Search NASA⌕ Search

SEARCH · Search NASA

Results for “ATTENTION”

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 145 records · Page 8
All sourcesNASA NTRSDOE OSTI
DOE OSTI2024

Adaptive Patching for High-resolution Image Segmentation with Transformers

Attention-based models are proliferating in the space of image analytics, including segmentation. The standard method of feeding images to transformer encoders is to divide the images into patches and then feed the patches to the model as a linear sequence of tokens. For high-resolution images, e.g. microscopic pathology images, the quadratic compute and memory cost prohibits the use of an attention-based model, if we are to use smaller patch sizes that are favorable in segmentation. The solution is to either use custom complex multi-resolution models or approximate attention schemes. We take inspiration from Adapative Mesh Refinement (AMR) methods in HPC by adaptively patching the images, as a pre-processing step, based on the image details to reduce the number of patches being fed to the model, by orders of magnitude. This method has a negligible overhead, and works seamlessly with any attention-based model, i.e. it is a pre-processing step that can be adopted by any attention-based model without friction. We demonstrate superior segmentation quality over SoTA segmentation models for realworld pathology datasets while gaining a geomean speedup of 6.9× for resolutions up to 64K2, on up to 2, 048 GPUs.

Zhang, Enzhi↗
DOE OSTI2026

MATEY: multiscale adaptive transformer models for spatiotemporal physical systems

Accurate representation of the multiscale features in spatiotemporal physical systems using vision transformer architectures requires extremely long, computationally prohibitive token sequences. To address this issue, we propose two novel adaptive tokenization schemes that dynamically adjust patch sizes based on local features: one ensures convergent behavior to uniform patch refinement, while the other offers better computational efficiency. Moreover, we present a set of spatiotemporal attention schemes, where the temporal or axial spatial dimensions are decoupled, to evaluate their baseline computational and data efficiencies and to determine whether adaptive tokenization can improve this performance. We assess the performance of the proposed multiscale adaptive model, MATEY, in a sequence of experiments. Compared to a full spatiotemporal attention scheme or a scheme that decouples only the temporal dimension, we find that fully decoupled axial attention is less efficient and expressive, requiring more training time and model parameters to achieve the same accuracy. The experiments on the adaptive tokenization schemes show that, compared to a uniformly refined model, the proposed schemes achieve comparable or improved accuracy at a much lower cost in the tested two-dimensional settings. While the asymptotic analysis suggests the potential for favorable scaling, empirical validation at substantially longer sequence lengths remains to be performed in future work. Finally, we demonstrate in two fine-tuning tasks featuring different physics that models pretrained on PDEBench data outperform the ones trained from scratch, especially in the low data regime with frozen attention.

adaptive tokenization↗
DOE OSTI2021

Donald J. Trump’s Presidency in Cyberspace: A Case Study of Social Perception and Social Influence in Digital Oligarchy Era

In the past few years, with the rapid growth of digital technologies, Facebook, Twitter, and other social media platforms have become the digital oligarchies, which have the enormous capabilities to potentially control what is discussed in cyberspace. In the digital oligarchy era, social perception and social influence in different complex social systems have evolved quickly. In this article, we conducted large-scale empirical studies on social perception and social influence regarding the Trump phenomenon from personal perception, media, and public attention perspectives. We found that there exist obvious correlations between the posting behavior of Trump and the attention of news media. By constructing public attention networks using complex networks based on Google search information, we further reveal that digital platforms could affect social perception and social influence significantly. Especially, we obtained that the public attention can always be influenced by the political moments.

digital oligarchy↗
DOE OSTI2023

A comparison of histopathology imaging comprehension algorithms based on multiple instance learning

Whole slide imaging (WSI), also called digital virtual microscopy, is a new imaging modality. It allows for the application of AI and machine learning methods to cancer pathology to help establish a means for the automatic diagnosis of cancer cases. However, designing machine-learning models for WSI is computationally challenging due to its required ultra-high resolution. The current state-of-the-art models use multiple instance learning (MIL). MIL is a weakly-supervised learning method in which the model uses an array of inferences from many smaller instances to make a final classification about the entire set. In the context of WSI, researchers divide the ultra-high-resolution image into many patches. The model then classifies the slide based on an array of inferences from the patches. Among several ways of making the final classification, attention-based mechanisms have resulted in superb accuracy scores. The Transformer, one attention-based algorithm, has reported substantial improvements for WSI comprehension tasks. In this project, we studied and compared several WSI comprehension algorithms. We used the following three datasets: CAMELYON16+17, TCGALung, and TCGA-Kidney. We found that attention-based MIL algorithms performed better than standard MIL algorithms for classifying WSI images, achieving a higher mean accuracy and AUC. However, none of the attention-based algorithms performed significantly better than the others, reporting accuracy scores that varied widely. Presumably, it is due to the limited availability of training samples in the data corpus. Since it is not easy to increase the samples from human subjects, some machine learning techniques like transfer learning could help mitigate this issue.

