Engineering Papers⌕ Search

SEARCH · Engineering Papers

Results for “constrained learning”

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 595 records · Page 33
All sourcesNASA NTRSDOE OSTI
DOE OSTI2024

Hamiltonian variational ansatz without barren plateaus

Variational quantum algorithms, which combine highly expressive parameterized quantum circuits (PQCs) and optimization techniques in machine learning, are one of the most promising applications of a near-term quantum computer. Despite their huge potential, the utility of variational quantum algorithms beyond tens of qubits is still questioned. One of the central problems is the trainability of PQCs. The cost function landscape of a randomly initialized PQC is often too flat, asking for an exponential amount of quantum resources to find a solution. This problem, dubbed barren plateaus , has gained lots of attention recently, but a general solution is still not available. In this paper, we solve this problem for the Hamiltonian variational ansatz (HVA), which is widely studied for solving quantum many-body problems. After showing that a circuit described by a time-evolution operator generated by a local Hamiltonian does not have exponentially small gradients, we derive parameter conditions for which the HVA is well approximated by such an operator. Based on this result, we propose an initialization scheme for the variational quantum algorithms and a parameter-constrained ansatz free from barren plateaus.

Physics↗
NASA NTRS2011

Reactive Oxygen Species on the Early Earth and Survival of Bacteria

An oxygen-rich atmosphere appears to have been a prerequisite for complex, multicellular life to evolve on Earth and possibly elsewhere in the Universe. However it remains unclear how free oxygen first became available on the early Earth. A potentially important, and as yet poorly constrained pathway, is the production of oxygen through the weathering of rocks and release into the near-surface environment. Reactive Oxygen Species (ROS), as precursors to molecular oxygen, are a key step in this process, and may have had a decisive impact on the evolution of life, present and past. ROS are generated from minerals in igneous rocks during hydrolysis of peroxy defects, which consist of pairs of oxygen anions oxidized to the valence state -1 and during (bio) transformations of iron sulphide minerals. ROS are produced and consumed by intracellular and extracellular reactions of Fe, Mn, C, N, and S species. We propose that, despite an overall reducing or neutral oxidation state of the macroenvironment and the absence of free O2 in the atmosphere, organisms on the early Earth had to cope with ROS in their microenvironments. They were thus under evolutionary pressure to develop enzymatic and other defences against the potentially dangerous, even lethal effects of oxygen and its derived ROS. Conversely it appears that microorganisms learned to take advantage of the enormous reactive potential and energy gain provided by nascent oxygen. We investigate how oxygen might be released through weathering. We test microorganisms in contact with rock surfaces and iron sulphides. We model bacteria such as Deionococcus radiodurans and Desulfotomaculum, Moorella and Bacillus species for their ability to grow or survive in the presence of ROS. We examine how early Life might have adapted to oxygen.

Balk, Melikea↗
DOE OSTI2024

Candidate ferroelectrics via ab initio high-throughput screening of polar materials

Ferroelectrics are a class of polar and switchable functional materials with diverse applications, from microelectronics to energy conversion. Computational searches for new ferroelectric materials have been constrained by accurate prediction of the polarization and switchability with electric field, properties that, in principle, require a comparison with a nonpolar phase whose atomic-scale unit cell is continuously deformable from the polar ground state. For most polar materials, such a higher-symmetry nonpolar phase does not exist or is unknown. Here, we introduce a general high-throughput workflow that screens polar materials as potential ferroelectrics. We demonstrate our workflow on 1978 polar structures in the Materials Project database, for which we automatically generate a nonpolar reference structure using pseudosymmetries, and then compute the polarization difference and energy barrier between polar and nonpolar phases, comparing the predicted values to known ferroelectrics. Focusing on a subset of 182 potential ferroelectrics, we implement a systematic ranking strategy that prioritizes candidates with large polarization and small polar-nonpolar energy differences. To assess stability and synthesizability, we combine information including the computed formation energy above the convex hull, the Inorganic Crystal Structure Database id number, a previously reported machine learning-based synthesizability score, and ab initio phonon band structures. To distinguish between previously reported ferroelectrics, materials known for alternative applications, and lesser-known materials, we combine this ranking with a survey of the existing literature on these candidates through Google Scholar and Scopus databases, revealing ~130 promising materials uninvestigated as ferroelectric. Our workflow and large-scale high-throughput screening lays the groundwork for the discovery of novel ferroelectrics, revealing numerous candidates materials for future experimental and theoretical endeavors.

