SEARCH · Engineering Papers

Results for “Training Analysis”

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.

SourceResearch connectionFrom yearThrough year

Quote a phrase for an exact phrase match. Source license links do not imply unrestricted reuse.

At least 343 records · Page 19

All sources NASA NTRS DOE OSTI

DOE OSTI2026

GADRAS + Geant4 Simulated Gamma Spectra Dataset

This dataset contains GADRAS and Geant4 simulated gamma-ray energy spectra used in the papers "Sim-to-real supervised domain adaptation for radioisotope identification" (https://doi.org/10.1016/j.nima.2025.171159) and "Unsupervised domain adaptation for radioisotope identification in gamma spectroscopy" (https://arxiv.org/abs/2603.05719). Each spectrum contains a mixture of several different radioisotopes, and the goal of the analysis is to leverage domain adaptation techniques to use a model trained on source-domain spectra (GADRAS) to classify target-domain spectra (Geant4).

Lalor, Peter William [Pacific Northwest National L↗

DOE OSTI2024

New Measurements of the Lyα Forest Continuum and Effective Optical Depth with LyCAN and DESI Y1 Data

Abstract We present the Ly α Continuum Analysis Network (LyCAN), a convolutional neural network that predicts the unabsorbed quasar continuum within the rest-frame wavelength range of 1040–1600 Å based on the red side of the Ly α emission line (1216–1600 Å). We developed synthetic spectra based on a Gaussian mixture model representation of nonnegative matrix factorization (NMF) coefficients. These coefficients were derived from high-resolution, low-redshift ( z < 0.2) Hubble Space Telescope/Cosmic Origins Spectrograph (COS) quasar spectra. We supplemented this COS-based synthetic sample with an equal number of DESI Year 5 mock spectra. LyCAN performs extremely well on testing sets, achieving a median error in the forest region of 1.5% on the DESI mock sample, 2.0% on the COS-based synthetic sample, and 4.1% on the original COS spectra. LyCAN outperforms principal component analysis (PCA) and NMF-based prediction methods using the same training set by 40% or more. We predict the intrinsic continua of 83,635 DESI Year 1 spectra in the redshift range of 2.1 ≤ z ≤ 4.2 and perform an absolute measurement of the evolution of the effective optical depth. This is the largest sample employed to measure the optical depth evolution to date. We fit a power law of the form τ ( z ) = τ 0 ( 1 + z ) γ to our measurements and find τ 0 = (2.46 ± 0.14) × 10 −3 and γ = 3.62 ± 0.04. Our results show particular agreement with high-resolution, ground-based observations around z = 2, indicating that LyCAN is able to predict the quasar continuum in the forest region with only spectral information outside the forest.

79 ASTRONOMY AND ASTROPHYSICS↗

DOE OSTI2021

ECAR-1944 AGC-1 AS RUN THERMAL RESULTS

Purpose of this ECAR is to provide documentation of as-run analysis results for each specimen temperature in the AGC-1 test train at specific time points during the irradiation.

22 GENERAL STUDIES OF NUCLEAR REACTORS↗

DOE OSTI2024

HAPPA: A Modular Platform for HPC Application Resilience Analysis with LLMs Embedded

High-performance computing (HPC) systems are increasingly vulnerable to soft errors, which pose significant challenges in maintaining computational accuracy and reliability. Predicting the resilience of HPC applications to these errors is crucial for robust code protection and detailed resilience analysis. In this study, we present HAppA, a modular platform designed for HPC Application Resilience Analysis. Embedding Large Language Models (LLMs), HAppA addresses understanding the context information of long code sequences typical in HPC applications. HAppA implements a novel code representation module that chunks the code into fixed-size segments and aggregates the embeddings of these segments. Three aggregation methods have been explored: MeanPooling, MaxPooling, and LSTM-based techniques. We built a DAtaset for REsilience analysis using Fault Injection (FI), named DARE. Using our DARE dataset, HAppA is trained for regression prediction tasks. Our evaluation results demonstrate the predictive accuracy of HAppA compared to other models, particularly noting that the LSTM-based aggregation method -- HAppA-LSTM -- achieves a mean squared error (MSE) of 0.078 for SDC prediction, surpassing the existing state-of-the-art PARIS model, which recorded an MSE of 0.1172. Additionally, HAppA with the KeyBERT model extracts a list of keywords representing the source code. A comprehensive importance analysis of these keywords further elucidates the code patterns contributing to the error rate. These findings highlight the effectiveness of HAppA in analyzing the resilience of HPC applications and establish a new benchmark for predictive accuracy in resilience.

