Engineering Papers⌕ Search

SEARCH · Engineering Papers

Results for “tokenization”

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 109 records · Page 6

RingX: Scalable Parallel Attention for Long-Context Learning on HPC

The attention mechanism has become foundational for remarkable AI breakthroughs since the introduction of the Transformer, driving the demand for increasingly longer context to power frontier models such as large-scale reasoning language models and high-resolution image/video generators. However, its quadratic computational and memory complexities present substantial challenges. Current state-of-the-art parallel attention methods, such as ring attention, are widely adopted for long-context training but utilize a point-to-point communication strategy that fails to fully exploit the capabilities of modern HPC network architectures. In this work, we propose ringX, a scalable family of parallel attention methods optimized explicitly for HPC systems. By enhancing workload partitioning, refining communication patterns, and improving load balancing, ringX achieves up to 3.4 × speedup compared to conventional ring attention on the Frontier supercomputer. Optimized for both bi-directional and causal attention mechanisms, ringX demonstrates its effectiveness through training benchmarks of a Vision Transformer (ViT) on a climate dataset and a Generative Pre-Trained Transformer (GPT) model, Llama3 8B. Our method attains an end-to-end training speedup of approximately 1.5 × in both scenarios. To our knowledge, the achieved 38% model FLOPs utilization (MFU) for training Llama3 8B with a 1M-token sequence length on 4,096 GPUs represents one of the highest training efficiencies reported for long-context learning on HPC systems. Our code implementation is available at https://github.com/jqyin/ringX-attention.

Yin, Junqi [ORNL] (ORCID:0000000338435520)↗

Distributed Cross-Channel Hierarchical Aggregation for Foundation Models

Vision-based scientific foundation models hold significant promise for advancing scientific discovery and innovation. This potential stems from their ability to aggregate images from diverse sources—such as varying physical groundings or data acquisition systems—and to learn spatio-temporal correlations using transformer architectures. However, tokenizing and aggregating images can be compute-intensive, a challenge not fully addressed by current distributed methods. In this work, we introduce the Distributed Cross-Channel Hierarchical Aggregation (D-CHAG) approach designed for datasets with a large number of channels across image modalities. Our method is compatible with any model-parallel strategy and any type of vision transformer architecture, significantly improving computational efficiency. We evaluated D-CHAG on hyperspectral imaging and weather forecasting tasks. When integrated with tensor parallelism and model sharding, our approach achieved up to a 75% reduction in memory usage and more than doubled sustained throughput on up to 1,024 AMD GPUs on the Frontier Supercomputer.

Tsaris, Aristeidis (aris) [ORNL] (ORCID:0000000277↗

ORBIT-2: Scaling Exascale Vision Foundation Models for Weather and Climate Downscaling

Sparse observations and coarse-resolution climate models limit effective regional decision-making, underscoring the need for robust downscaling. However, existing AI methods struggle with generalization across variables and geographies and are constrained by the quadratic complexity of Vision Transformer (ViT) self-attention. We introduce ORBIT-2, a scalable foundation model for global, hyper-resolution climate downscaling. ORBIT-2 incorporates two key innovations: (1) Residual Slim ViT (Reslim), a lightweight architecture with residual learning and Bayesian regularization for efficient, robust prediction; and (2) TILES, a tile-wise sequence scaling algorithm that reduces self-attention complexity from quadratic to linear, enabling long-sequence processing and massive parallelism. ORBIT-2 scales to 10 billion parameters across 65,536 GPUs, achieving up to 4.1 ExaFLOPS sustained throughput and 74–98% strong scaling efficiency. It supports downscaling to 0.9 km global resolution and processes sequences up to 4.2 billion tokens. On 7 km resolution benchmarks, ORBIT-2 achieves high accuracy with R2 scores in range of 0.98–0.99 against observation data.

