Results for “Computational pipelines” — Page 46

NASA NTRSDate not supplied

The NASA Open Science Data Repository: Biomedical Fair Data, Analysis Tools, User Communities, Publications, and Discoveries for Deep Space Missions

Increased biomedical risks and challenges associated with deep space missions require new knowledge discovery, new health countermeasures, and development of novel ecosystems, life support, crop production, and biomedical support capabilities. To meet NASA’s Moon to Mars strategic program goals for Human and Biological Sciences, findable, accessible, interoperable, reusable (FAIR), and maximally open-access data is going to be required to enable humanity to thrive in deep space. Indeed, this cornerstone perspective on FAIR and maximally open access data was also recommended in the recent 2023-2032 Decadal Survey from the National Academies of Sciences, Engineering, and Medicine. The NASA Open Science Data Repository (OSDR) is a maximally open access and FAIR database, and meets various scientific, technical, and operational spaceflight needs. It offers public users and submitters the ability to upload, download, search, share, analyze, and visualize data across ‘omics, physiological, phenotypic, behavioral, bioimaging, video, and environmental monitoring telemetry datasets. OSDR includes NASA GeneLab, NASA Ames Life Sciences Data Archive, and the NASA Biological Institutional Scientific Collection. OSDR has >455 studies with datasets from model organisms and non-NASA human astronauts. There are ~12 datasets from the Inspiration 4 (I4) mission, spanning metagenomics, comprehensive metabolic panels, clonal hematopoiesis, spatial transcriptomics, proteomics, and cytokine panels. In the interest of data privacy, two I4 datasets have raw FASTQ and FASTA files relating to the epitranscriptome, and a new request feature is live in OSDR (with a backend review process established) which was developed based on industry norms. OSDR also recently began a collaboration with the European Space Agency (ESA) to scientifically curate and make available >200 terabytes of human and model organism space-relevant data. The OSDR submission portal is designed to ingest and curate ~25 ‘omics assay data types, and ~50 physiological-phenotypic-imaging assay data types, spanning ultrasonography, micro-computed tomography, histology, morphometric photography, rebound tonometry, gait analysis, optical coherence tomography, novel object recognition, flow cytometry, and immunohistochemistry. A suite of analysis tools are available for OSDR users including: 1) an Environmental Data Application to compare radiation, CO2, relative humidity, temperature, and other telemetry across missions and subjects, 2) the RadLab database, a collaboration between NASA, ESA, the German and Italian Space Agencies, and the Bulgarian Academy of Sciences, which compiles radiation measurements relevant to human spaceflight and provides tools for accessing and manipulating the data, and 3) a Multi-study visualization tool which enables users to look across and combine GeneLab’s omics datasets across different experiments and missions. There are ~600 volunteer OSDR Analysis Working Group (AWG) members who: 1) provide feedback on scientific standards for reuse (subject and assay metadata; processing pipelines; dataset formats and uniformed structures for machine-readability), and 2) collaborate to mine-reuse OSDR data conducting scientific analysis. OSDR has enabled 60 publications as of September 2023, many directly from AWG collaborations most notably the Cell Press package in 2020. Lastly, there are at least 15 articles which mine OSDR data part of a package of ~50 articles across Nature Portfolio with research stemming from I4, the Japan Aerospace Exploration Agency, NASA Space Biology, and the NASA Human Research Program.

space biology↗

NASA NTRSDate not supplied

NASA Open Science Data Repository: Biomedical FAIR Data, Analysis Tools, User Communities, and Discoveries for Deep Space Missions