36 MATERIALS SCIENCE↗
DOE OSTI2024

Neural network based emulation of galaxy power spectrum covariances: A reanalysis of BOSS DR12 data

We train neural networks to quickly generate redshift-space galaxy power spectrum covariances from a given parameter set (cosmology and galaxy bias). This covariance emulator utilizes a combination of traditional fully connected network layers and transformer architecture to accurately predict covariance matrices for the high redshift, north galactic cap sample of the BOSS DR12 galaxy catalog. We run simulated likelihood analyses with emulated and brute-force computed covariances, and we quantify the network’s performance via two different metrics: (1) difference in Χ 2 and (2) likelihood contours for simulated BOSS DR 12 analyses. We find that the emulator returns excellent results over a large parameter range. We then use our emulator to perform a reanalysis of the BOSS HighZ NGC galaxy power spectrum, and find that varying covariance with cosmology along with the model vector produces Ω m = $0.27⁢6$$^{+0.013}_{–0.015}$, H 0 = 70.2 ± 1.9 km/s/Mpc, and σ 8 = $0.67⁢4$$^{+0.058}_{–0.077}$. These constraints represent an average 0.46⁢σ shift in best-fit values and a 5% increase in constraining power compared to fixing the covariance matrix (Ω m = 0.293 ± 0.017, H 0 = 70.3 ± 2.0 km/s/Mpc, σ 8 = $0.70⁢2$$^{+0.063}_{–0.075}$). As a result, this work demonstrates that emulators for more complex cosmological quantities than second-order statistics can be trained over a wide parameter range at sufficiently high accuracy to be implemented in realistic likelihood analyses.

79 ASTRONOMY AND ASTROPHYSICS↗
DOE OSTI2020

A Hybrid Deep Learning Approach to Cosmological Constraints from Galaxy Redshift Surveys

We present a deep machine learning (ML)–based technique for accurately determining σ g and Ω m from mock 3D galaxy surveys. The mock surveys are built from the AbacusCosmos suite of N -body simulations, which comprises 40 cosmological volume simulations spanning a range of cosmological parameter values, and we account for uncertainties in galaxy formation scenarios through the use of generalized halo occupation distributions (HODs). We explore a trio of ML models: a 3D convolutional neural network (CNN), a power spectrum–based fully connected network, and a hybrid approach that merges the two to combine physically motivated summary statistics with flexible CNNs. We describe best practices for training a deep model on a suite of matched-phase simulations, and we test our model on a completely independent sample that uses previously unseen initial conditions, cosmological parameters, and HOD parameters. Despite the fact that the mock observations are quite small (~0.07 h -3 Gpc 3 ) and the training data span a large parameter space (six cosmological and six HOD parameters), the CNN and hybrid CNN can constrain estimates of σ g and Ω m to ~3% and ~4%, respectively.

79 ASTRONOMY AND ASTROPHYSICS↗
DOE OSTI2026

WellPINN: Accurate Well Representation for Transient Fluid Pressure Diffusion in Subsurface Reservoirs With Physics‐Informed Neural Networks

