Search NASA⌕ Search

SEARCH · Search NASA

Results for “evolutionary”

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 271 records · Page 15
All sourcesNASA NTRSDOE OSTI
DOE OSTI2024

Structural Insights into Mechanisms Underlying Mitochondrial and Bacterial Cytochrome c Synthases

Mitochondrial holocytochrome c synthase (HCCS) is an essential protein in assembling cytochrome c (cyt c) of the electron transport system. HCCS binds heme and covalently attaches the two vinyls of heme to two cysteine thiols of the cyt c CXXCH motif. Human HCCS recognizes both cyt c and cytochrome c1 of complex III (cytochrome bc1). HCCS is mutated in some human diseases and it has been investigated recombinantly by mutational, biochemical, and reconstitution studies in the past decade. Here, we employ structural prediction programs (e.g., AlphaFold 3) on HCCS and its two substrates, heme and cytochrome c. The results, when combined with spectroscopic and functional analyses of HCCS and variants, provide insights into the structural basis for heme binding, apocyt c binding, covalent attachment, and release of the holocyt c product. Results from in vitro reconstitution of purified human HCCS using cyt c and cyt c1 peptides as acceptors are consistent with the structural modeling of substrate binding. Reconstitution of HCCS and cyt c1 provides an approach to studying cyt c1 assembly, which has been refractile to recombinant in vivo reconstitution (unlike HCCS and cyt c). We propose a structural basis for release of the holocyt c product from HCCS based on in vitro studies and on cryoEM structures of the bacterial cyt c synthase (CcsBA) active site. We analyze the kinetoplastid mitochondrial synthase (KCCS), and hypothesize a molecular evolutionary path from mitochondrial endosymbiosis to the current HCCS.

Biochemistry & Molecular Biology↗
DOE OSTI2024

StructuredFuzzer: Fuzzing Structured Text-Based Control Logic Applications

Rigorous testing methods are essential for ensuring the security and reliability of industrial controller software. Fuzzing, a technique that automatically discovers software bugs, has also proven effective in finding software vulnerabilities. Unsurprisingly, fuzzing has been applied to a wide range of platforms, including programmable logic controllers (PLCs). However, current approaches, such as coverage-guided evolutionary fuzzing implemented in the popular fuzzer American Fuzzy Lop Plus Plus (AFL++), are often inadequate for finding logical errors and bugs in PLC control logic applications. They primarily target generic programming languages like C/C++, Java, and Python, and do not consider the unique characteristics and behaviors of PLCs, which are often programmed using specialized programming languages like Structured Text (ST). Furthermore, these fuzzers are ill suited to deal with complex input structures encapsulated in ST, as they are not specifically designed to generate appropriate input sequences. This renders the application of traditional fuzzing techniques less efficient on these platforms. To address this issue, this paper presents a fuzzing framework designed explicitly for PLC software to discover logic bugs in applications written in ST specified by the IEC 61131-3 standard. The proposed framework incorporates a custom-tailored PLC runtime and a fuzzer designed for the purpose. We demonstrate its effectiveness by fuzzing a collection of ST programs that were crafted for evaluation purposes. We compare the performance against a popular fuzzer, namely, AFL++. The proposed fuzzing framework demonstrated its capabilities in our experiments, successfully detecting logic bugs in the tested PLC control logic applications written in ST. On average, it was at least 83 times faster than AFL++, and in certain cases, for example, it was more than 23,000 times faster.

47 OTHER INSTRUMENTATION↗
DOE OSTI2024

Light-Induced Charge Separation in Photosystem I from Different Biological Species Characterized by Multifrequency Electron Paramagnetic Resonance Spectroscopy