Saunders, Adam↗
DOE OSTI2021

Integrated Framework of Vehicle Dynamics, Instabilities, Energy Models, and Sparse Flow Smoothing Controllers

This work presents an integrated framework of: vehicle dynamics models, with a particular attention to instabilities and traffic waves; vehicle energy models, with particular attention to accurate energy values for strongly unsteady driving profiles; and sparse Lagrangian controls via automated vehicles, with a focus on controls that can be executed via existing technology such as adaptive cruise control systems. This framework serves as a key building block in developing control strategies for human-in-the-loop traffic flow smoothing on real highways. In this contribution, we outline the fundamental merits of integrating vehicle dynamics and energy modeling into a single framework, and we demonstrate the energy impact of sparse flow smoothing controllers via simulation results.

Lee, Jonathan W.↗
DOE OSTI2024

STFM: Accurate Spatio-Temporal Fusion Model for Weather Forecasting

Meteorological prediction is crucial for various sectors, including agriculture, navigation, daily life, disaster prevention, and scientific research. However, traditional numerical weather prediction (NWP) models are constrained by their high computational resource requirements, while the accuracy of deep learning models remains suboptimal. In response to these challenges, we propose a novel deep learning-based model, the Spatiotemporal Fusion Model (STFM), designed to enhance the accuracy of meteorological predictions. Our model leverages Fifth-Generation ECMWF Reanalysis (ERA5) data and introduces two key components: a spatiotemporal encoder module and a spatiotemporal fusion module. The spatiotemporal encoder integrates the strengths of convolutional neural networks (CNNs) and recurrent neural networks (RNNs), effectively capturing both spatial and temporal dependencies. Meanwhile, the spatiotemporal fusion module employs a dual attention mechanism, decomposing spatial attention into global static attention and channel dynamic attention. This approach ensures comprehensive extraction of spatial features from meteorological data. The combination of these modules significantly improves prediction performance. Experimental results demonstrate that STFM excels in extracting spatiotemporal features from reanalysis data, yielding predictions that closely align with observed values. In comparative studies, STFM outperformed other models, achieving a 7% improvement in ground and high-altitude temperature predictions, a 5% enhancement in the prediction of the u/v components of 10 m wind speed, and an increase in the accuracy of potential height and relative humidity predictions by 3% and 1%, respectively. This enhanced performance highlights STFM’s potential to advance the accuracy and reliability of meteorological forecasting.

54 ENVIRONMENTAL SCIENCES↗
DOE OSTI2021

Hardware-Based Emulator with Deep Learning Model for Building Energy Control and Prediction Based on Occupancy Sensors’ Data

Heating, ventilation, and air conditioning (HVAC) is the largest source of residential energy consumption. Occupancy sensors’ data can be used for HVAC control since it indicates the number of people in the building. HVAC and sensors form a typical cyber-physical system (CPS). In this paper, we aim to build a hardware-based emulation platform to study the occupancy data’s features, which can be further extracted by using machine learning models. In particular, we propose two hardware-based emulators to investigate the use of wired/wireless communication interfaces for occupancy sensor-based building CPS control, and the use of deep learning to predict the building energy consumption with the sensor data. We hypothesize is that the building energy consumption may be predicted by using the occupancy data collected by the sensors, and question what type of prediction model should be used to accurately predict the energy load. Another hypothesis is that an in-lab hardware/software platform could be built to emulate the occupancy sensing process. The machine learning algorithms can then be used to analyze the energy load based on the sensing data. To test the emulator, the occupancy data from the sensors is used to predict energy consumption. The synchronization scheme between sensors and the HVAC server will be discussed. We have built two hardware/software emulation platforms to investigate the sensor/HVAC integration strategies, and used an enhanced deep learning model—which has sequence-to-sequence long short-term memory (Seq2Seq LSTM)—with an attention model to predict the building energy consumption with the preservation of the intrinsic patterns. Because the long-range temporal dependencies are captured, the Seq2Seq models may provide a higher accuracy by using LSTM architectures with encoder and decoder. Meanwhile, LSTMs can capture the temporal and spatial patterns of time series data. The attention model can highlight the most relevant input information in the energy prediction by allocating the attention weights. The communication overhead between the sensors and the HVAC control server can also be alleviated via the attention mechanism, which can automatically ignore the irrelevant information and amplify the relevant information during CNN training. Our experiments and performance analysis show that, compared with the traditional LSTM neural network, the performance of the proposed method has a 30% higher prediction accuracy.