Accurate representation of pumping wells is essential for reliable reservoir characterization and simulation of operational scenarios in subsurface flow models. Physics-informed neural networks (PINNs) are emerging as a promising alternative to numerical models for reservoir modeling, offering seamless integration of monitoring data and governing physical equations. However, existing PINN-based studies face major challenges in capturing fluid pressure near wells when using a source/sink term, particularly during the early stages after pumping begins. We address this problem by introducing WellPINN, a workflow in which an initially trained PINN infers fluid pressure across the entire reservoir domain using a large equivalent well radius. This initial PINN solution is then locally refined around the well by a set of subdomain PINNs that are trained for smaller equivalent well radii. Continuity across these subdomain interfaces as well as at the initial condition is ensured by hard-constraining each PINN on its subdomain boundary. Our results demonstrate WellPINN as the first workflow of its kind to focus on accurate inference of fluid pressure from pumping rates throughout the entire injection period, significantly advancing the potential of PINNs for inverse modeling and operational scenario simulations. All data and code for this paper are openly available at https://doi.org/10.20350/DIGITALCSIC/17260.

58 GEOSCIENCES↗
NASA NTRS2019

Supporting Crew Autonomy in Deep Space Exploration: Preliminary Onboard Capability Requirements and Proposed Research Questions. Technical Report of the Autonomous Crew Operations Technical Interchange Meeting

Communication delays are a critical challenge posed by long duration deep space exploration. Space missions historically have relied on an ever-present Mission Control Center (MCC) to direct operations in near real-time. As unanticipated anomalies that defeat fault detection and resolution systems do arise, the lack of real-time communication will significantly weaken what the MCC support represents: a reliable safety net for the flight crew through its deep and diverse areas of expertise and investigative resources. As a consequence, future space vehicles and habitats need to be equipped with capabilities to support the flight crew to operate with little or no ground support. Considerations must be given to vehicle and mission designs that will fortify the traditionally ground-centered safety net and forge new support systems, when communication delays exist. In August 2018, NASA’s Human Research Program, through its Human Factors and Behavioral Performance Element, convened a Technical Interchange Meeting (TIM) on Autonomous Crew Operations at NASA Ames Research Center. The goal of the meeting was to gather input from NASA centers, industry, academia, and branches of the Department of Defense (DoD) to address how intelligent technologies can be applied to augment onboard capabilities to support crew anomaly response. The TIM featured 24 presentations by 29 speakers and hosted a total of 59 attendees, including 43 from 5 NASA centers (Ames, Johnson, Langley, Marshall, and Jet Propulsion Lab) and 4 from the DoD (3 from Army Research Lab and 1 from Naval Postgraduate School), with remaining attendees from academia (e.g., UC Davis, CMU) and industry (e.g., IBM, Siemens). Discussions were centered around three themes: standards and guidelines, lessons learned in analog environments, and technologies. To help provide a framework for discussion, a concept matrix describing anomaly response processes was created prior to the TIM (Figure 1, page 6). The matrix captures the steps involved (monitoring and detection, diagnosis, solution development and evaluation, solution implementation and verification, resolution documentation) as well as the resources and capabilities required to support these steps (data, knowledge, analysis, synthesis, resource management). A wallpaper size printout of the matrix was utilized at the TIM to solicit attendee inputs along the three themes; the activity garnered 108 submissions of ideas. Overall, what emerged from TIM discussions was a picture of mismatch between crew anomaly response needs and support that can be provided by existing intelligent technologies. The needs are broad, spanning multiple steps and processes/resources, with many of which lacking support from existing technologies, such as knowledge management throughout the steps of problem solving (especially in resolution documentation) and manpower management. The solutions provided by existing intelligent technologies are specific to the steps/processes that they are designed to support and constrained to solving only problems similar to those that have occurred before. What is lacking from technologies is typically made up by humans, specifically their complex critical thinking, creative problem solving, and domain expertise. In the end, the TIM highlighted the pressing need to support responses to onboard anomalies during autonomous crew operations, particularly those that have eluded the system tests, inspection, and other assurance processes. Such anomalies can potentially threaten crew and vehicle safety, as well as significantly impact overall operations with additional workload. These fairly rare events are difficult to anticipate and prepare for, given the state-of-the-art in intelligent technologies. This is true even for anomalies that stem from “unknown knowns”—cases in which there is sufficient external information to characterize the problem but the overall pattern fails to be recognized by the problem solver, or in which the internal knowledge needed to solve a problem is held tacitly and potentially accessible by the problem solver but not articulated. It follows that the ability to tackle anomalies lies not only with the availability of relevant information and knowledge but also their accessibility in times of need. To that end, we propose research questions along the following three broad themes: • How intelligent technologies can help make relevant knowledge and information available? • How intelligent technologies can help make relevant knowledge and information accessible? • How intelligent technologies can help support the crew operating as a team in anomaly response processes?