Wang, Xiao [ORNL] (ORCID:0000000165451943)↗

BrickQA: Bridging the Semantic Gap in Building Operations with Dynamic Graph Exploration

While standardized ontologies like the Brick schema address data heterogeneity in Building Automation Systems (BAS), accessing this semantic data remains a challenge as domain experts often lack the expertise to formulate complex SPARQL queries. To bridge this gap, we present BrickQA, a Large Language Model (LLM)-based framework that translates natural language into executable SPARQL queries through structured query decomposition, dynamic schema exploration, and inline validation. BrickQA utilizes an iterative reasoning agent to actively navigate graph topology through dynamic exploration actions without requiring exhaustive context injection or model fine-tuning. This approach effectively mitigates hallucinations, particularly in large-scale building knowledge graphs. Empirical evaluation on BuildingQA, a standardized benchmark, demonstrates that BrickQA significantly outperforms ReAct baselines, delivering a 0.291–0.355 absolute F1 improvement while achieving 3 × –12.7 × higher token cost-efficiency. Beyond these metrics, the framework maintains structural fidelity across heterogeneous buildings and remains resilient to ambiguous queries without requiring site-specific fine-tuning. Furthermore, a case study on operational analytics validates the framework’s capability to handle temporal and aggregation constraints, effectively transforming abstract semantic models into actionable facility management insights.1

Ko, Yun-Dam↗

Gafaelfawr

Gafaelfawr is a FastAPI application for the authorization and management of tokens, including their issuance and revocation.

Allbery, Russ↗

ORBIT-2 Weather and Climate Downscaling Software Repository

ORBIT-2 is a scalable foundation model for global, hyper-resolution climate and weather downscaling. ORBIT-2 incorporates two key innovations: (1) Residual Slim ViT (Reslim), a lightweight architecture with residual learning and Bayesian regularization for efficient, robust prediction; and (2) TILES, a tile-wise sequence scaling algorithm that reduces self-attention complexity from quadratic to linear, enabling long-sequence processing and massive parallelism. ORBIT-2 scales to 10 billion parameters across 65,536 GPUs, achieving up to 4.1 ExaFLOPS sustained throughput and 74–98% strong scaling efficiency. It supports downscaling to 0.9 km global resolution and processes sequences up to 4.2 billion tokens. On 7 km resolution benchmarks, ORBIT-2 achieves high accuracy with 𝑅2 scores in range of 0.98–0.99 against observation data.

Wang, Xiao [Oak Ridge National Laboratory]↗

genomeocean: a pretrained microbial genome foundational model (genomeoceanLLM) v1.0

We present Genomeocean, a foundational genome language model that represents the microbial genome sequences from complex environmental samples. By training on a large, diverse metagenomic dataset, Genomeocean learns species-specific sequence composition and can generate long, realistic open reading frames (ORFs). Our model employs a Byte-pair-encoding (BPE) tokenization strategy, allowing it to efficiently process large genomic datasets and generate long sequences up to 50kb. We demonstrate that fine-tuning Genomeocean can generate novel gene clusters encoding biosynthetic pathways, showcasing its ability to model both fundamental and complex biological processes. Our work establishes Genomeocean as a powerful tool for understanding microbial genome biology and paves the way for its application in a range of fields, from synthetic biology to microbiome research.

Wang, Zhong [Lawrence Berkeley National Laboratory↗

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↗

An Open-source Llm Enhanced-tool Specialized In Helping Moose Related Problems And Tasks

MOOSEenger is an open-source, terminal-first chat application for the MOOSE ecosystem that couples specialized parsing of MOOSE documentation and “.i” input files with retrieval-augmented generation to deliver grounded answers about multiphysics modeling and workflows. It includes dedicated readers for MOOSE-style HTML and a pyhit-based parser that uses the MOOSE syntax tree to preserve block structure and attach retrieval metadata. A data-ingestion pipeline performs semantic chunking into atomic facts and stores them hierarchically in a local Chroma vector database that maintains parent–child relationships across documents; the system can ingest directories, individual files, and single-page web content, and it provides CRUD operations (insert, update, delete) to manage the corpus. At query time, relevant chunks are embedded, retrieved, and fused into the model context, with interactive features such as token streaming, persistent chat history, and dynamic RAG (retrieval triggered by user input or intermediate model output). Deployment is flexible: MOOSEenger runs with local Ollama models or remote Hugging Face/OpenAI backends—typically coordinating generation, lightweight tagging/summarization, and embeddings across three models—and it also supports a server mode and integration with the VS Code Continue interface.

