Engineering Papers⌕ Search

SEARCH · Engineering Papers

Results for “Surprisal”

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 1,189 records · Page 66
All sourcesNASA NTRSDOE OSTI
DOE OSTI2024

Human Subjects in Energy Technology and Policy Research (Symposium Report)

The U.S. Department of Energy (DOE) has a unique and wide-ranging research portfolio. One focus of the department’s research mission is to develop and promote the adoption of technologies that reduce greenhouse gas emissions and result in renewable and clean energy for the country. This research encompasses a surprisingly diverse array of studies, many of which involve the inclusion of human participants that go beyond traditional biomedical investigations. For example, individuals play an important role in DOE research on clean energy sources, development and testing of new energy-efficient technologies, and partnerships with local communities for equitable climate solutions.

29 ENERGY PLANNING, POLICY, AND ECONOMY↗
DOE OSTI2024

Fermilab Booster Beam Emittances from Quadrupole Modes Measured BPMs

The measurement of beam emittances by extracting the quadrupole mode signal from a 4 plate beam position monitor (BPM) was published at least 40 years ago. Unfortunately, in practice, this method suffers from poor signal to noise ratio and requires a lot of tuning to extract out the emittances. In this paper, an improved method where multiple BPMs are used together with better mathematical analysis is described. The BPM derived emittances are then compared with those measured by the Ion Profile Monitor (IPM). Surprisingly, the BPM measured emittances behave very well and are more realistic than those measured by the IPM.

43 PARTICLE ACCELERATORS↗
DOE OSTI2024

One Year In: Tracking the Impacts of NEM 3.0 on California’s Residential Solar Market

On December 15, 2022, the California Public Utilities Commission passed an overhaul of the net metering program for the state’s investor-owned utilities. The changes replaced the long-standing net energy metering (NEM) tariffs with a net billing tariff (NBT) structure—colloquially known as “NEM 3.0”—which significantly reduces the compensation for behind-the-meter solar photovoltaic (PV) systems. The NEM tariffs remained open for new interconnection applications until April 15, 2023, but after that date, all new interconnection applications were submitted under NBT. Now, one year later, we have an opportunity to evaluate how the California solar market has evolved under this new compensation regime. As a precursor to its annual Tracking the Sun report, Berkeley Lab has released a short technical brief describing key trends in the California residential solar market since the roll-out of the new NBT structure. The purpose of this analysis is to provide empirical insights into how the market has evolved over the past year, confirming some expectations while also revealing several striking surprises.

14 SOLAR ENERGY↗
DOE OSTI2024

Study of the Neutrino Magnetic Moment with the NOvA Near Detector

Predicted by the Standard Model as theoretical, massless particles, neutrinos have been the subject of many experiments since their first detection. It is now experimentally confirmed that neutrinos do have mass necessitating an extension to the Standard Model. Such an extension allows for other surprising neutrino properties, such as a neutrino magnetic moment. While neutrinos are observed to be neutral and do not couple to photons at leading order, higher order expansion of the interaction allows for coupling to the photon to occur and gives rise to a neutrino magnetic moment through quantum loop effects. This is a useful property for studying the Dirac or Majorana nature of neutrinos, as the predicted value of the magnetic moment would differ. This talk focuses on an introduction to the neutrino magnetic moment as well as discusses the current status of work being done on NOvA utilizing the Near Detector to obtain an upper limit on the neutrino magnetic moment value.

43 PARTICLE ACCELERATORS↗
DOE OSTI2024

X-ray Security Imaging on Personal Dosimetry [Slides]

Federal regulations mandate that personal dosimetry devices — like optically stimulated luminescence (OSLs) — must be worn by all US Department of Energy (DOE) and associated radiation workers to track their occupational dose. Unfortunately, the inadvertent passage of OSLs through x-ray security scanners can compromise their validity. With the advent of high energy, advanced resolution security technology used in airports, this once insignificant issue now requires that Radiological Control (RadCon) be able to accurately discern non-occupational dose to effected OSLs. This presentation will discuss the principles, methods, and, rather surprising, models for establishing the corrective dose estimates at the Idaho National Laboratory (INL), and its implication across the DOE.