Increased biomedical risks and challenges associated with deep space missions require new knowledge discovery, new health countermeasures, and development of novel ecosystems, life support, crop production, and biomedical support capabilities. To meet NASA’s Moon to Mars strategic program goals for Human and Biological Sciences, findable, accessible, interoperable, reusable (FAIR), and maximally open-access data is going to be required to enable humanity to thrive in deep space. Indeed, this cornerstone perspective on FAIR and maximally open access data was also recommended in the recent 2023-2032 Decadal Survey from the National Academies of Sciences, Engineering, and Medicine. The NASA Open Science Data Repository (OSDR) is a maximally open access and FAIR database, and meets various scientific, technical, and operational spaceflight needs. It offers public users and submitters the ability to upload, download, search, share, analyze, and visualize data across ‘omics, physiological, phenotypic, behavioral, bioimaging, video, and environmental monitoring telemetry datasets. OSDR includes NASA GeneLab, NASA Ames Life Sciences Data Archive, and the NASA Biological Institutional Scientific Collection. OSDR has >455 studies with datasets from model organisms and non-NASA human astronauts. There are ~12 datasets from the Inspiration 4 (I4) mission, spanning metagenomics, comprehensive metabolic panels, clonal hematopoiesis, spatial transcriptomics, proteomics, and cytokine panels. In the interest of data privacy, two I4 datasets have raw FASTQ and FASTA files relating to the epitranscriptome, and a new request feature is live in OSDR (with a backend review process established) which was developed based on industry norms. OSDR also recently began a collaboration with the European Space Agency (ESA) to scientifically curate and make available >200 terabytes of human and model organism space-relevant data. The OSDR submission portal is designed to ingest and curate ~25 ‘omics assay data types, and ~50 physiological-phenotypic-imaging assay data types, spanning ultrasonography, micro-computed tomography, histology, morphometric photography, rebound tonometry, gait analysis, optical coherence tomography, novel object recognition, flow cytometry, and immunohistochemistry. A suite of analysis tools are available for OSDR users including: 1) an Environmental Data Application to compare radiation, CO2, relative humidity, temperature, and other telemetry across missions and subjects, 2) the RadLab database, a collaboration between NASA, ESA, the German and Italian Space Agencies, and the Bulgarian Academy of Sciences, which compiles radiation measurements relevant to human spaceflight and provides tools for accessing and manipulating the data, and 3) a Multi-study visualization tool which enables users to look across and combine GeneLab’s omics datasets across different experiments and missions. There are ~600 volunteer OSDR Analysis Working Group (AWG) members who: 1) provide feedback on scientific standards for reuse (subject and assay metadata; processing pipelines; dataset formats and uniformed structures for machine-readability), and 2) collaborate to mine-reuse OSDR data conducting scientific analysis. OSDR has enabled 60 publications as of September 2023, many directly from AWG collaborations most notably the Cell Press package in 2020. Lastly, there are at least 15 articles which mine OSDR data part of a package of ~50 articles across Nature Portfolio with research stemming from I4, the Japan Aerospace Exploration Agency, NASA Space Biology, and the NASA Human Research Program.

open access↗

DOE OSTI2026

AI-Ready Data Pilot Project Report

The proliferation of artificial intelligence in scientific research has created an urgent need to define "AI-ready data" for researchers and, more importantly, provide resources to help them produce AI-ready data. At Pacific Northwest National Laboratory, we conducted a pilot study with three data scientists evaluating three CSV datasets from different scientific domains, followed by semi-structured interviews capturing assessment practices. Our findings reveal that AI-readiness evaluation is intuition-based, with practitioners asking "How fast can I go from raw data to my machine learning pipeline?" Data scientists consistently prioritized workflow efficiency, human interpretability, and quality stewardship signals. From these insights, we developed a practical evaluation framework comprising data requirements, metadata standards, and validation tests that provides actionable criteria for producing and curating AI-ready datasets, addressing the gap between theoretical understanding and practical implementation.

97 MATHEMATICS AND COMPUTING↗

NASA NTRS1993

Data fusion with artificial neural networks (ANN) for classification of earth surface from microwave satellite measurements