Jiang, Hailong [Kent State University]↗

DOE OSTI2026

Image processing pipeline for AI-driven nanoparticle megalibrary characterization

Recent innovations have made it possible to produce megalibraries, millions of structurally and compositionally distinct nanoparticles on a chip. These megalibraries yield vast volumes of data that are impossible to analyze manually, necessitating the development of automated tools. In previous work, we created a binary classification machine learning model to select quality nanoparticle images for downstream analysis. In this work, we show that adding a custom image processing step before training can produce significantly higher-performing models in a fraction of the time and make them more robust to different image noise levels and microscope acquisition settings. The image processing pipeline proposed here effectively cleans raw nanoparticle images, enhances key features, and allows us to use much lower resolution images and simpler neural network model architectures. These features result in higher performance and significant cost savings. Experiments demonstrate superior performance relative to baseline, including an 18.2% improvement in recall and a 13.1% increase in accuracy. Given the high cost of downstream analysis, it is critical to minimize false positives, and our best-performing model reaches a precision of 95.9% and a weighted F-score of 95.1% on an unseen test set. Additionally, model training time is reduced from hours to less than a minute. We also show that, using this custom image processing pipeline, model performance is significantly improved at lower pixel resolutions compared to downsizing alone. We expect that adopting this pipeline for AI-driven automated nanoparticle characterization will allow researchers to rapidly and accurately analyze much greater volumes of data, thereby accelerating materials discovery.

77 NANOSCIENCE AND NANOTECHNOLOGY↗

DOE OSTI2025

Microscopic Imprints of Learned Solutions in Tunable Networks

In physical networks trained using supervised learning, physical parameters are adjusted to produce desired responses to inputs. An example is an electrical contrastive local learning network of nodes connected by edges that adjust their conductances during training. When an edge conductance changes, it upsets the current balance of every node. In response, physics adjusts the node voltages to minimize the dissipated power. Learning in these systems is therefore a coupled double-optimization process, in which the network descends both a cost landscape in the high-dimensional space of edge conductances and a physical landscape—the power dissipation—in the high-dimensional space of node voltages. Because of this coupling, the physical landscape of a trained network contains information about the learned task. Here, we derive a structure-function relation for trained tunable networks and demonstrate that all the physical information relevant to the trained input-output relation can be captured by a tuning susceptibility, an experimentally measurable quantity. We supplement our theoretical results with simulations to show that the tuning susceptibility is correlated with functional importance and that we can extract physical insight into how the system performs the task from the conductances of highly susceptible edges. Our analysis is general and can be applied directly to mechanical networks, such as networks trained for protein-inspired function such as allostery.

36 MATERIALS SCIENCE↗

DOE OSTI2026

Opportunities to Expand Building Efficiency Programming at Community Colleges

According to the most recent U.S. Energy and Employment Report, more than 2.3 million workers in the United States are involved in activities that reduce energy usage in buildings. This workforce supports energy efficiency from the design of buildings and their systems through the manufacturing and trade of components and supplies involved in these systems to the installation, repair, and maintenance of these systems. Less than 10% of the workers in key building efficiency occupations have a bachelor’s or higher degree, compared to ~40% of the general workforce. Thus, the community college system is a key stakeholder in training and educating a large portion of the building efficiency workforce. Despite this, the literature review conducted for this report found almost no research focused on better understanding and supporting the role of community colleges as they train this workforce at scale. This report seeks to understand how and to what extent building efficiency and advanced building technology concepts are being addressed in community colleges as well as potential pathways for schools to consider to better prepare students to enter the building efficiency industry. The first section presents information from a literature review and data analysis to provide background on the building efficiency workforce, the types of building efficiency training and education available from community colleges, and the barriers and challenges that exist in the workforce. The second section offers a series of case studies that illustrate the various ways that building efficiency content can be addressed at community colleges. The final section provides an overview of the opportunities available to community colleges as well as considerations for schools that want to increase building efficiency programming.