Photosystem I (PSI) serves as a model system for studying fundamental processes such as electron transfer (ET) and energy conversion, which are not only central to photosynthesis but also have broader implications for bioenergy production and biomimetic device design. In this study, we employed electron paramagnetic resonance (EPR) spectroscopy to investigate key light-induced charge separation steps in PSI isolated from several green algal and cyanobacterial species. Following photoexcitation, rapid sequential ET occurs through either of two quasi-symmetric branches of donor/acceptor cofactors embedded within the protein core, termed the A and B branches. Using high-frequency (130 GHz) time-resolved EPR (TR-EPR) and deuteration techniques to enhance spectral resolution, we observed that at low temperatures prokaryotic PSI exhibits reversible ET in the A branch and irreversible ET in the B branch, while PSI from eukaryotic counterparts displays either reversible ET in both branches or exclusively in the B branch. Furthermore, we observed a notable correlation between low-temperature charge separation to the terminal [4Fe-4S] clusters of PSI, termed F A and F B , as reflected in the measured F A /F B ratio. These findings enhance our understanding of the mechanistic diversity of PSI’s ET across different species and underscore the importance of experimental design in resolving these differences. Though further research is necessary to elucidate the underlying mechanisms and the evolutionary significance of these variations in PSI charge separation, this study sets the stage for future investigations into the complex interplay between protein structure, ET pathways, and the environmental adaptations of photosynthetic organisms.

37 INORGANIC, ORGANIC, PHYSICAL, AND ANALYTICAL CH↗
DOE OSTI2026

Recent progress in atomic-scale controlled plasma processing

Atomic-scale control in plasma processing is becoming increasingly critical for fabricating of advanced semiconductor devices, particularly as the industry shifts toward three-dimensional (3D) architectures and high-aspect-ratio (HAR) structures. This review presents a comprehensive overview of recent developments in atomic-scale controlled plasma processes, organized along two key directions: the hierarchical structure of plasma–surface interactions and the generational evolution of atomic layer processing (ALP) technologies. We examined the gas phase, where molecular design enables selective generation of ions and radicals; the boundary layer, where transport phenomena govern species delivery into nanoscale features, and the surface, where temperature-dependent reactions and cyclic processing determine etching selectivity and precision. Building on this foundation, we outline five generations of ALP—from thermal atomic layer deposition to transport-aware, temporally and structurally decoupled processes—highlighting the increasing sophistication of process control. The review further explores the transition from empirical recipe development to science-based, data-driven methodologies. By integrating quantum-chemical modeling, advanced diagnostics, and machine learning, we demonstrated how predictive models can link plasma species composition to process outcomes, enabling autonomous and adaptive control strategies. Finally, this review discusses the broader societal implications of plasma process innovation through the E4 quartet: energy and resource efficiency, environmental sustainability, evolutionary advancement, and educational promotion. These principles guide the development of sustainable and intelligent atomic-scale manufacturing technologies that are not only technically advanced but also socially responsible.

Ishikawa, Kenji [Nagoya Univ. (Japan)] (ORCID:0000↗
DOE OSTI2024

A Full Accounting of the Visible Mass in SDSS MaNGA Disk Galaxies

We present a study of the ratio of visible mass to total mass in spiral galaxies to better understand the relative amount of dark matter present in galaxies of different masses and evolutionary stages. Using the velocities of the Hα emission line measured in spectroscopic observations from the Sloan Digital Sky Survey (SDSS) MaNGA Data Release 17 (DR 17), we evaluate the rotational velocity of over 5500 disk galaxies at their 90% elliptical Petrosian radii, R 90 . We compare this to the velocity expected from the total visible mass, which we compute from the stellar, H i, H 2 , and heavy metals and dust masses. H 2 mass measurements are available for only a small subset of galaxies observed in SDSS MaNGA DR17, so we derive a parameterization of the H 2 mass as a function of absolute magnitude in the r band using galaxies observed as part of SDSS DR7. With these parameterizations, we calculate the fraction of visible mass within R 90 that corresponds to the observed velocity. Based on statistically analyzing the likelihood of this fraction, we conclude that the null hypothesis (no dark matter) cannot be excluded at a confidence level better than 95% within the visible extent of the disk galaxies. We also find that when all mass components are included, the ratio of visible to total mass within the visible extent of star-forming disk galaxies increases with galaxy luminosity.