autonomous crew operations↗
DOE OSTI2024

Enhancing Solar Power Forecasting with Regularized Constrained Quantile Regression Averaging and Bootstrapping Techniques

Probabilistic solar power forecasting (SPF) plays an essential role in optimizing power-grid operations by quantifying the forecast uncertainty. To improve the accuracy and robustness of probabilistic SPF, this paper introduces the regularized constrained quantile regression averaging (rCQRA) method to combine outputs from multiple PSPF models. In addition, a bootstrapping method was used to quantify model uncertainty, providing insights into the reliability and significance of each ensemble component. To evaluate its efficacy, the proposed rCQRA method is used to integrate four PSPF methods. The resulting SPF models are trained and validated using a real-world six-year dataset from a rooftop solar plant in the USA. The performance of the proposed rCQRA method is evaluated and compared with two benchmark methods under three categories of weather conditions. It is shown that the rCQRA method has superior performance in its forecast reliability, sharpness, and accuracy.

Ensemble learning, probabilistic solar power forec↗
NASA NTRS2006

Deep X-ray and UV Surveys of Galaxies with Chandra, XMM-Newton, and GALEX

Only with the deepest Chandra surveys has X-ray emission from normal and star forming galaxies (as opposed to AGN, which dominate the X-ray sky) been accessible at cosmologically interesting distances. The X-ray emission from accreting binaries provide a critical glimpse into the binary phase of stellar evolution and studies of the hot gas reservoir constrain past star formation. UV studies provide important, sensitive diagnostics of the young star forming populations and provide the most mature means for studying galaxies at 2 < zeta < 4. This talk will review current progress on studying X-ray emission in concert with UV emission from normal/star-forming galaxies at higher redshift. We will also report on our new, deep surveys with GALEX and XMM-Newton in the nearby Coma cluster. These studies are relevant to DEEP06 as Coma is the nearest rich cluster of galaxies and provides an important benchmark for high-redshift studies in the X-ray and UV wavebands. The 30 ks GALEX (note: similar depth to the GALEX Deep Imaging Survey) and the 110 ks XMM observations provide extremely deep coverage of a Coma outskirts field, allowing the construction of the UV and X-ray luminosity function of galaxies and important constraints on star formation scaling relations such as the X-ray-Star Formation Rate correlation and the X-ray/Stellar Mass correlation. We will discuss what we learn from these deep observations of Coma, including the recently established suppression of the X-ray emission from galaxies in the Coma outskirts that is likely associated with lower levels of past star formation and/or the results of tidal gas stripping.

Hornschemeier, Ann↗
DOE OSTI2023

Uniform-in-phase-space data selection with iterative normalizing flows

Improvements in computational and experimental capabilities are rapidly increasing the amount of scientific data that are routinely generated. In applications that are constrained by memory and computational intensity, excessively large datasets may hinder scientific discovery, making data reduction a critical component of data-driven methods. Datasets are growing in two directions: the number of data points and their dimensionality. Whereas dimension reduction typically aims at describing each data sample on lower-dimensional space, the focus here is on reducing the number of data points. A strategy is proposed to select data points such that they uniformly span the phase-space of the data. The algorithm proposed relies on estimating the probability map of the data and using it to construct an acceptance probability. An iterative method is used to accurately estimate the probability of the rare data points when only a small subset of the dataset is used to construct the probability map. Instead of binning the phase-space to estimate the probability map, its functional form is approximated with a normalizing flow. Therefore, the method naturally extends to high-dimensional datasets. The proposed framework is demonstrated as a viable pathway to enable data-efficient machine learning when abundant data are available.