32 ENERGY CONSERVATION, CONSUMPTION, AND UTILIZATI↗

DOE OSTI2025

In-Transit Data Transport Strategies for Coupled AI-Simulation Workflow Patterns

Coupled AI-Simulation workflows are becoming the major workloads for HPC facilities, and their increasing complexity necessitates new tools for performance analysis and prototyping of new in-situ workflows. We present SimAI-Bench, a tool designed to both prototype and evaluate these coupled workflows. In this paper, we use SimAI-Bench to benchmark the data transport performance of two common patterns on the Aurora supercomputer: a one-to-one workflow with co-located simulation and AI training instances, and a many-to-one workflow where a single AI model is trained from an ensemble of simulations. For the one-to-one pattern, our analysis shows that node-local and DragonHPC data staging strategies provide excellent performance compared Redis and Lustre file system. For the many-to-one pattern, we find that data transport becomes a dominant bottleneck as the ensemble size grows. Our evaluation reveals that file system is the optimal solution among the tested strategies for the many-to-one pattern.

Tummalapalli, Harikrishna [Argonne National Labora↗

DOE OSTI2025

Wide‐Field Bond Quality Evaluation Using Frequency Domain Thermoreflectance with Deep Neural Network Feature Reconstruction

Heterogeneous integration of microelectronic components provides a pathway to improve circuit/component performance; however, this comes with assembly challenges, in particular due to complex interfaces via subsurface bump bonds. The ability of these bonds to transmit electrical signals and conduct heat to the carrier substrate limits component performance. In this work, hyperspectral frequency‐domain thermoreflectance (FDTR) imaging is demonstrated as a robust technique for evaluating the quality of subsurface indium bump bonds in a surrogate microelectronic sample. By performing microscale FDTR imaging with coarse motion image stitching, thermal phase maps that cover a 4 mm by 4 mm field‐of‐view with subsurface feature sensitivity at depths greater than 50 µm are obtained. The resulting FDTR hyperspectral data contains more than three million pixels and reveal the quality of subsurface microbump arrays. Wide‐field analysis of bonded versus gap regions is enabled by deep neural network feature reconstruction, that after training, rapidly provides an interpretable representation of bond quality. Utility of noisy higher frequency FDTR phase maps, i.e., near the computationally predicted sensing depth limit, results in an average prediction error of 11%. Taken together, FDTR with neural network‐based analysis demonstrates subsurface bond monitoring at length scales relevant for heterogeneously integrated microelectronics.

FDTR↗

DOE OSTI2025

Quantitative Imaging of Cobalt Phthalocyanine Distribution on Carbon Nanotubes: A Deep Learning Approach to Catalyst Characterization

Electrochemical reduction of carbon dioxide (CO 2 ) offers a pathway to valuable products, with catalysts playing a crucial role. This study investigates the distribution of cobalt tetraaminophthalocyanine (CoPc-NH 2 ) immobilized on carbon nanotubes (CNTs), utilizing high-angle annular dark-field scanning transmission electron microscopy (HAADF-STEM) to characterize CoPc-NH 2 distribution. A challenge in the quantitative HAADF-STEM analysis is the introduction of bias from manual Co atom identification. To address this, we developed and trained a convolutional neural network (CNN) using a data set generated from images of CoPc-NH 2 /CNT samples with varying Co loadings. The CNN, implemented in TensorFlow and Keras, facilitated Co atom detections. Analysis of the CNN-generated data confirmed a correlation between Co loading and surface density, consistent with findings from UV–vis spectroscopy. Furthermore, the application of Ripley’s L(d) function highlighted the presence of slight Co atom clustering. Furthermore, this work demonstrates the utility of the combined HAADF-STEM and CNN approach for providing spatially resolved information about catalyst distribution on nonplanar supports, revealing structural details that are typically lost through other characterization methods.