79 ASTRONOMY AND ASTROPHYSICS↗
DOE OSTI2024

Distance Estimate Method for Asymptotic Giant Branch Stars Using Infrared Spectral Energy Distributions

We present a method to estimate distances to asymptotic giant branch (AGB) stars in the Galaxy, using spectral energy distributions (SEDs) in the near- and mid-infrared. By assuming that a given set of source properties (initial mass, stellar temperature, composition, and evolutionary stage) will provide a typical SED shape and brightness, sources are color matched to a distance-calibrated template and thereafter scaled to extract the distance. The method is tested by comparing the distances obtained to those estimated from very long baseline interferometry or Gaia parallax measurements, yielding a strong correlation in both cases. Additional templates are formed by constructing a source sample likely to be close to the Galactic center, and thus with a common, typical distance for calibration of the templates. These first results provide statistical distance estimates to a set of almost 15,000 Milky Way AGB stars belonging to the Bulge Asymmetries and Dynamical Evolution (BAaDE) survey, with typical distance errors of ±35%. With these statistical distances, a map of the intermediate-age population of stars traced by AGBs is formed, and a clear bar structure can be discerned, consistent with the previously reported inclination angle of 30° to the GC–Sun direction vector. These results motivate deeper studies of the AGB population to tease out the intermediate-age stellar distribution throughout the Galaxy, as well as determining statistical properties of the AGB population luminosity and mass-loss-rate distributions.

79 ASTRONOMY AND ASTROPHYSICS↗
DOE OSTI2025

A Semi-analytical Model for Stellar Evolution in AGN Disks

Disks of gas accreting onto supermassive black holes may host numerous stellar-mass objects, formed within the disk or captured from a nuclear star cluster. We present a simplified model of stellar evolution in these dense environments, which exhibits exceptional agreement with full stellar evolution calculations at a minuscule fraction of the cost. Although the model presented here is limited to stars burning hydrogen in their cores, it is sufficient to determine the evolutionary fate of disk-embedded stars: whether they proceed to later stages of nuclear burning and leave behind a compact remnant, reach a quasi-steady state where mass loss and accretion balance one another, or whether accretion proceeds faster than stellar structure can adjust, causing a runaway. We highlight how various disk parameters and phenomena such as gap opening affect stellar evolution outcomes. We also highlight how our model can accommodate time-varying conditions, such as those experienced by a star on an eccentric orbit, and can couple to N -body integrations. This model will enable more detailed studies of stellar populations and their interaction with accretion disks than have previously been possible.

79 ASTRONOMY AND ASTROPHYSICS↗
DOE OSTI2025

Aggressively Dissipative Dark Dwarfs: The Effects of Atomic Dark Matter on the Inner Densities of Isolated Dwarf Galaxies

We present the first suite of cosmological hydrodynamical zoom-in simulations of isolated dwarf galaxies for a dark sector that consists of cold dark matter and a strongly dissipative subcomponent. The simulations are implemented in GIZMO and include standard baryons following the FIRE-2 galaxy formation physics model. The dissipative dark matter is modeled as atomic dark matter (aDM), which forms a dark hydrogen gas that cools in direct analogy to the Standard Model. Our suite includes seven different simulations of ∼10 10 M ⊙ systems that vary over the aDM microphysics and the dwarf’s evolutionary history. We identify a region of aDM parameter space where the cooling rate is aggressive and the resulting halo density profile is universal. In this regime, the aDM gas cools rapidly at high redshifts, and only a small fraction survives in the form of a central dark gas disk; the majority collapses centrally into collisionless dark “clumps,” which are clusters of subresolution dark compact objects. These dark clumps rapidly equilibrate in the inner galaxy, resulting in an approximately isothermal distribution that can be modeled with a simple fitting function. Even when only a small fraction (∼5%) of the total dark matter is strongly dissipative, the central densities of classical dwarf galaxies can be enhanced by over an order of magnitude, providing a sharp prediction for observations.

cold dark matter↗
DOE OSTI2025

Primordial Black Hole Triggered Type Ia Supernovae. I. Impact on Explosion Dynamics and Light Curves

Primordial black holes (PBHs) in the asteroid-mass window are compelling dark matter candidates, made plausible by the existence of black holes and by the variety of mechanisms of their production in the early Universe. If a PBH falls into a white dwarf (WD), the strong tidal forces can generate enough heat to trigger a thermonuclear runaway explosion, depending on the WD’s mass and the PBH’s orbital parameters. In this work, we investigate the WD explosion triggered by the passage of a PBH. We perform 2D simulations of the WD undergoing thermonuclear explosion in this scenario, with the predicted ignition site as a parameter assuming the deflagration–detonation transition model. We study the explosion dynamics, and predict the associated light curves and nucleosynthesis. We find that the model sequence predicts light curves which align with the Phillips relation (B max versus ΔM 15 ). Our models hint at a unifying approach in triggering Type Ia supernovae without involving two distinctive evolutionary tracks.