97 MATHEMATICS AND COMPUTING↗
DOE OSTI2021

EXPRES. II. Searching for Planets around Active Stars: A Case Study of HD 101501

By controlling instrumental errors to below 10 cm s{sup −1}, the EXtreme PREcision Spectrograph (EXPRES) allows for a more insightful study of photospheric velocities that can mask weak Keplerian signals. Gaussian processes (GP) have become a standard tool for modeling correlated noise in radial velocity data sets. While GPs are constrained and motivated by physical properties of the star, in some cases they are still flexible enough to absorb unresolved Keplerian signals. We apply GP regression to EXPRES radial velocity measurements of the 3.5 Gyr old chromospherically active Sun-like star, HD 101501. We obtain tight constraints on the stellar rotation period and the evolution of spot distributions using 28 seasons of ground-based photometry, as well as recent Transiting Exoplanet Survey Satellite data. Light-curve inversion was carried out on both photometry data sets to reveal the spot distribution and spot evolution timescales on the star. We find that the >5 m s{sup −1} rms radial velocity variations in HD 101501 are well modeled with a GP stellar activity model without planets, yielding a residual rms scatter of 45 cm s{sup −1}. We carry out simulations, injecting and recovering signals with the GP framework, to demonstrate that high-cadence observations are required to use GPs most efficiently to detect low-mass planets around active stars like HD 101501. Sparse sampling prevents GPs from learning the correlated noise structure and can allow it to absorb prospective Keplerian signals. We quantify the moderate to high-cadence monitoring that provides the necessary information to disentangle photospheric features using GPs and to detect planets around active stars.

46 INSTRUMENTATION RELATED TO NUCLEAR SCIENCE AND ↗
NASA NTRS2004

Cosmic Ray Propagation Models

Astrophysics of cosmic rays and gamma rays depends very much on the quality of the data, which become increasingly accurate each year and therefore more constraining. While direct measurements of cosmic rays are possible in only one location on the outskirts of the Milky Way, the Galactic diffuse gamma-ray emission provides insights into the spectra of cosmic rays in distant locations, therefore complementing the local cosmic-ray studies. This connection, however, requires extensive modeling and is yet to be explored in detail. The GUST mission, which is scheduled for launch in 2007 and is capable of measuring gamma-rays in the range 20 MeV - 300 GeV, will change the status quo dramatically. Galactic diffuse gamma-ray emission gathered by GUST will require adequate theoretical models. The efforts will be rewarded by the wealth of information on cosmic ray spectra and fluxes in remote locations. In its turn, a detailed cosmic ray propagation model will provide a reliable basis for other studies such as search for dark matter signals in cosmic rays and diffuse gamma rays, spectrum and origin of the extragalactic gamma-ray emission, theories of nucleosynthesis and evolution of elements etc. In this talk, I will discuss what we can learn studying the cosmic ray propagation and diffuse gamma-ray emission.