HAADF-STEM↗

DOE OSTI2022

Deep learning to estimate permeability using geophysical data

Time-lapse electrical resistivity tomography (ERT) is a popular geophysical method to estimate three-dimensional (3D) permeability fields from electrical potential difference measurements. Traditional inversion and data assimilation methods are used to ingest this ERT data into hydrogeophysical models to estimate permeability. Due to ill-posedness and the curse of dimensionality, existing inversion strategies provide poor estimates and low resolution of the 3D permeability field. Recent advances in deep learning provide us with powerful algorithms to overcome this challenge. This paper presents a deep learning (DL) framework to estimate the 3D subsurface permeability from time-lapse ERT data. To test the feasibility of the proposed framework, we train DL-enabled inverse models on simulation data. Each measurement in both synthetic and field data is standardized by removing the mean and scaling the time-series to unit variance. This pre-processing step is necessary to bring simulation data closer to field observations. Subsurface process models based on hydrogeophysics are used to generate this synthetic data. Training performed on limited simulation data resulted in the DL model over-fitting. An advanced data augmentation based on mixup is implemented to generate additional training samples to overcome this issue. This mixup technique creates weakly labeled (low-fidelity) samples from strongly labeled (high-fidelity) data. The weakly labeled training data is then used to develop DL-enabled inverse models and reduce over-fitting. As both time-lapse ERT (1133048 features/realization) and 3D permeability (585453 features/realization) data samples are from a high-dimensional space, principal component analysis (PCA) is employed to reduce dimensionality. Encoded ERT and encoded permeability are generated using the trained PCA estimators. A deep neural network is then trained to map the encoded ERT to encoded permeability. This mixup training and unsupervised learning allowed us to build a fast and reasonably accurate DL-based inverse model under limited simulation data. Results show that proposed weak supervised learning can capture salient spatial features in the 3D permeability field. Quantitatively, the average mean squared error (in terms of the natural log) on the strongly labeled training, validation, and test datasets is less than 0.5. The R 2 -score (global metric) is greater than 0.75, and the percent error in each cell (local metric) is less than 10%. Finally, an added benefit in terms of computational cost is that the proposed DL-based inverse model is at least O(10 4 ) times faster than running a forward model once it is trained. Data generation, DL model training, and hyperparameter tuning to identify optimal neural network architectures utilized high-performance computing resources while the DL inference is performed on a standard laptop. Approximately, O(10 5 ) processor hours are used for generating data and DL tuning and training. We acknowledge that the data generation and DL model development are expensive. But once a DL model is trained, it can be re-used for inversion rapidly for the given system, with set physics and domain. Note that traditional inversion may require multiple forward model simulations (e.g., in the order of 10 to 1000), which are very expensive. This computational savings ≈ O(10 5 ) – O(10 7 )) makes the proposed DL-based inverse model attractive for subsurface imaging and real-time ERT monitoring applications due to fast and yet reasonably accurate estimations of permeability field.

58 GEOSCIENCES↗

DOE OSTI2024

Collaborative Research: Vlasov-Maxwell Simulations to Resolve Electron Heating and Dissipation, in Quasi-Perpendicular Shocks (Final Report)