Dark matter↗
DOE OSTI2026

Hubble Space Telescope Imaging of Three Isolated Faint Dwarf Galaxies beyond the Local Group: Pavo, Corvus A, and Kamino

We present new Hubble Space Telescope (HST) imaging of three recently discovered star-forming dwarf galaxies beyond the Local Group: Pavo, Corvus A, and Kamino. The discovery of Kamino is reported here for the first time. They rank among the most isolated faint dwarf galaxies known; hence they provide unique opportunities to study galaxy evolution at the smallest scales, free from the environmental effects of more massive galaxies. Our HST data reach ∼2–4 magnitudes below the tip of the red giant branch (TRGB) for each dwarf, allowing us to measure their distances, structural properties, and recent star formation histories (SFHs). All three galaxies contain a complex stellar population of young and old stars, and are typical of field galaxies in this mass regime (M V = −10.62 ± 0.08 and $D = 2.16_{-0.07}^{+0.08}$ Mpc for Pavo, M V = −10.91 ± 0.10 and D = 3.34 ± 0.11 Mpc for Corvus A, and M V = −12.02 ± 0.12 and $D = 6.50_{-0.11}^{+0.15}$ Mpc for Kamino). Our HST-derived SFHs reveal differences among the three dwarfs: Pavo and Kamino show relatively steady, continuous star formation, while Corvus A formed ∼60% of its stellar mass by 10 Gyr ago. These results align with theoretical predictions of diverse evolutionary pathways for isolated low-mass galaxies.

Mutlu-Pakdil, Burçin [Dartmouth College, Hanover, ↗
DOE OSTI2026

Constraints on White Dwarf Hydrogen Layer Masses Using Gravitational Redshifts

The hydrogen envelope is the outermost layer of a DA white dwarf; it makes up the entirety of the stellar photosphere, and yet its typical extent is difficult to model theoretically and remains poorly observationally constrained. As a result, hydrogen envelope mass is a substantial source of systematic uncertainty in the physical properties of white dwarfs, including overall masses and cooling ages. In this work, we fit a Gaussian mixture model to gravitational redshifts from high-resolution spectroscopy, paired with radius measurements from Gaia BP/RP spectra, to measure the mass–radius relation for a sample of 468 white dwarfs. Our results are in excellent agreement with the predicted mass–radius relations of state-of-the-art evolutionary models, including those from the MESA Isochrones and Stellar Tracks (MIST) library. We find that mass–radius relations such as those from MIST that assume a thick and mass-dependent hydrogen envelope are preferred by the observed probability density function over models that assume a hydrogen envelope of constant mass. Proper treatment of the evolution of white dwarf progenitors is thus important for accurately modeling the mass–radius relation. Our results indicate that gravitational redshift measurements of large samples of white dwarfs in wide binaries are promising probes of the hydrogen envelope masses of DA white dwarfs.

Astronomy and AstroPhysics↗
DOE OSTI2026

[C/N] Ages for Red Giants and Their Implications for Galactic Archaeology

Red giants undergo the first dredge-up, a mixing event that creates a connection between their surface [C/N] and their mass and age. We derive a [C/N]–age relationship for red giants calibrated on Apache Point Observatory Galactic Evolution Experiment (APOGEE) DR17 abundances and APOKASC-3 asteroseismic ages. We find that we can use [C/N] to reliably recover asteroseismic ages between 1 and 10 Gyr with average uncertainties of 1.64 Gyr. We find that [C/N] yields concordant ages, with modest offsets, for stars in different evolutionary states. We also find that the [C/N]–birth mass relationship is robust for luminous giants, and argue that this is an advantage over direct asteroseismology for these stars. We use our ages to infer Galactic birth abundance trends in [Fe/H] and [Mg/H] as a function of position in the Galactic disk. We filter out stars with kinematic or chemical properties consistent with migrators and find the number of migrators to be much lower than expected by standard radial migration prescriptions. The remaining population shows weak chemical evolution trends, on the order of 0.01 dex Gyr −1 , over the last 10 Gyr across a wide range of radii.