A data fusion system with artificial neural networks (ANN) is used for fast and accurate classification of five earth surface conditions and surface changes, based on seven SSMI multichannel microwave satellite measurements. The measurements include brightness temperatures at 19, 22, 37, and 85 GHz at both H and V polarizations (only V at 22 GHz). The seven channel measurements are processed through a convolution computation such that all measurements are located at same grid. Five surface classes including non-scattering surface, precipitation over land, over ocean, snow, and desert are identified from ground-truth observations. The system processes sensory data in three consecutive phases: (1) pre-processing to extract feature vectors and enhance separability among detected classes; (2) preliminary classification of Earth surface patterns using two separate and parallely acting classifiers: back-propagation neural network and binary decision tree classifiers; and (3) data fusion of results from preliminary classifiers to obtain the optimal performance in overall classification. Both the binary decision tree classifier and the fusion processing centers are implemented by neural network architectures. The fusion system configuration is a hierarchical neural network architecture, in which each functional neural net will handle different processing phases in a pipelined fashion. There is a total of around 13,500 samples for this analysis, of which 4 percent are used as the training set and 96 percent as the testing set. After training, this classification system is able to bring up the detection accuracy to 94 percent compared with 88 percent for back-propagation artificial neural networks and 80 percent for binary decision tree classifiers. The neural network data fusion classification is currently under progress to be integrated in an image processing system at NOAA and to be implemented in a prototype of a massively parallel and dynamically reconfigurable Modular Neural Ring (MNR).

Lure, Y. M. Fleming↗

DOE OSTI2025

Multistage economic MPC for systems with a cyclic steady state: A gas network case study

Multistage model predictive control (MPC) provides a robust control strategy for dynamic systems with uncertainties and a setpoint tracking objective. Moreover, extending MPC to minimize an economic cost instead of tracking a pre-calculated optimal setpoint improves controller performance. This paper presents a novel multistage economic nonlinear model predictive control (E-NMPC) framework for dynamic systems operating under uncertainty, with specific application to natural gas transmission networks. A key innovation lies in the integration of cyclic steady-state (CSS) constraints within the multistage MPC formulation, enabling the controller to manage periodic operating conditions commonly observed in energy systems. A Lyapunov-based descent condition is enforced to ensure robust stability of the controller. The multistage economic MPC framework is validated on two gas pipeline case studies, where it successfully minimizes net energy consumption, respects operational constraints under uncertain demand profiles, and guides the network to optimal cyclic operation. The Lyapunov function remains bounded in both case studies, validating the robust stability of multistage E-NMPC.

03 NATURAL GAS↗

DOE OSTI2022

Root system architecture and environmental flux analysis in mature crops using 3D root mesocosms

Current methods of root sampling typically only obtain small or incomplete sections of root systems and do not capture their true complexity. To facilitate the visualization and analysis of full-sized plant root systems in 3-dimensions, we developed customized mesocosm growth containers. While highly scalable, the design presented here uses an internal volume of 45 ft 3 (1.27 m 3 ), suitable for large crop and bioenergy grass root systems to grow largely unconstrained. Furthermore, they allow for the excavation and preservation of 3-dimensional root system architecture (RSA), and facilitate the collection of time-resolved subterranean environmental data. Sensor arrays monitoring matric potential, temperature and CO 2 levels are buried in a grid formation at various depths to assess environmental fluxes at regular intervals. Methods of 3D data visualization of fluxes were developed to allow for comparison with root system architectural traits. Following harvest, the recovered root system can be digitally reconstructed in 3D through photogrammetry, which is an inexpensive method requiring only an appropriate studio space and a digital camera. We developed a pipeline to extract features from the 3D point clouds, or from derived skeletons that include point cloud voxel number as a proxy for biomass, total root system length, volume, depth, convex hull volume and solidity as a function of depth. Ground-truthing these features with biomass measurements from manually dissected root systems showed a high correlation. We evaluated switchgrass, maize, and sorghum root systems to highlight the capability for species wide comparisons. We focused on two switchgrass ecotypes, upland (VS16) and lowland (WBC3), in identical environments to demonstrate widely different root system architectures that may be indicative of core differences in their rhizoeconomic foraging strategies. Finally, we imposed a strong physiological water stress and manipulated the growth medium to demonstrate whole root system plasticity in response to environmental stimuli. Hence, these new “3D Root Mesocosms” and accompanying computational analysis provides a new paradigm for study of mature crop systems and the environmental fluxes that shape them.