61 RADIATION PROTECTION AND DOSIMETRY↗
DOE OSTI2023

Self-healing, self-assembling islanded power systems using only local measurements

This SAND report collects the results from the LDRD project “SHAZAM”, which aimed to push the limits of performance for self-healing, self-assembling power systems whose sectionalizing and load-control agents rely on local measurements only (i.e., only what they can measure at their own terminals, with no data sharing between agents). This work includes self-networking microgrids. The key objectives of this work were a) to demonstrate how high the performance of local-measurement-only self-assembling power systems can be; and b) to solve certain technical problems associated with such systems, such as their inability to prevent the accidental formation of closed loops and their tendency to thermally overload some conductors. “SHAZAM” investigators a) demonstrated that the performance of such systems can be surprisingly high, b) demonstrated that such systems are quite robust to all kinds of variations, and c) developed and demonstrated solutions to several key challenges associated with this type of system.

24 POWER TRANSMISSION AND DISTRIBUTION↗
DOE OSTI2021

Correspondence of NNGP Kernel and the Matérn Kernel

Kernels representing limiting cases of neural network architectures have recently gained popularity. However, the application and performance of these new kernels compared to existing options, such as the Matérn kernel, is not well studied. We take a practical approach to explore the neural network Gaussian process (NNGP) kernel and its application to data in Gaussian process regression. We first demonstrate the necessity of normalization to produce valid NNGP kernels and explore related numerical challenges. We further demonstrate that the predictions from this model are quite inflexible, and therefore do not vary much over the valid hyperparameter sets. We then demonstrate a surprising result that the predictions given from the NNGP kernel correspond closely to those given by the Matérn kernel under specific circumstances, which suggests a deep similarity between overparameterized deep neural networks and the Matérn kernel. Finally, we demonstrate the performance of the NNGP kernel as compared to the Matérn kernel on three benchmark data cases, and we conclude that for its flexibility and practical performance, the Matérn kernel is preferred to the novel NNGP in practical applications.

97 MATHEMATICS AND COMPUTING↗
DOE OSTI2024

Photoinitiated Reactions of Molecules and Radicals in Molecular Beams

The UV photochemistry of organic molecules is a fundamental process that governs reactions in the atmosphere, synthetic chemistry, processing of organic aerosols, and biological damage in living tissues. In many organic molecules, the ensuing evolution involves pathways that are in competition, giving rise to different products, isomerization, coupling to other electronic states, and secondary reactive collisions. Because photodissociation is usually fast (picoseconds to microseconds), these processes are far from equilibrium and are controlled by kinetic competition and dynamical forces. The work accomplished was focused primarily on the photochemistry of alpha-keto carboxylic acids, a group of acids produced from natural sources, which are implicated in aerosol formation and biological processes. In the atmosphere, they are destroyed mainly by solar radiation. Studying their photochemistry has been surprisingly difficult because of the complexity of their excited electronic states, and the effect of collisions and secondary reactions. The second project investigated the lifetime of excited electronic states of pyrazine and picoline and had both experimental and theoretical aspects. The work focused on a model molecule, pyruvic acid (PA), because it was hypothesized that it should be a good source of the unstable carbene, methylhydroxycarbene (MHC). PA has an internal hydrogen bond that controls the evolution of its decomposition. The decomposition was studied following excitation to two of its lowest excited states reached by laser irradiation at 351 and 193 nm. The photodissociation dynamics in the absence and presence of collisions was monitored and compared. Understanding the UV photochemistry requires the use of complementary experimental approaches. Two methods were enlisted, which together generated a comprehensive and detailed set of results: (i) The time-sliced velocity map imaging (SVMI) instrument at USC was exploited to determine kinetic energy release of fragments, fast dissociation timescales, and internal state distributions of fragments for which Resonance Enhanced Multiphoton Ionization (REMPI) schemes exist; and (ii) The multiplexed photoionization mass spectrometer (MPIMS) setup developed at the Sandia Combustion Research Facility was used for product discovery, achieved by exploiting tunable narrowband VUV radiation at the Advanced Light Source (ALS), and to follow in real time their subsequent unimolecular and bimolecular reactions. The experimental conditions ranged from collisionless molecular beams to study nascent products to flow reactors of variable pressures to study subsequent bimolecular reactions of the products. The goals stated above were successfully accomplished. By exciting PA to its lowest excited state, MHC was identified as the only primary product and its isomerization to vinylalcohol and acetaldehyde was directly observed in real time. Moreover, we reported the first bimolecular reaction of MHC, which was with acetaldehyde, and identified its reaction product. This paves the way for the study of other reactions of this important carbene intermediate. When excitation was carried out at 193 nm, which imparted much higher energy to PA, many more products were directly observed in real time, some deriving from three-body dissociation. Quantitative branching ratios and secondary reactions of radical products were also determined and analyzed. The studies of pyrazine and picoline focused on the second excited state of these molecules and showed (via ionization studies) that their excited states were very short lived (<100 fs) because of efficient couplings to lower electronic states via vibronic coupling. Theoretical work on modeling the absorption spectrum and decay mechanisms of the excited states are in progress in collaboration with theoreticians.