Collisionless shocks are a grand challenge problem in plasma physics and have been the subject of study for more than six decades. Shocks are ubiquitous phenomena in the universe and are responsible for transforming high energy and momentum flows into thermal energy and energetic particles. Understanding how shocks operate is of primary importance to understand the sun-earth coupling, protecting manned missions and spacecrafts from high energy particles, inertial confinement fusion, and radiation observed from astrophysical plasmas, such as supernova remnants and astrophysical jets. A major unanswered question on this frontier is, how does a collisionless plasma transform flow energy and momentum into electron thermal energy? Many potential mechanisms have been proposed to perform the conversion between flow and thermal energy in a collisionless plasma, but the answer has been elusive. To solve the mystery, non-relativistic perpendicular and quasi-perpendicular shocks are studied via a coordinated approach employing cutting-edge numerical simulations with the continuum kinetic code Gkeyll and spacecraft observations supplied by NASA's Magnetospheric Multiscale (MMS) mission. In both sets of data, we employ novel techniques to identify how and where energy is exchanged between electromagnetic fields and plasma particles. By combining cutting-edge spacecraft data and numerical simulations with these diagnostics, we have been able to identify some of the primary mechanisms responsible electron heating and dissipation in heliospheric shocks. In addition to electron dynamics, we have complete access to proton simulation and in situ data, enabling us to characterize proton heating and acceleration in quasi-perpendicular shocks. Many astrophysical quasi-perpendicular shocks have similar parameters to those in this study, including supernova remnants and shocks in galaxy clusters. Additionally, the simulation code, Gkeyll, and associated analysis scripts are openly available for community use. This award has supported a collaborative effort between Princeton University, the University of Arizona - Tucson, and the University of Maryland - College Park. The award funded three early career scientists, including the funding and training of one female postdoctoral researcher, who was involved in all aspects of the MMS data analysis and some portions of the simulation analysis.

70 PLASMA PHYSICS AND FUSION TECHNOLOGY↗

DOE OSTI2026

Uncertainty-Aware Machine Learning for Small-Angle X-ray Scattering Analysis in Autonomous Experimentation

Small-angle X-ray scattering (SAXS) is a powerful high-throughput characterization tool for probing nanoscale structure in native sample environments, providing real-time morphological information such as nanoparticle size and shape during synthesis. However, automated SAXS data analysis for extracting meaningful structural parameters is non-trivial and remains a bottleneck in closed-loop experimentation towards autonomous materials discovery, which demands fast, reliable, and uncertainty-aware data analysis. Here, we develop a machine-learning approach for automated SAXS analysis tailored to closed-loop nanoparticle synthesis. A Random Forest (RF) regression model is trained on 100,000 synthetic SAXS curves generated from polydisperse spherical nanoparticles with realistic background contributions. Using normalized one-dimensional SAXS intensity profiles as input, the RF model directly predicts nanoparticle radius, size polydispersity, and background parameters, while the ensemble standard deviation across trees provides built-in uncertainty quantification (UQ). On synthetic data, we show that combining fit-quality metrics (R 2 , MAE) with thresholds on prediction uncertainty reliably identifies accurate parameter estimates without access to ground truth. We then apply the trained model to 365 experimental SAXS profiles of citrate-reduced gold nanoparticles synthesized using an automated droplet-flow microreactor with in situ SAXS at a synchrotron beamline, classifying the results into high- and low-confidence subsets based on UQ metrics. Finally, we integrate RF-based SAXS analysis into a simulated closed-loop optimization campaign using Gaussian process Bayesian optimization to minimize nanoparticle polydispersity, benchmarking against conventional automated Levenberg–Marquardt fitting. The RF-guided campaign exhibits substantially faster convergence and lower relative opportunity cost (∼0.07 vs ∼0.3), demonstrating that uncertainty-aware machine-learning SAXS analysis significantly enhances the efficiency and robustness of autonomous nanomaterials synthesis workflows.

Bayesian optimization↗

DOE OSTI2025

LTAU-FF: Loss Trajectory Analysis for Uncertainty in atomistic Force Fields

Model ensembles are effective tools for estimating prediction uncertainty in deep learning atomistic force fields. However, their widespread adoption is hindered by high computational costs and overconfident error estimates. In this work, we address these challenges by leveraging distributions of per-sample errors obtained during training and employing a distance-based similarity search in the model latent space. Our method, which we call LTAU (Loss Trajectory Analysis for Uncertainty), efficiently estimates the full probability distribution function of errors for any test point using the logged training errors, achieving speeds that are 2–3 orders of magnitudes faster than typical ensemble methods and allowing it to be used for tasks where training or evaluating multiple models would be infeasible. We apply LTAU towards estimating parametric uncertainty in atomistic force fields (LTAU-FF), demonstrating that it produces well-calibrated confidence intervals and predicts errors that correlate strongly with the true errors for data near the training domain. Furthermore, we show that the errors predicted by LTAU-FF can be used in practical applications for detecting out-of-domain data, tuning model performance, and predicting failure during simulations. We believe that LTAU will be a valuable tool for uncertainty quantification in atomistic force fields and is a promising method that should be further explored in other domains of machine learning.