59 BASIC BIOLOGICAL SCIENCES↗

DOE OSTI2022

Multiple Mutations Associated with Emergent Variants Can Be Detected as Low-Frequency Mutations in Early SARS-CoV-2 Pandemic Clinical Samples

Genetic analysis of intra-host viral populations provides unique insight into pre-emergent mutations that may contribute to the genotype of future variants. Clinical samples positive for SARS-CoV-2 collected in California during the first months of the pandemic were sequenced to define the dynamics of mutation emergence as the virus became established in the state. Deep sequencing of 90 nasopharyngeal samples showed that many mutations associated with the establishment of SARS-CoV-2 globally were present at varying frequencies in a majority of the samples, even those collected as the virus was first detected in the US. A subset of mutations that emerged months later in consensus sequences were detected as subconsensus members of intra-host populations. Spike mutations P681H, H655Y, and V1104L were detected prior to emergence in variant genotypes, mutations were detected at multiple positions within the furin cleavage site, and pre-emergent mutations were identified in the nucleocapsid and the envelope genes. Because many of the samples had a very high depth of coverage, a bioinformatics pipeline, “Mappgene”, was established that uses both iVar and LoFreq variant calling to enable identification of very low-frequency variants. This enabled detection of a spike protein deletion present in many samples at low frequency and associated with a variant of concern.

60 APPLIED LIFE SCIENCES↗

NASA NTRS1995

The AAMP5/AAMP-FV project

This presentation describes a project, formal verification of the microcode in the AAMP5 microprocessor, conducted to explore how formal techniques for specification and verification could be introduced into an industrial process. Sponsored by the Systems Validation Branch of NASA Langley and by Collins Commercial Avionics, a division of Rockwell International, it was conducted by Collins and the SRI International Computer Science Laboratory. The project consisted of specifying in the PVS language developed by SRI a portion of a Rockwell proprietary microprocessor, the AAMP5, at both the instruction set and register-transfer levels and using the PVS theorem prover to prove the microcode correct for a representative subset of instructions. While this presentation includes a brief technical overview, its emphasis is on the lessons learned in using PVS for an example of this size and the implications for using formal methods in an industrial setting. The central result of this project was to demonstrate the feasibility of formally specifying a commercial microprocessor and the use of mechanical proofs of correctness to verify microcode. This is particularly significant since the AAMP5 was not designed for formal verification, but to provide a more than three fold performance improvement, by pipelining instruction execution, while remaining object code compatible with the earlier AAMP2. As a consequence, the AAMP5 is one of the most complex microprocessors to which formal methods have been applied. Another key result was the discovery of both actual and seeded errors. Two actual microcode errors were discovered and corrected during development of the formal specification, illustrating the value of simply creating a precise specification. Two seeded errors were systematically uncovered while doing correctness proofs. One of these was an actual error that had been discovered after first fabrication but left in the microcode provided to SRI. The other error was designed to be unlikely to be detected by walkthroughs, testing, or simulation. Several other results emerged during the project, including the ease with which practicing engineers became comfortable with PVS, the need for libraries of general purpose theories, the usefulness of formal specification in revealing errors, the natural fit between formal specification and inspections, the difficulty of selecting the best style of specification for a new problem domain, the high level of assurance provided by proofs of correctness, and the need to engineer proof strategies for reuse.

Miller, Steven P.↗

DOE OSTI2020

Panel-Segmentation [SWR-21-18]

Panel-Segmentation contains the scripts for automated metadata extraction of solar PV installations, using satellite imagery coupled with computer vision techniques. In this package, the user can perform the following actions: *Automatically generate a satellite image using a set of lat-long coordinates, and a Google Maps API key. Users would need to set up a Google Cloud account and get a Maps Static API key. Please refer to Setting Up Google Maps Static API Key section for this process. *Perform image segmentation on the satellite image, to locate the solar array(s) in the image on a pixel-by-pixel basis, using an image segmentation model (panel_detection_model.pth). Get classification of the installation (rooftop, ground mounted fixed-tilt or tracking, carport, etc). *Perform azimuth estimation on each solar array cluster in the masked image. *Detect solar panels and get its latitude, longitude, and address within a geographic bounding box through the SOL-Searcher Pipeline. *Detect and calculate hurricane damage on solar installations given pre-hurricane and post-hurricane satellite imagery through the Hurricane Detection Pipeline. *Detect and calculate hail damage on solar installations given satellite imagery through the Hail Detection pipeline. *Convert NOAA MESH (Maximum Estimated Size of Hail) grib2 files into kml or geojson files. *Estimate tilt and azimuth of a solar array by processing USGS LiDAR data for the array’s location.