Moskalenko, I. V.↗
NASA NTRS2022

Mars 2020 Entry, Descent, and Landing Software Implementation

On February 18th, 2021, the Mars 2020 project's Perseverance Rover successfully touched down on the Martian surface after nearly eight years of development. The Mars 2020 Entry, Descent, and Landing (EDL) System largely leveraged heritage from the Mars Science Laboratory (MSL) EDL System while employing targeted technological advancements. The landing process is autonomously directed by a software behavior implemented in the rover's primary flight computer called the EDL Timeline that assumes control of the vehicle six days before atmospheric entry. In addition to performing the critical function of landing the rover on the Martian surface, the EDL timeline behavior must co-exist in a non-partitioned software and system environment with other high-level functions that accomplish the goals for the rest of the mission. Due to the criticality of EDL, the potential for loss of mission, and a need for complete system autonomy, the standard for how the EDL Timeline interacts with other functions in the system is highly constrained. This paper first walks through the basics of the EDL Timeline mechanics and how the behavior is designed to account for internal system variations and environmental unknowns. It then summarizes the interactions between the EDL timeline and other high-level system behaviors like spacecraft mode transitions and system fault protection, focusing on the complications that arise when passing spacecraft control between executive functions. Although the MSL-inherited EDL System is reliable and capable, targeted updates and a thorough verification and validation program were required for Mars 2020. This paper discusses changes made to close vulnerabilities discovered during both MSL and Mars 2020 development cycles, landing system capability enhancements that were enabling for Mars 2020's mission, and how these updates were integrated with the heritage system. It then describes how both analysis and testing campaigns were utilized to verify and validate all aspects of EDL and system behaviors that run during the six days before landing, as well as the operational workarounds that were needed to address problems found during the development and commissioning process. Finally, this paper imparts lessons learned from Mars 2020 EDL development, implementation, and operations, emphasizing how systems designed to conduct time-critical mission events with low margin of error can be improved in the future.

Stehura, Aaron↗
DOE OSTI2022

A Search for Correlated Low-energy Electron Antineutrinos in KamLAND with Gamma-Ray Bursts

We present the results of a time-coincident event search for low-energy electron antineutrinos in the KamLAND detector with gamma-ray bursts (GRBs) from the Gamma-ray Coordinates Network and Fermi Gamma-ray Burst Monitor. Using a variable coincidence time window of ±500 s plus the duration of each GRB, no statistically significant excess above the background is observed. We place the world’s most stringent 90% confidence level upper limit on the electron antineutrino fluence below 17.5 MeV. Assuming a Fermi–Dirac neutrino energy spectrum from the GRB source, we use the available redshift data to constrain the electron antineutrino luminosity and effective temperature.

79 ASTRONOMY AND ASTROPHYSICS↗
NASA NTRSDate not supplied

In-Situ XRD/XRF to Support Life Detection on Mars

X-ray diffraction / X-ray fluorescence (XRD/XRF) analysis provides the most comprehensive mineralogical / compositional characterization of rocks and soils of any flight-capable technique. XRD data provide quantitative mineralogy (including abundance of X-ray amorphous materials) and crystal chemistry (structure, elemental composition and valence state), and XRF data provide complimentary major, minor, and some trace element abundances. Both types of data are important in evaluating habitability (environment of formation) and biosignature preservation/degradation (post-depositional diagenetic change). Whether or not a relict biosignature is detected, the mineral assemblage and its geochemistry can be used to determine the habitability of an ancient environment (e.g., salinity, pH, temperature), and to identify potential sources of energy for life (e.g., elements in different redox states). In this respect, a null result (a habitable environment lacking evidence of life) can play an important role in constraining the parameters of the search. Conversely, diagenetic alteration (taphonomic change) resulting from post-depositional variations in temperature, pressure or fluid chemistry can preserve evidence of biogenicity, erase such evidence completely or indeed can provide for post-depositional habitable conditions in the subsurface. XRD / XRF data are critical to these determinations. The CheMin instrument on the Mars Science Laboratory (MSL) Curiosity rover is the first XRD instrument flown in space. CheMin operates in transmission geometry with a Co X-ray source to minimize fluorescence from iron. Diffracted photons are collected with an energy-sensitive charge-coupled device (CCD). The position of the diffracted photons provides structural information for minerals, whereas the energy of sample-generated X-ray fluorescence photons provides elemental information, though these XRF data are qualitative. Mineralogical data from the CheMin XRD identified the three circumstances above: habitable depositional environments (e.g., Yellowknife Bay), habitable subsurface/diagenetic environments (e.g., throughout the Murray formation), and diagenetic conditions that may destroy evidence of habitability (e.g., oxidative and acidic environments at Vera Rubin ridge). Technological advances in X-ray technology and lessons learned from the operation of CheMin on Mars have resulted in a next-generation XRD/XRF, called CheMinX. Replacement of CheMin’s CCD with an array of hybrid pixel detectors and improvements in focusing optics dramatically decrease analysis time (15 minutes vs. 22 hours for MSL-CheMin) and result in a better angular resolution (0.18 vs. 0.30 °2θ for MSL-CheMin). This increased resolution improves mineral detection, including discrimination between types of pyroxenes, which is not possible with MSL-CheMin data. The hybrid pixel detectors do not require cooling like the MSL-CheMin CCD, therefore reducing the power needed to operate CheMinX. CheMinX has a silicon-drift detector (SDD) to measure fluoresced photons, enabling the quantification of major, minor, and some trace elements via XRF. XRD/XRF data are collected simultaneously in CheMinX, obviating the need for multiple compositional instruments. The CheMinX design also improves upon MSL-CheMin’s sample handling. Instead of sample cells on wheel, which are often not reusable and add complexity in commanding the instrument, CheMinX has single-use cells in a cartridge/dispenser configuration. Because of these improvements, CheMinX is an ideal instrument for Discovery-class life-detection missions, including Mars Life Explorer that was recommended for development in the Planetary Science and Astrobiology Decadal Survey 2023-2032.