37 INORGANIC, ORGANIC, PHYSICAL, AND ANALYTICAL CH↗
DOE OSTI2024

Radiometric Testing of Germicidal UV Products, Round 2: Upper-Room Luminaires (CALiPER Report)

This report analyzes the independently tested performance of eight germicidal ultraviolet (GUV) upper-room luminaires marketed for use in occupied spaces and purchased between March and June 2023. This type of product is mounted to upper walls or ceilings to treat air in the portion of the room above occupants; this allows for safe use of the room when the device is operating, but requires sufficient air mixing between upper and lower portions of the room. Three of the luminaires used UV-emitting LEDs, and the remaining five luminaires used low-pressure mercury (LPM) lamps. Product testing covered radiometric and electrical performance for each luminaire. Initial performance was measured for all eight products, and four were additionally measured after 100 h and 500 h of operation. Measured performance data allowed for comparison against manufacturer or vendor claims if the tested products included such claims. Some products had no performance data available for a given quantity (e.g., UV-C output power), and only four of the eight luminaires had radiant intensity distribution data files in a standard format (e.g., IES LM-63) available for download from product websites. The lack of publicly available performance data makes it difficult for potential buyers and specifiers to identify suitable products and design GUV systems for their specific applications. When products had performance claims, they were sometimes contradictory (e.g., unexplained differences between multiple power values) or ambiguous (e.g., measurement units conflict with quantity, unclear whether luminaire power or lamp power, unclear whether UV output power or UV-C output power). Three of the eight tested luminaires had claimed output power (i.e., radiant flux) values that exceeded measured values by more than an order of magnitude. There was substantial variation in UV-C radiant efficiency, with a measured range of 0.3–1.9% for LED and 0.4–2.1% for LPM, as shown in Figure 1. For example, the LPM luminaire with 0.4% radiant efficiency would need 5 times the amount of electrical energy used by the LPM luminaire with 2.1% radiant efficiency to produce the same amount of UV-C output power. LPM luminaires that had parabolic reflectors aligned with inclined louvers exhibited substantially higher UV-C radiant efficiency than tested luminaires with other designs, potentially cutting energy use by 75%. These results indicate a substantial opportunity for more energy efficient LPM luminaire designs, while demonstrating that UV LED luminaires can offer comparable UV-C radiant efficiency in this application. This may seem surprising, given that LED emitters have lower UV-C radiant efficiency than LPM lamps, but the efficiency-throttling louvers that are generally required for LPM luminaires typically are not needed for LEDs thanks to their directionality. However, lateral beam angles (which describe beam width as viewed from above) were 41–83° for LED luminaires versus 89–110° for LPM luminaires. More luminaires may be required if their lateral beam angles are relatively small, and coverage may be poor if UV-C radiant intensity distribution (i.e., beam shape) is not considered when designing systems.

32 ENERGY CONSERVATION, CONSUMPTION, AND UTILIZATI↗
DOE OSTI2024

Effective Strength of Sinusoidally-Varying Focussing

Particle beam focussing and other confined physical systems often obey the equation ¨x = kx, where k < 0 implies stable motion. A constant negative value of k yields simple harmonic motion (constant focussing), while positive values give unstable (defocussing) motion. Alternating gradient focussing is the surprising principle that if k(t) is varied periodically, the overall motion can sometimes be stable even if the average value of k is zero or negative. It is used where focussing in one direction implies defocussing in one or more other directions, such as with electromagnetic fields in free space that have to satisfy ∇ · E = 0 and similar.