Edun, Ayobami↗

NASA NTRS1990

On recursive least-squares filtering algorithms and implementations

In many real-time signal processing applications, fast and numerically stable algorithms for solving least-squares problems are necessary and important. In particular, under non-stationary conditions, these algorithms must be able to adapt themselves to reflect the changes in the system and take appropriate adjustments to achieve optimum performances. Among existing algorithms, the QR-decomposition (QRD)-based recursive least-squares (RLS) methods have been shown to be useful and effective for adaptive signal processing. In order to increase the speed of processing and achieve high throughput rate, many algorithms are being vectorized and/or pipelined to facilitate high degrees of parallelism. A time-recursive formulation of RLS filtering employing block QRD will be considered first. Several methods, including a new non-continuous windowing scheme based on selectively rejecting contaminated data, were investigated for adaptive processing. Based on systolic triarrays, many other forms of systolic arrays are shown to be capable of implementing different algorithms. Various updating and downdating systolic algorithms and architectures for RLS filtering are examined and compared in details, which include Householder reflector, Gram-Schmidt procedure, and Givens rotation. A unified approach encompassing existing square-root-free algorithms is also proposed. For the sinusoidal spectrum estimation problem, a judicious method of separating the noise from the signal is of great interest. Various truncated QR methods are proposed for this purpose and compared to the truncated SVD method. Computer simulations provided for detailed comparisons show the effectiveness of these methods. This thesis deals with fundamental issues of numerical stability, computational efficiency, adaptivity, and VLSI implementation for the RLS filtering problems. In all, various new and modified algorithms and architectures are proposed and analyzed; the significance of any of the new method depends crucially on specific application.

Hsieh, Shih-Fu↗

DOE OSTI2025

Geospatial Data Workflow Orchestration and Architecture

In an era characterized by explosive growth in geospatial data, the selection of appropriate technologies for data storage, processing, and orchestration is critical for organizations aiming to maintain competitive advantages. This white paper provides a comprehensive analysis of how Oak Ridge National Laboratory (ORNL) has effectively employed various cloud technologies, including containerized applications, container orchestrators, and workflow orchestrators, to develop robust geospatial data processing solutions. We explore the fundamental concepts behind these technologies and compare multiple deployment models tailored to diverse use cases. Our findings conclude that while Kubernetes has emerged as the preferred platform for truly scalable and fault-tolerant production workflows, the choice of workflow orchestration tool requires careful consideration of team needs, pipeline complexity, and deployment environments. This paper aims to serve as a strategic guide for organizations leveraging geospatial data, articulating the balance between technology choices and practical implementation to enhance workflow efficacy and scalability.

97 MATHEMATICS AND COMPUTING↗

DOE OSTI2025

Event Classifications on DNE2 Main Experiment Data using a Convolutional Neural Network Ensemble

The Dynamic Networks (DN) Experiment for FY24 (DNE2) is an experiment within DN with the goal of quantitatively evaluating the effectiveness of solutions developed so far by various researchers under the Low Yield Nuclear Monitoring (LYNM) program using a shared set of metrics and datasets. A key component of this experiment is the mimicking of a signature processing pipeline, and comparing currently accepted and standard-use processing methods to more state-of-the-art processes developed under DN. In this work, we focus specifically on the Event Characterization (EC) Focus Area (FA) of the pipeline, where a seismic event’s magnitude, yield and class are identified. We use Deep Learning (DL) to classify the type of events being processed as either earthquakes (EQs) or explosions (EXs) for three iterations of experiment datasets. The model is noticeably more confident and accurate in classifying explosions than earthquakes, reflecting a known shortcoming of the model, that being of a bias towards predicting explosions over earthquakes in the west coast due to training data biases.