Roberts, John D. [The Ohio State Univ., Columbus, ↗
DOE OSTI2025

Discovery of Powerful Multivelocity Ultrafast Outflows in the Starburst Merger Galaxy IRAS 05189–2524 with XRISM

We observed the X-ray-bright ultraluminous infrared galaxy IRAS 05189−2524 with XRISM during its performance verification phase. The unprecedented energy resolution of the onboard X-ray microcalorimeter revealed complex spectral features at ∼7–9 keV, which can be interpreted as blueshifted Fe XXV/XXVI absorption lines with various velocity dispersions, originating from ultrafast outflow (UFO) components with multiple bulk velocities of ∼0.076c, ∼0.101c, and ∼0.143c. In addition, a broad Fe–K emission line was detected around ∼7 keV, forming a P Cygni profile together with the absorption lines. The onboard X-ray CCD camera revealed a 0.4–12 keV broadband spectrum characterized by a neutrally absorbed power-law continuum with a photon index of ∼2.3 and intrinsic flare-like variability on timescales of ∼10 ks, both of which are likely associated with near-Eddington accretion. We also found potential variability of the UFO parameters on a timescale of ∼140 ks. Using these properties, we propose new constraints on the outflow structure and suggest the presence of multiple outflowing regions on scales of about tens to 100 Schwarzschild radii, located within roughly 2000 Schwarzschild radii. Since both the estimated momentum and energy outflow rates of the UFOs exceed those of galactic molecular outflows, our results indicate that powerful, multivelocity UFOs are already well developed during a short-lived evolutionary phase following a major galaxy merger, characterized by intense starburst activity and likely preceding the quasar phase. This system is expected to evolve into a quasar, sustaining strong UFO activity and suppressing star formation in the host galaxy.

Astronomy and AstroPhysics↗
DOE OSTI2025

HD 143811 AB b: A Directly Imaged Planet Orbiting a Spectroscopic Binary in Sco-Cen

We present confirmation of HD 143811 AB b, a substellar companion to spectroscopic binary HD 143811 AB through direct imaging with the Gemini Planet Imager (GPI) and Keck NIRC2. HD 143811 AB was observed as a part of the GPI Exoplanet Survey in 2016 and 2019 and is a member of the Sco-Cen star formation region. The exoplanet is detected ∼430 mas from the host star by GPI. With two GPI epochs and one from Keck/NIRC2 in 2022, we confirm through common proper motion analysis that the object is bound to its host star. We derive an orbit with a semimajor axis of $64^{+32}_{-14}$ au and eccentricity $0.23^{+0.24}_{-0.16}$. Spectral analysis of the GPI H-band spectrum and NIRC2 L′ photometry provides additional proof that this object is a substellar companion. We compare the spectrum of HD 143811 AB b to PHOENIX stellar models and Exo-Radioactive-Convective Equilibrium Model (REM) exoplanet atmosphere models and find that Exo-REM models provide the best fits to the data. From the Exo-REM models, we derive an effective temperature of $1042^{+178}_{-132}$ K for the planet and translate the derived luminosity of the planet to a mass of 5.6 ± 1.1 M Jup assuming hot-start evolutionary models. HD 143811 AB b is the first directly imaged planet around a binary that is not on an ultrawide orbit. Future characterization of this object will shed light on the formation of planets around binary star systems.

Astronomy and AstroPhysics↗
DOE OSTI2026

Use of Digital Real-Time Simulation and Optimization to Identify Maximum Real Power Injection on the Banshee Distribution Network: Preprint