Ye, Zhijing↗
DOE OSTI2024

Interpreting Transformers for Jet Tagging

Machine learning (ML) algorithms, particularly attention-based transformer models, have become indispensable for analyzing the vast data generated by particle physics experiments like ATLAS and CMS at the CERN LHC. Particle Transformer (ParT), a state-of-the-art model, leverages particle-level attention to improve jet-tagging tasks, which are critical for identifying particles resulting from proton collisions. This study focuses on interpreting ParT by analyzing attention heat maps and particle-pair correlations on the $\eta$-$\phi$ plane, revealing a binary attention pattern where each particle attends to at most one other particle. At the same time, we observe that ParT shows varying focus on important particles and subjets depending on decay, indicating that the model learns traditional jet substructure observables. These insights enhance our understanding of the model's internal workings and learning process, offering potential avenues for improving the efficiency of transformer architectures in future high-energy physics applications.

72 PHYSICS OF ELEMENTARY PARTICLES AND FIELDS↗
DOE OSTI2021

Publication and Impact of Preprints Included in the First 100 Editions of the CDC COVID-19 Science Update: Content Analysis

Preprints are publicly available manuscripts posted to various servers that have not been peer reviewed. Although preprints have existed since 1961, they have gained increased popularity during the COVID-19 pandemic due to the need for immediate, relevant information. The aim of this study is to evaluate the publication rate and impact of preprints included in the Centers for Disease Control and Prevention (CDC) COVID-19 Science Update and assess the performance of the COVID-19 Science Update team in selecting impactful preprints. All preprints in the first 100 editions (April 1, 2020, to July 30, 2021) of the Science Update were included in the study. Preprints that were not published were categorized as “unpublished preprints.” Preprints that were subsequently published exist in 2 versions (in a peer-reviewed journal and on the original preprint server), which were analyzed separately and referred to as “peer-reviewed preprint” and “original preprint,” respectively. Time to publish was the time interval between the date on which a preprint was first posted and the date on which it was first available as a peer-reviewed article. Impact was quantified by Altmetric Attention Score and citation count for all available manuscripts on August 6, 2021. Preprints were analyzed by publication status, publication rate, preprint server, and time to publication. Of the 275 preprints included in the CDC COVID-19 Science Update during the study period, most came from three servers: medRxiv (n=201, 73.1%), bioRxiv (n=41, 14.9%), and SSRN (n=25, 9.1%), with 8 (2.9%) coming from other sources. Additionally, 152 (55.3%) were eventually published. The median time to publish was 2.3 (IQR 1.4-3.7). When preprints posted in the last 2.3 months were excluded (to account for the time to publish), the publication rate was 67.8%. Moreover, 76 journals published at least one preprint from the CDC COVID-19 Science Update, and 18 journals published at least three. The median Altmetric Attention Score for unpublished preprints (n=123, 44.7%) was 146 (IQR 22-552) with a median citation count of 2 (IQR 0-8); for original preprints (n=152, 55.2%), these values were 212 (IQR 22-1164) and 14 (IQR 2-40), respectively; for peer-review preprints, these values were 265 (IQR 29-1896) and 19 (IQR 3-101), respectively. Prior studies of COVID-19 preprints found publication rates between 5.4% and 21.1%. Preprints included in the CDC COVID-19 Science Update were published at a higher rate than overall COVID-19 preprints, and those that were ultimately published were published within months and received higher attention scores than unpublished preprints. These findings indicate that the Science Update process for selecting preprints had a high fidelity in terms of their likelihood to be published and their impact. The incorporation of high-quality preprints into the CDC COVID-19 Science Update improves this activity’s capacity to inform meaningful public health decision-making.