97 MATHEMATICS AND COMPUTING↗

DOE OSTI2022

Transitioning from File-Based HPC Workflows to Streaming Data Pipelines with openPMD and ADIOS2

This paper aims to create a transition path from file-based IO to streaming-based workflows for scientific applications in an HPC environment. By using the openPMP-api, traditional workflows limited by filesystem bottlenecks can be overcome and flexibly extended for in situ analysis. The openPMD-api is a library for the description of scientific data according to the Open Standard for Particle-Mesh Data (openPMD). Its approach towards recent challenges posed by hardware heterogeneity lies in the decoupling of data description in domain sciences, such as plasma physics simulations, from concrete implementations in hardware and IO. The streaming backend is provided by the ADIOS2 framework, developed at Oak Ridge National Laboratory. This paper surveys two openPMD-based loosely-coupled setups to demonstrate flexible applicability and to evaluate performance. In loose coupling, as opposed to tight coupling, two (or more) applications are executed separately, e.g. in individual MPI contexts, yet cooperate by exchanging data. This way, a streaming-based workflow allows for standalone codes instead of tightly-coupled plugins, using a unified streaming-aware API and leveraging high-speed communication infrastructure available in modern compute clusters for massive data exchange. We determine new challenges in resource allocation and in the need of strategies for a flexible data distribution, demonstrating their influence on efficiency and scaling on the Summit compute system. The presented setups show the potential for a more flexible use of compute resources brought by streaming IO as well as the ability to increase throughput by avoiding filesystem bottlenecks.

Poeschel, Franz↗

DOE OSTI2022

CORAL: A framework for rigorous self-validated data modeling and integrative, reproducible data analysis

Abstract Background Many organizations face challenges in managing and analyzing data, especially when relevant datasets arise from multiple sources and methods. Analyzing heterogeneous datasets and additional derived data requires rigorous tracking of their interrelationships and provenance. This task has long been a Grand Challenge of data science and has more recently been formalized in the FAIR principles: that all data objects be Findable, Accessible, Interoperable, and Reusable, both for machines and for people. Adherence to these principles is necessary for proper stewardship of information, for testing regulatory compliance, for measuring the efficiency of processes, and for facilitating reuse of data-analytical frameworks. Findings We present the Contextual Ontology-based Repository Analysis Library (CORAL), a platform that greatly facilitates adherence to all 4 of the FAIR principles, including the especially difficult challenge of making heterogeneous datasets Interoperable and Reusable across all parts of a large, long-lasting organization. To achieve this, CORAL's data model requires that data generators extensively document the context for all data, and our tools maintain that context throughout the entire analysis pipeline. CORAL also features a web interface for data generators to upload and explore data, as well as a Jupyter notebook interface for data analysts, both backed by a common API. Conclusions CORAL enables organizations to build FAIR data types on the fly as they are needed, avoiding the expense of bespoke data modeling. CORAL provides a uniquely powerful platform to enable integrative cross-dataset analyses, generating deeper insights than are possible using traditional analysis tools.

97 MATHEMATICS AND COMPUTING↗

DOE OSTI2026

ML–Enabled FPGA Framework for Fast Quantum State Discrimination in Mid-Circuit Measurement Regimes

Accurate and low-latency quantum state discrimination is essential for protocols involving mid-circuit measurement (MCM) and conditional feed-forward. In superconducting quantum systems, conventional readout pipelines transfer measurement data to host processors for post-processing, introducing millisecond-scale delays that far exceed qubit coherence times. To overcome this bottleneck, we present an in-situ machine learning (ML) inference engine implemented on an FPGA for real-time quantum state discrimination. Our design performs inference directly on digitized readout signals with 40 ns latency, supports both qubit and qutrit readout, and enables conditional operations without host-side intervention. This capability is critical for MCM and for feedback-driven protocols such as quantum error correction. We validate the system on superconducting transmon hardware, demonstrating robust discrimination fidelity across multiple qubit and qutrit channels. We further demonstrate conditional qutrit logic driven by FPGA-resident classification, highlighting the potential of low-latency ML-on-FPGA control for NISQ applications and scalable fault-tolerant quantum computing.