97 MATHEMATICS AND COMPUTING↗

DOE OSTI2022

Integrating Deep Learning and Hydrodynamic Modeling to Improve the Great Lakes Forecast

The Laurentian Great Lakes, one of the world’s largest surface freshwater systems, pose a modeling challenge in seasonal forecast and climate projection. While physics-based hydrodynamic modeling is a fundamental approach, improving the forecast accuracy remains critical. In recent years, machine learning (ML) has quickly emerged in geoscience applications, but its application to the Great Lakes hydrodynamic prediction is still in its early stages. This work is the first one to explore a deep learning approach to predicting spatiotemporal distributions of the lake surface temperature (LST) in the Great Lakes. Our study shows that the Long Short-Term Memory (LSTM) neural network, trained with the limited data from hypothetical monitoring networks, can provide consistent and robust performance. The LSTM prediction captured the LST spatiotemporal variabilities across the five Great Lakes well, suggesting an effective and efficient way for monitoring network design in assisting the ML-based forecast. Furthermore, we employed an explainable artificial intelligence (XAI) technique named SHapley Additive exPlanations (SHAP) to uncover how the features impact the LSTM prediction. Our XAI analysis shows air temperature is the most influential feature for predicting LST in the trained LSTM. The relatively large bias in the LSTM prediction during the spring and fall was associated with substantial heterogeneity of air temperature during the two seasons. In contrast, the physics-based hydrodynamic model performed better in spring and fall yet exhibited relatively large biases during the summer stratification period. Finally, we developed a statistical integration of the hydrodynamic modeling and deep learning results based on the Best Linear Unbiased Estimator (BLUE). The integration further enhanced prediction accuracy, suggesting its potential for next-generation Great Lakes forecast systems.

Xue, Pengfei (ORCID:000000025702421X)↗

DOE OSTI2022

Finding Hidden Patterns in High Resolution Wind Flow Model Simulations

Wind flow data is critical in terms of investment decisions and policy making. High resolution data from wind flow model simulations serve as a supplement to the limited resource of original wind flow data collection. Given the large size of data, finding hidden patterns in wind flow model simulations are critical for reducing the dimensionality of the analysis. In this work, we first perform dimension reduction with two autoencoder models: the CNN-based autoencoder (CNN-AE) [1], and hierarchical autoencoder (HIER-AE) [2], and compare their performance with the Principal Component Analysis (PCA). We then investigate the super-resolution of the wind flow data. By training a Generative Adversarial Network (GAN) with 300 epochs, we obtained a trained model with 2× resolution enhancement. We compare the results of GAN with Convolutional Neural Network (CNN), and GAN results show finer structure as expected in the data field images. Also, the kinetic energy spectra comparisons show that GAN outperforms CNN in terms of reproducing the physical properties for high wavenumbers and is critical for analysis where high-wavenumber kinetics play an important role.

97 MATHEMATICS AND COMPUTING↗

DOE OSTI2023

Lateral Variations Across the Southern San Andreas Fault Zone Revealed From Analysis of Traffic Signals at a Dense Seismic Array

Abstract We image the shallow seismic structure across the Southern San Andreas Fault (SSAF) using signals from freight trains and trucks recorded by a dense nodal array, with a linear component perpendicular to SSAF and two 2D subarrays centered on the Banning Fault and Mission Creek Fault (MCF). Particle motion analysis in the frequency band 2–5 Hz shows that the examined traffic sources can be approximated as moving single‐ or multi‐point sources that primarily induce Rayleigh waves. Using several techniques, we resolve strong lateral variations of Rayleigh wave velocities and Q ‐values across the SSAF, including 35% velocity reduction across MCF toward the northeast and strong attenuation around the two fault strands. We further resolve 10% mass density reduction and 45% shear modulus decrease across the MCF. These findings suggest that the MCF is currently the main strand of the SSAF in the area with important implications for seismic hazard assessments.

58 GEOSCIENCES↗

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)↗