59 BASIC BIOLOGICAL SCIENCES↗
DOE OSTI2025

Using Eye Tracking to Elucidate the Mechanisms Underlying Stimulation-Enhanced Visual Target Detection

Transcranial direct current stimulation (tDCS) is a noninvasive form of brain stimulation that involves passing a weak electrical current between electrodes on the scalp to modulate underlying neural tissue. TDCS has been shown to modulate cognition in a variety of domains, including memory, attention, and visual processing. Prior work from our laboratory has shown positive effects of tDCS on learning to detect target objects hidden in complex naturalistic visual scenes and learn rules for categorizing images, though the mechanism for these benefits remains unknown. One possibility is that tDCS optimizes visual search by modulating visual attention or via the reduction in search errors. One method of quantifying visual attention is to use eye tracking to record search patterns to determine if and how visual search is adjusted under verum stimulation conditions. Eye tracking data allows classification of errors into error types, including sampling errors (failing to look in the relevant region), recognition errors (looking at the critical portion of a scene, but failing to recognize it as such as evidenced by visual fixation), and decision-making errors (fixating on the relevant portion of a scene, but making the wrong determination). Our results indicate that the benefit tDCS confers on visual search for targets stems from the reduction in decision-making errors when targets are present (Cohen’s d = 0.86). Also reported is a replication of previous findings showing a tDCS-dependent improvement in learning this task, learning score (Cohen’s d = 0.88); d’ (Cohen’s d = 1.00). This provides support for moving tDCS into the application space by pairing it with analysts who are concerned with the type of search error that is corrected via stimulation.

attention↗
DOE OSTI2021

Industrial Energy Assessment Training Effectiveness Evaluation: An Eye-Tracking Study

It is essential to understand the effectiveness of any training program so it can be improved accordingly. Various studies have applied standard metrics for the evaluation of visual behavior to recognize the areas of interest that attract individuals’ attention as there is a high correlation between attentional behavior and where one is focusing on. However, through reviewing the literature, we believe that studies that applied eye-tracking technologies for training purposes are still limited, especially in the industrial energy assessment training field. In this paper, the effectiveness of industrial energy assessment training was quantitatively evaluated by measuring the attentional allocation of trainees using eye-tracking technology. Moreover, this study identifies the areas that require more focus based on evaluating the performance of subjects after receiving the training. Additionally, this research was conducted in a controlled environment to remove the distractions that may be caused by environmental factors to only concentrate on variables that influence the learning behavior of subjects. The experiment results showed that after receiving the training, the subjects’ performance in energy assessment was significantly improved in two areas: production, and recycling and waste management, and the designed training program enhanced the knowledge of participants in identifying energy-saving opportunities to the knowledge level of experienced participants.

32 ENERGY CONSERVATION, CONSUMPTION, AND UTILIZATI↗
DOE OSTI2023

TransformerG2G: Adaptive time-stepping for learning temporal graph embeddings using transformers

Dynamic graph embedding has emerged as a very effective technique for addressing diverse temporal graph analytic tasks (i.e., link prediction, node classification, recommender systems, anomaly detection, and graph generation) in various applications. Such temporal graphs exhibit heterogeneous transient dynamics, varying time intervals, and highly evolving node features throughout their evolution. Hence, incorporating long-range dependencies from the historical graph context plays a crucial role in accurately learning their temporal dynamics. In this paper, we develop a graph embedding model with uncertainty quantification, TransformerG2G, by exploiting the advanced transformer encoder to first learn intermediate node representations from its current state (t) and previous context (over timestamps [t–1,t–l], l is the length of context). Moreover, we employ two projection layers to generate lower-dimensional multivariate Gaussian distributions as each node's latent embedding at timestamp t. We consider diverse benchmarks with varying levels of "novelty" as measured by the TEA (Temporal Edge Appearance) plots. Here, our experiments demonstrate that the proposed TransformerG2G model outperforms conventional multi-step methods and our prior work (DynG2G) in terms of both link prediction accuracy and computational efficiency, especially for high degree of novelty. Furthermore, the learned time-dependent attention weights across multiple graph snapshots reveal the development of an automatic adaptive time stepping enabled by the transformer. Importantly, by examining the attention weights, we can uncover temporal dependencies, identify influential elements, and gain insights into the complex interactions within the graph structure. For example, we identified a strong correlation between attention weights and node degree at the various stages of the graph topology evolution.