E B Rampe↗
NASA NTRS2004

Propagation of Cosmic Rays

Astrophysics of cosmic rays and gamma rays depends very much on the quality of the data, which become increasingly accurate each year and therefore more constraining. While direct measurements of cosmic rays are possible in only one location on the outskirts of the Milky Way, the Galactic diffuse gamma-ray emission provides insights into the spectra of cosmic rays in distant locations, therefore complementing the local cosmic-ray studies. This connection, however, requires extensive modeling and is yet to be explored in detail. The GLAST mission, which is scheduled for launch in 2007 and is capable of measuring gamma-rays in the range 20 MeV - 300 GeV, will change the status quo dramatically. The detailed spectra and skymaps of the Galactic diffuse gamma-ray emission gathered by GLAST will require adequate theoretical models. The efforts will be rewarded by the wealth of information on cosmic ray spectra and fluxes in remote locations. In its turn, a detailed cosmic ray propagation model will provide a reliable basis for other studies such as search for dark matter signals in cosmic rays and diffuse gamma rays, spectrum and origin of the extragalactic gamma-ray'emission, theories of nucleosynthesis and evolution of elements etc. In this talk, I will discuss what we can learn studying the cosmic ray propagation and diffuse gamma-ray emission.

Moskalenko, I. V.↗
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 OSTI2022

Tackling the Challenges in Scene Graph Generation With Local-to-Global Interactions

In this work, we seek new insights into the underlying challenges of the scene graph generation (SGG) task. Quantitative and qualitative analysis of the visual genome (VG) dataset implies: 1) ambiguity: even if interobject relationship contains the same object (or predicate), they may not be visually or semantically similar; 2) asymmetry: despite the nature of the relationship that embodied the direction, it was not well addressed in previous studies; and 3) higher-order contexts: leveraging the identities of certain graph elements can help generate accurate scene graphs. Motivated by the analysis, we design a novel SGG framework, Local-to-global interaction networks (LOGINs). Locally, interactions extract the essence between three instances of subject, object, and background, while baking direction awareness into the network by explicitly constraining the input order of subject and object. Globally, interactions encode the contexts between every graph component (i.e., nodes and edges). Finally, Attract and Repel loss is utilized to fine-tune the distribution of predicate embeddings. By design, our framework enables predicting the scene graph in a bottom-up manner, leveraging the possible complementariness. To quantify how much LOGIN is aware of relational direction, a new diagnostic task called Bidirectional Relationship Classification (BRC) is also proposed. Overall, experimental results demonstrate that LOGIN can successfully distinguish relational direction than existing methods (in BRC task), while showing state-of-the-art results on the VG benchmark (in SGG task).

97 MATHEMATICS AND COMPUTING↗