43 PARTICLE ACCELERATORS↗
DOE OSTI2024

Picosecond Carrier Dynamics in InAs and GaAs Revealed by Ultrafast Electron Microscopy

Understanding the limits of spatio-temporal carrier dynamics, especially in III-V semiconductors, is key to designing ultrafast and ultra-small optoelectronic components. However, identifying such limits and the properties controlling them has been elusive. Here, using scanning ultrafast electron microscopy (SUEM), in bulk n-GaAs and p-InAs, we simultaneously measure picosecond carrier dynamics along with three related quantities: sub-surface band bending, above-surface vacuum potentials, and surface trap densities. We make two surprising observations. First, we uncover a negative-time contrast in secondary electrons resulting from an interplay among these quantities. Second, despite dopant concentrations and surface state densities differing by many orders of magnitude between the two materials, their carrier dynamics, measured by photo-excited band bending and filling of surface states, occur at a seemingly common timescale of about 100 ps. This observation may indicate fundamental kinetic limits tied to a multitude of material and surface properties of optoelectronic III-V semiconductors, and highlights the need for techniques that simultaneously measure electro-optical kinetic properties.

36 MATERIALS SCIENCE↗
DOE OSTI2025

Implications of Aerosol Physicochemical Properties Including Ice Nucleation at ARM Mega Sites for Improved Understanding of Microphysical Atmospheric Cloud Processes (Final Technical Report)

Continuous, long-term measurements of atmospheric ice-nucleating particles (INPs) that influence clouds and precipitation were conducted to investigate the abundance and variability of ground-level INPs across the world. Three field campaigns were organized by the DOE Atmospheric Radiation Measurement (ARM) program, including Examining INP from Southern Great Plains (ExINP-SGP, 2019), Examining INP from Eastern North Atlantic (ExINP-ENA, 2020 – 2021), and Examining INP from North Slope of Alaska (ExINP-NSA, 2021 – 2024). Additional small-scale supporting field experiments were performed in 2019 and 2021 to collect airborne particulate matter at SGP for complementary laboratory characterization of the particles’ physical and chemical properties [Aerosol-Ice Formation Closure Pilot Study (AEROICESTUDY), 2019; ExINP-SGP II, 2021]. In these studies, the PI’s team measured INP concentration with both real-time and laboratory measurements in a wide range of freezing temperatures ($T$ from 0 °C to about –30 °C). This project elucidated spatial variability and seasonality in the abundance of immersion mode active INPs across three ARM sites using a single instrument, a Portable Ice Nucleation Experiment (PINE) chamber version 03 (PINE-03 hereafter). Collocated aerosol and meteorological data were analyzed to assess the correlation between ambient INP abundance, air mass origin region, and meteorological variability. Our findings suggest very high freezing efficiency of INPs at the NSA site across the measured temperatures (ice nucleation active surface site density, $n_s(T)$, $\approx 2 \times 10^{8} - 10^{10}$ m -2 for from –16 to –31 °C), which is a factor of 10 – 1000 times greater efficiency as compared to that found in the previous mid-latitude INP measurements in autumn using the same instrument; surprisingly high INP abundance ($\ge 1 \text{ L}^{-1}$ at –25 °C) for the examined temperatures throughout the year that PINE-03 did not measure at other sites; and high INP concentration in spring, possibly related to arctic haze episodes.