97 MATHEMATICS AND COMPUTING↗
DOE OSTI2026

Transformer-based operator learning framework for self-energy in strongly correlated systems

We introduce Σ-Attention, a transformer-based operator-learning framework for approximating the self-energy operator of strongly correlated electronic systems. By creating a batched dataset that combines results from three complementary approaches, i.e., many-body perturbation theory, strong-coupling expansion, and exact diagonalization, each effective in specific parameter regimes, Σ-Attention is applied to learn an accurate approximation for the self-energy operator that is valid across a wide range of parameter regimes. This hybrid strategy leverages the strengths of existing methods while relying on the transformer's ability to generalize beyond individual limitations. More importantly, the scalability of the transformer architecture allows the learned self-energy to be extended to systems with larger sizes, leading to much improved computational scaling. Using the one-dimensional Hubbard model, we demonstrate that Σ-Attention can accurately predict the Matsubara Green's function of large systems with a wide range of coupling strength. Our framework offers a promising and scalable pathway for studying strongly correlated systems with many possible generalizations.

Zhu, Yuanran↗
DOE OSTI2026

FFTSF: Revisiting Sub-Seasonal Streamflow Forecasting with Simple Feedforward Network

Accurate short-to-subseasonal streamflow forecasts are vital for water management, including flood preparedness, drought mitigation, hydropower scheduling, and ecosystem protection. However, extending a forecast beyond a few days remains challenging due to complexity of hydrological processes. While recent self-attention based transformer architectures such as iTransformer have gained traction in time-series forecasting, these models suffer from several critical limitations: (1) significant computational overhead that scales quadratically with sequence length, (2) vulnerability to overfitting on limited hydrological datasets, (3) degraded performance on long-horizon forecasts due to attention decay, and (4) excessive architectural complexity that hampers interpretability and operational deployment. In this study, we propose a simple Feedforward Time Series Forecasting (FFTSF) network that directly addresses these limitations through its lightweight architecture and long-range forecasting capabilities. We evaluate FFTSF across 178 USGS stream gauges spanning diverse climate regimes by forecasting lead times of 1-, 7-, 14-, and 30-days. Our results demonstrate that FFTSF achieves competitive performance at short lead times (NSE of 0.778 for 1-day forecasts) while substantially outperforming complex baselines at longer forecast period, achieving the highest NSE (0.271) at 30-day forecasts with greater robustness and stability. For 30-day forecasts, FFTSF achieves a 71% improvement over NLinear, 57% improvement over DLinear and 12% improvement over the computationally intensive iTransformer while requiring fewer computational resources. Our findings reveal that architectural complexity is not necessary for hydrological forecasting, demonstrating that well-designed simple models can outperform attention mechanisms for subseasonal streamflow forecasting. The computational efficiency and consistent long-range performance of FFTSF make it suitable for water management applications where reliable extended forecasts are essential.

Krishnan Kutty Ambika, Anukesh [ORNL] (ORCID:00000↗
DOE OSTI2022

Background-Aware 3-D Point Cloud Segmentation With Dynamic Point Feature Aggregation

With the proliferation of LiDAR sensors and 3-D vision cameras, 3-D point cloud analysis has attracted significant attention in recent years. In this article, we propose a novel 3-D point cloud learning network, referred to as dynamic point feature aggregation network (DPFA-Net), by selectively performing the neighborhood feature aggregation (FA) with dynamic pooling and an attention mechanism. DPFA-Net has two variants for semantic segmentation and classification of 3-D point clouds. As the core module of the DPFA-Net, we propose an FA layer, in which features of the dynamic neighborhood of each point are aggregated via a self-attention mechanism. In contrast to other segmentation models, which aggregate features from fixed neighborhoods, our approach can aggregate features from different neighbors in different layers providing a more selective and broader view to the query points and focusing more on the relevant features in a local neighborhood. In addition, to further improve the performance of semantic segmentation, we exploit the background–foreground (BF) information and present two novel approaches, namely, two-stage BF-Net and BF regularization. Experimental results show that the proposed DPFA-Net achieves the state-of-the-art overall accuracy score of 89.22% for semantic segmentation on the Stanford large-scale 3-D Indoor Spaces (S3DIS) dataset and provides consistently satisfactory performance across different tasks of semantic segmentation, part segmentation, and 3-D object classification. Furthermore, our model achieves 93.1% accuracy on the ModelNet40 dataset and provides a mean shape intersection-over-union (IoU) value of 85.5% for part segmentation on the ShapeNet-Part dataset. It is a also computationally more efficient compared to other methods.