Vora, Neel [Lawrence Berkeley National Laboratory ↗

DOE OSTI2026

Persistent DynAMICS

Persistent DynAMICS is a lightweight software platform that enables distributed sensing systems to operate as a unified, responsive network. It integrates with existing sensors, data systems, and processing tools as an overlay, without requiring hardware replacement or increasing user burden. As data is collected, the system identifies situations that require additional attention and directs selected sensors to gather more information. For example, if a sensor detects an unusual condition along a pipeline, the system can task nearby sensors or different sensing modalities to confirm the event and focus data collection where it is needed. To support this, the platform forms temporary groups of sensors that work together to address specific objectives. Once complete, these groups are dissolved and resources are reassigned as needed. Communication is handled through an integrated messaging framework with client and server components, enabling reliable operation and data continuity even in low-bandwidth or intermittent connectivity environments.

97 MATHEMATICS AND COMPUTING↗

DOE OSTI2026

HARMONY: Large-Scale Architecture Search for Efficient Hybrid Language Models

As large language models scale to trillions of parameters, their computational and memory requirements present critical challenges for efficient training and deployment. While Mixture of Experts (MoE) architectures enable efficient scaling through sparse parameter activation, and state-space models like Mamba offer linear-time complexity, principled methods for combining these paradigms remain undeveloped. We introduce HARMONY (Hybrid Architecture Research for Mamba, Optimized with Neural efficiencY), a multi-objective evolutionary neural architecture search framework for discovering efficient hybrid language models that integrate Transformer attention mechanisms, Mixture-of-Experts routing, and Mamba state-space components. Through large-scale distributed search using 16,384 MI250X GPUs on the Frontier supercomputer, HARMONY explores a comprehensive design space encompassing six attention variants (MHA, MQA, GQA, MLA, SWA, and Mamba-2), variable MoE configurations with both routed and shared experts, and extensive Mamba hyperparameters. Our framework discovers heterogeneous architectures that balance training performance with computational efficiency through multi-objective optimization incorporating latency penalties and fitness-based selection. Analysis of discovered architectures reveals that optimal hybrid designs favor heterogeneous component mixing rather than homogeneous patterns, with Mamba-2 and Multi-Head Latent Attention (MLA) emerging as preferred mechanisms. Discovered architectures demonstrate superior training efficiency: our best configuration achieves a final perplexity of 1.0874 with 2.38B parameters while processing 4,320 tokens/second, outperforming significantly larger manually designed models. Full-scale evaluation shows HARMONY's top architectures achieve better loss trajectories than equivalently-sized models using state-of-the-art configurations including Mixtral, Jamba, and Samba. Additionally, we demonstrate 91% weak scaling efficiency when training discovered 36B-parameter models across 1,024 GPUs. HARMONY is released as an open framework with comprehensive tools for building and training hybrid models using expert-data-pipeline parallelism, democratizing access to automated architecture design for next-generation language models.

Herron, Emily [ORNL] (ORCID:0000000273008172)↗

DOE OSTI2025

A Gigaparsec-scale Hydrodynamic Volume Reconstructed with Deep Learning

The next generation of spectroscopic surveys will map the large-scale structure of the Universe at high redshifts (2 ≤ z ≤ 5) using millions of quasar spectra, enabling major advances in constraining both the standard cosmological model and its extensions. Robust cosmological analyses of these data sets require numerical simulations that both cover gigaparsec volumes and resolve features on ∼10 kpc scales and smaller. However, running such large-volume, high-resolution hydrodynamic simulations is computationally prohibitive. We present a generative deep learning model that enhances a low-resolution, gigaparsec-scale (960 h −1 Mpc) hydrodynamic simulation using a smaller (80 h −1 Mpc) high-resolution input hydrodynamic simulation as training data. The resulting enhanced simulation reproduces the line-of-sight power spectrum to within ∼10% and the three-dimensional power spectrum at the ∼20% level at intermediate to small scales (k ≲ 2 h Mpc −1 ). Our method shows strong promise for producing realistic simulations for cosmological analyses with current surveys such as the Dark Energy Spectroscopic Instrument and upcoming next-generation experiments, but further improvements are needed to accurately recover the large-scale modes. We publicly release the enhanced hydrodynamic simulation, along with a halo catalog from a companion N-body dark matter simulation to support the calibration of data analysis pipelines for these large-scale surveys.

Convolutional neural networks↗