58 GEOSCIENCES↗
DOE OSTI2025

SCGSR Final Report

Dual phase noble liquid time projection chambers (TPCs) have an unprecedented sensitivity towards dark matter searches and neutrinos, owing to their ability to detect single electrons. When a particle deposits energy into one of these detectors, electrons are released and drifted into a gaseous region of high electric field. In the gas, the electrons are accelerated to produce extra light into what is called an S2 pulse. This light is far easier to detect than the current induced from a single electron. The two widely used elements for noble liquid detectors are Xe and Ar. Xe has a larger chance of being hit by neutrinos or dark matter, but ends with a lower (i.e. harder to detect) energy when struck. However, Ar light is hard to reflect and detect due to its short wavelength which is absorbed by most materials. While one may use TPB to re-emit this light at a longer wavelength, spatial variations of TPB thickness can smear the amount of re-emitted light, and thus cause a poor S2 energy resolution and spatial resolution. Furthermore, a single electron pulse in Ar is spread over a much longer time than one in Xe, which makes it harder to identify single electrons in Ar. The goal of Xe doping is to combine the best aspects of Ar with the best aspects of Xe. More specifically, this means a detector with a larger yield of photons and electrons produced per energy deposition, an ability to collect more of the light given by those signals, and a narrower pulse from the electron signal. During this project, we developed and operated a detector that can overcome the thermodynamic challenges of Xe doping – namely, preventing Xe from freezing. In the first experimental run during this award period, we found that we can achieve at least 5% of Xe in LAr, and measure the Xe concentration in the gas. In the second run, we installed our detector and saw the first S2 pulses in a dual phase Xe doped LAr TPC. With the data we took, we quantified the effect of Xe doping on the detectability of wavelength-shifted light produced in a Xe doped Ar mixture, as well as the single electron pulse shape. While the Ar light has successfully been wavelength shifted, we learned that much to our surprise, Xe doping actually makes the S2 pulse wider before it makes it narrower.

72 PHYSICS OF ELEMENTARY PARTICLES AND FIELDS↗
DOE OSTI2025

A Data-Agnostic, Continuous Machine Learning Framework for Application in High Energy Physics and Beyond: Phase 1 Final Scientific/Technical Report

This Phase 1 effort has focused on the development of continual learning frameworks for use in machine learning, specifically in the applied context of High Energy Physics (HEP). Machine learning (ML) is a transformative technology by which computers, typically through the use of neural networks, are able to perform tasks with proficiency that rivals or surpasses that of human users. Model Degradation & Catastrophic Forgetting are two undesired phenomena which can occur in ML where the performance of a model degrades when either deployed on novel data streams, or trained on novel data which are sufficiently different than the data the models were initially trained on. A natural example where these sorts of effects can be observed is in the performance of detectors in harsh environments, where the detector signature may change over the lifetime of the detector as it ages and deteriorates — precisely what occurs in the experiments conducted in HEP. Real world HEP data is therefore an excellent test-ground and use-case for Continual Learning paradigms, which are techniques used in ML to counteract these problems. Ensemble learning is one such technique, where multiple smaller models are trained on subsets of the overall data and are ensembled together during inference. The intuition behind this technique is that, although there are shifts in the distributions which govern the incoming data streams, these shifts are not expected to be homogeneous or global. If a sufficient diversity in solutions within the various sub-models has been achieved, then at least one sub-model is expected to retain its performance within the overall ensemble. One further strength of this approach is that the architectures of the various models do not need to be identical, and in fact even different modalities of data can naturally be combined in this way. This work focused on applying ensemble learning techniques to derive results using two main datasets, anomaly detection in HEP data & time-series forecasting in semiconductor manufacturing data. Semiconductor manufacturing involves data with surprising similarity to that of HEP (e.g. wafer maps look very similar to digi-occupancy maps) and Cerium Lab’s prominence within the semiconductor industry makes semiconductor manufacturing a natural opportunity for commercialization of this work. Our efforts have led to two strong results. The first is that we evaluated the proposed ensembling techniques using previously proposed machine learning architectures for use in anomaly detection, namely AutoEncoder based models and their derivatives. We also developed new architectures which have not been evaluated in this context before. In fact, this work marks the first use of Vision Transformers for anomaly detection in HEP. Second, we demonstrated that ensemble learning significantly improves model performance in scenarios prone to degradation, validating its effectiveness across both HEP and semiconductor datasets. These results further support ensemble learning as a powerful strategy for mitigating catastrophic forgetting and maintaining robust performance in evolving data environments.