3-D↗
DOE OSTI2025

Knowledge Oriented Graph Unified Transformer (KOGUT) v0.1

KOGUT — Knowledge Oriented Graph Unified Transformer KOGUT implements the Relational Graph Transformer (RelGT) architecture for knowledge graph link prediction in biological domains, with a primary focus on microbial growth media prediction. While the original RelGT (arXiv:2505.10960) targets relational tables, time series, and multi-table databases, KOGUT adapts this architecture for heterogeneous biological knowledge graphs, providing first-in-class AI predictive models for microbial cultivation. Key Adaptations Beyond Original RelGT: - Knowledge Graph Focus: Applied to biological KGs with semantic node types (taxa, chemicals, media, phenotypes, environments) versus generic relational database tables, trained on the KG-Microbe knowledge graph (1.3M entities, 2.9M edges, 24 relation types). - Multimodal Node Encoding: Integrates node labels, categories, descriptions, and synonyms from KG metadata through learned embedding layers—adapting relational column features to graph node attributes with textual semantics. - Extended K-Hop Subgraph Strategy: Optimized neighborhood sampling (3-hop default, configurable up to 200 nodes) tuned for sparse biological networks, building on the original local-global attention framework with biological relation preservation. - Biolink Predicate Preservation: Type-specific transformations for 24 biological edge semantics (occurs_in, consumes, produces, has_phenotype, subclass_of) beyond standard relational foreign keys, enabling multi-relation link prediction. - Inductive Learning Support: Enables zero-shot predictions for novel taxa through feature-based embeddings (temperature, oxygen requirements, gram stain, cell shape), extending the original transductive relational benchmark scope to uncultured microorganisms. CheapSOTA Performance Optimizations (This Distribution): - VQ-EMA Centroid Attention: Vector quantization with exponential moving average for improved global context modeling (+5-10% MRR improvement). - HDF5 Precomputed Data Loading: One-time preprocessing of k-hop subgraphs to eliminate redundant graph traversals (2-5× training speedup). - Distributed Data Parallel Training: Multi-GPU support for scaling to larger knowledge graphs (tested on 4× NVIDIA A100 GPUs at NERSC Perlmutter). - Mixed Precision Training: Automatic mixed precision (AMP) for memory efficiency and faster training. Advantages Over Standard Knowledge Graph Embedding Models: Combines RelGT's proven multi-element tokenization (features, type, hop, structure) with graph-native biological representations, enabling interpretable link prediction across heterogeneous entities that standard embedding models (TransE, RotatE, ComplEx) and table-based transformers cannot directly model. Achieves near-perfect performance on microbial growth media prediction (MRR: 0.9966, Precision@1: 0.9932, Hit@10: 1.0000) while maintaining explainability through attention-based reasoning over biological pathways. Training Data: - KG-Microbe merged knowledge graph: 1,379,337 nodes, 2,960,472 edges - 24 biological relation types including taxonomic hierarchies, metabolic interactions, phenotype associations, and environmental relationships - Primary prediction task: Growth media suitability for microbial taxa (biolink:occurs_in, 50K edges) - Multi-relation capability: Predicts links for any of the 24 relation types, including chemical consumption/production, phenotype associations, and taxonomic classification Citation: Original RelGT Architecture: Dwivedi et al., "Relational Graph Transformer", arXiv:2505.10960, 2025 KOGUT Implementation: Knowledge Oriented Graph Unified Transformer for Microbial Growth Media Prediction Developed at Lawrence Berkeley National Laboratory (LBNL) Trained on NERSC Perlmutter supercomputer

Joachimiak, Marcin [Lawrence Berkeley National Lab↗
DOE OSTI2025

Spatially Aware Linear Transformer (SAL-T) for Particle Jet Tagging

Transformers are very effective in capturing both global and local correlations within high-energy particle collisions, but they present deployment challenges in high-data-throughput environments, such as the CERN LHC. The quadratic complexity of transformer models demands substantial resources and increases latency during inference. In order to address these issues, we introduce the Spatially Aware Linear Transformer (SAL-T), a physics-inspired enhancement of the linformer architecture that maintains linear attention. Our method incorporates spatially aware partitioning of particles based on kinematic features, thereby computing attention between regions of physical significance. Additionally, we employ convolutional layers to capture local correlations, informed by insights from jet physics. In addition to outperforming the standard linformer in jet classification tasks, SAL-T also achieves classification results comparable to full-attention transformers, while using considerably fewer resources with lower latency during inference. Experiments on a generic point cloud classification dataset (ModelNet10) further confirm this trend. Our code is available at https://github.com/aaronw5/SAL-T4HEP.

Wang, Aaron [Illinois U., Chicago] (ORCID:00000003↗
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)↗