalization capabilities and robustness-particularly in predicting high-order derivatives. In image and text classification tasks, ZENN demonstrates superior generalization by introducing a learnable temperature variable that models latent multisource heterogeneity, allowing it to surpass state-of-the-art models on CIFAR-10/100, BBC News, and AG News. As a practical application in materials science, we employ ZENN to reconstruct the Helmholtz energy landscape of Fe3Pt using data generated from density functional theory and capture key material behaviors, including negative thermal expansion and the critical point in the temperature–pressure space. Overall, this work presents a zentropy-grounded framework for data-driven machine learning, positioning ZENN as a versatile and robust approach for scientific problems involving complex, heterogeneous datasets.

36 MATERIALS SCIENCE↗
DOE OSTI2026

Measuring the Energy Consumption and Efficiency of Deep Neural Networks: An Empirical Analysis and Design Recommendations

Addressing the "Red-AI" trend of rising energy consumption by large-scale neural networks, this study investigates the measured energy consumption of training various fully connected neural network architectures. We introduce the BUTTER-E dataset, an augmentation to the BUTTER Empirical Deep Learning dataset, containing energy consumption and performance data from 41,129 individual experimental runs spanning 30,582 distinct configurations: 13 datasets, 20 sizes (trainable parameters), 8 "shapes", and 14 depths on both CPUs and GPUs using node-level watt-meters. This dataset reveals the complex relationship between dataset size, network structure, and energy use. Our analysis uncovers a surprising, hardware-mediated non-linear relationship between energy efficiency and network design, challenging the assumption that reducing the number of parameters or FLOPs is the best way to achieve greater energy efficiency. We propose a straightforward and effective energy model that accounts for network size, computing, and memory hierarchy. Highlighting the need for cache-considerate algorithm development, we suggest a codesign approach to energy efficient network, algorithm, and hardware design. This work contributes to the fields of sustainable computing and Green AI, offering practical guidance for creating more energy-efficient neural networks and promoting sustainable AI.

97 MATHEMATICS AND COMPUTING↗