72 PHYSICS OF ELEMENTARY PARTICLES AND FIELDS↗
DOE OSTI2023

Emergent Behaviors and the Limits of Large Language Model Generalization

Over the past couple years, large language models (LLMs) have rapidly grown in size and performance across wide numbers of tasks, leading some to herald them as machine learning models with truly “general” capabilities. We address some of the ways that the language used to describe these models can be both deceptive and a poor framing for making progress towards understanding their capabilities, and also review arguments on to what extent these models can represent and interact with the meaning of language. Emergent behaviors in these highly complex models are unpredictable and research into understanding them is still in early phases, which limits the strength of claims of generality about them. Moreover, while these models demonstrate surprising emergent reasoning capabilities on many tasks, there are still many limitations in generalization that overall high benchmark performance can hide. We review the literature showcasing limitations of LLMs as well as techniques to mitigate or overcome these challenges, while also highlighting how fundamental problems in model evaluation may prevent true claims of generality. We conclude with a practical section on using LLMs on real world problems while appropriately evaluating and validating these tasks to prevent or mitigate the downstream impacts of incorrect results.

97 MATHEMATICS AND COMPUTING↗
DOE OSTI2025

Develop a Simulation Framework for Understanding Physico-Chemical Processes and Optimization of CHZ' Plastic Thermolyze Technology (CRADA Final Report)

Thermolyzer(TM) technology is a third-generation, multi-reactor, oxygen-free, low pressure, slow pyrolysis process. It has proven to be remarkably versatile in successfully processing all types of hydrocarbon waste. These include all seven types of plastics, tires, auto shredder residue, carpet, electronic waste, and composites. Working with the IACMI, Thermolyzer(TM) technology has recovered the glass and carbon fibers from wind turbine blades for reuse into new applications. Its primary output is a synthesis gas that is rich in H 2 , CH 4 , and small amounts of C 2-4 aliphatic hydrocarbons. Other components are CO and CO 2 . Three unique features set Thermolyzer(TM) apart from other pyrolysis systems: the syngas is clean enough to run in Siemens or Solar gas turbines or IC engines without fouling, clean, salable, byproducts are produced, and CO 2 emissions are lower than natural gas power plants. Research results have been obtained from an operating 7 ton/day facility. A 44 ton/day plant has successfully operated as noted above. The process works because it makes use of the recoverable embodied energy in the feedstock. Surprisingly, a pound of some plastics has the same recoverable energy content as a pound of gasoline. Thus, it is imperative to develop a process that economically converts scrap plastics (including ocean plastics) into energy and thereby conserve the non-renewable fossil fuels for future generations. The Thermolyzer(TM) technology can also be used to create liquid fuels like gasoline and diesel from hydrocarbon wastes. Because of the high hydrogen content of the synthesis gas, hydrogen can be recovered more inexpensively than the current solar or wind energy being used to electrolyze water. That hydrogen can be used for fuel cell powered vehicles or converted into ammonia for agriculture or as a hydrogen storage medium. Extension of this technology to other wastes such as tires, auto shredder residue or wood wastes would expand the circular economy.

36 MATERIALS SCIENCE↗
DOE OSTI2026

Improving perovskite solar module stability by understanding and mitigating scribe-induced chemo-thermomechanical degradation

Recognizing the correlation between the design of perovskite solar module (PSM) and their degradation mechanisms under environmental stressors helps to enable their long- term reliability. This work reports on the impact of the hole-transport layer (HTL) on thermal cyclability and degradation mechanisms of p-i-n PSMs by comparing 3 HTLs: NiO x , MeO-2PACz (a self-assembled monolayer), and a bilayer HTL of MeO-2PACz on NiO x . We observe surprising thermal cyclability from a performance standpoint despite generating clear degradation products. We find that without a fully dense HTL, seemingly insignificant moisture/oxygen ingress through the edge of the PSMs can lead to rapid destabilization of the metal halide perovskite (MHP) layer due to reactions between MHP degradation byproducts and the indium-tin oxide layer. We also show that illumination-induced cation phase segregation is dependent on the presence of an inactive area (regions of the substrate uncovered by the top electrode). As such, we find that a dense interface between the MHP and ITO is necessary for chemically robust PSMs and highlight criteria for testing field-relevant configurations in lab-scale architectures.

14 SOLAR ENERGY↗
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↗