This study investigates the hosting capacity of the Banshee Distribution Network by optimizing the real power injection at carefully selected Distributed Energy Resource (DER) locations. The analysis is conducted within the framework of power system operational constraints, including bus voltage ranges, thermal line ratings, and transformer loading limits. A Python-based Genetic Algorithm (GA), implemented using the PyGAD library, is employed to iteratively identify the optimal power injection configuration that maximizes network utilization while preserving system reliability. The methodology integrates a real-time simulation environment using the Real-Time Digital Simulator (RTDS), allowing high-fidelity evaluation of power flow and voltage behavior under each proposed injection scenario. By coupling the optimization algorithm with real-time simulation feedback, this approach ensures that both static and dynamic constraints are enforced during the evaluation process. The GA leverages evolutionary operators such as selection, crossover, and mutation to navigate the nonlinear search space efficiently. The results of the study delineate the feasible hosting capacity at three targeted buses, reflecting maximum real power levels that can be injected without causing voltage violations, transformer overloading, or line congestion. These findings provide a decision- support tool for distribution planners and utilities aiming to integrate higher penetrations of DERs in existing infrastructure. Additionally, the work lays the foundation for extending such optimization techniques to multi-objective formulations, including economic dispatch and reactive power coordination, in future studies.

24 POWER TRANSMISSION AND DISTRIBUTION↗
DOE OSTI2025

DEVELOPMENT AND APPLICATION OF RISK ANALYSIS TOOLKIT FOR PLANT RESOURCE OPTIMIZATION

This paper presents the development of methods and tools that are being designed to optimize plant operations (e.g., maintenance/replacement schedules and optimal maintenance postures for plant components) in a manner that is more cost effective than current approaches and makes better use of available component health and cost data. These methods include both data- and model-based optimization methods. Model-based optimization methods directly include reliability and cost models to determine an optimal plant operational strategy. We consider gradient-based and evolutionary (based on genetic algorithms) optimization methods. The second class of methods target more specific use cases (e.g., project schedule optimization) and are not based on reliability models directly, but they require specific component reliability and cost data. This class of methods is based on variants of the knapsack problem with an aim to determine an optimal project schedule that maximizes the overall NPV. This paper also presents multi-objective methods designed to identify an optimal maintenance posture based on a Pareto frontier analysis. Rather than dictating the “right” tradeoff (i.e., identify the absolute best posture), we show how it is possible to perform a trade space exploration approach (i.e., identify value and costs of several postures and let the analysis account for desired value and cost metrics). This is performed by identifying maintenance postures that maximize value (e.g., system availability) and minimize operational costs, i.e., the Pareto frontier in a value-cost trade space. For all these methods we present detailed applicative examples that show their validity from a decision-making perspective.

97 - MATHEMATICS AND COMPUTING↗
DOE OSTI2026

A New Record Census of Dwarf AGN and a Bimodal $M_{\rm BH}$-$M_{\star}$ Scaling Relation with DESI DR1

Using the first spectroscopic data release from the Dark Energy Spectroscopic Instrument (DESI DR1), we search for AGN signatures in 1,678,787 low-redshift ($0.001 \le z \le 0.45$) line-emitting galaxies. Based on the [NII]-BPT emission-line ratio diagnostic, we identify AGN in 314,245/1,211,573 (25.9%) high-mass ($\log (M_{\star}/M_{\odot}) > 9.5$) and 9648/467,214 (2.1%) dwarf ($\log (M_{\star}/M_{\odot}) \le 9.5$) galaxies. Among these AGN, 17,949 are broad-line candidates (BL-AGN) with broad H$α$ emission, enabling black hole (BH) mass estimates using single-epoch virial methods. We find that the AGN fraction in line-emitting galaxies increases monotonically with stellar mass, rising from $\sim$1.4% at the low-mass end to $\sim$93.3% at the high-mass end. Using the large BL-AGN sample, we extend the $M_{\rm BH} - M_{\star}$ scaling relation down to $\log (M_{\star}/M_{\odot}) \approx 7.8$ and $\log (M_{\rm BH}/M_{\odot}) \approx 4.4$. In the context of high-redshift overmassive BHs, our results suggest that galaxies and their central BHs may follow two distinct evolutionary pathways across cosmic time. With this paper, we release the EmFit value-added catalog, containing emission-line flux and width measurements for $\sim$7.4 million galaxies, the largest catalog with emission-line decomposition into narrow, broad, and outflow components to date. This work significantly expands upon the early DESI results and provides a statistical sample for probing the galaxy$-$BH connection in the low-mass galaxy regime.

Pucha, Ragadeepika [Utah U.; Arizona U., Astron. D↗
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)↗