Li, Mengnan [Idaho National Laboratory (INL), Idah↗

PyJMAK: An Open-Source Python Toolkit for Modeling Solid-State Metallurgical Phase Transformations

Accurate prediction of metallurgical phase transformations is an essential basis for autonomous optimization and rapid part qualification. Several methods can be used to estimate the evolution of phase fractions such as JMAK kinetics-based models, phase-field models, thermodynamic models, and data-driven machine learning models. Thermodynamic and phase-field-based methodologies solve multiphysics equations requiring numerous calibration parameters and significant computational resources. As a result, the computation domain is limited to a point or on order of micron-meters. The data-driven models rely on large datasets from experiments and simulations. While the JMAK model only provides information about phase fraction evolution, it can predict this evolution in near real-time using thermal history and thermodynamic data without restriction on the domain. JMAK models have been popularly used by researchers to model phase transformations occuring during additive manufacturing or over arbitrary temperature profiles. Commercial proprietary software such as Abaqus and Ansys or closed-source in-house implementations offer the ability to model JMAK based kinetics to predict phase transformation. However, these software packages are not open-source or freely available for use and development in conjunction with manufacturing machines, sensors, and machine learning algorithms. In addition, the use of the model is restricted by a license token. In contrast, given temperature profiles at multiple points in the domain, this Python-based PyJMAK model can compute phase evolution in parallel due to its stand-alone modular, voxel-based structure, and it can be executed on high-performance computing resources without any license restrictions.

Prabhune, Bhagya [Oak Ridge National Laboratory (O↗

LATTE_SciFM

LATTE: LAtent Token Transformer for Evolutionary dynamics

Most, Alex↗

Adaptive language model training for molecular design

Abstract The vast size of chemical space necessitates computational approaches to automate and accelerate the design of molecular sequences to guide experimental efforts for drug discovery. Genetic algorithms provide a useful framework to incrementally generate molecules by applying mutations to known chemical structures. Recently, masked language models have been applied to automate the mutation process by leveraging large compound libraries to learn commonly occurring chemical sequences (i.e., using tokenization) and predict rearrangements (i.e., using mask prediction). Here, we consider how language models can be adapted to improve molecule generation for different optimization tasks. We use two different generation strategies for comparison, fixed and adaptive. The fixed strategy uses a pre-trained model to generate mutations; the adaptive strategy trains the language model on each new generation of molecules selected for target properties during optimization. Our results show that the adaptive strategy allows the language model to more closely fit the distribution of molecules in the population. Therefore, for enhanced fitness optimization, we suggest the use of the fixed strategy during an initial phase followed by the use of the adaptive strategy. We demonstrate the impact of adaptive training by searching for molecules that optimize both heuristic metrics, drug-likeness and synthesizability, as well as predicted protein binding affinity from a surrogate model. Our results show that the adaptive strategy provides a significant improvement in fitness optimization compared to the fixed pre-trained model, empowering the application of language models to molecular design tasks.

37 INORGANIC, ORGANIC, PHYSICAL, AND ANALYTICAL CH↗

Advancing representations of equity and justice in climate mitigation futures

THIS PAPER WAS PRIMARILY COMPLETED PRIOR TO THE AUTHOR JOINING PNNL AND NO DOE FUNDING WAS USED FOR THIS PAPER. In this work, we review how equity and justice issues in global climate mitigation scenarios are addressed within Integrated Assessment Models (IAMs) and propose a new research agenda to strengthen their integration in model development and application. We begin by examining prominent concerns at the science-policy interface. We introduce a typology of equity and justice limitations in climate mitigation scenarios, distinguishing among structural, methodological, and epistemological biases that shape what integrated assessment models can reveal at policy-relevant scales. Reflecting on these concerns, we propose a research agenda that describes new avenues of work and draws together distinct emerging initiatives. This agenda is based on the feasibility and depth of required interventions, from incremental improvements to structural reforms and alternative participatory approaches. Drawing on reflexive insights from integrated assessment practitioners, it addresses the operational challenges of translating justice concepts into metrics, including risks of reductionism, tokenism, and narrow definitions. Underlying this research agenda is a recognition that modeling communities must engage more critically with implicit assumptions in model design and use that have equity and justice implications. Achieving equitable climate futures will require transformative actions that integrate diverse justice concerns, advance sustainable development goals, and confront systemic inequities across both human and ecological dimensions. Although models will never capture all these aspects, they can be significantly enhanced to support more informed discussion and practical application. Our contribution proposes a way forward to achieving this goal.

Pachauri, Shonali↗

Enabling AI in synthetic biology through Construction File specification

The Construction File (CF) specification establishes a standardized interface for molecular biology operations, laying a foundation for automation and enhanced efficiency in experiment design. It is implemented across three distinct software projects: PyDNA_CF_Simulator, a Python project featuring a ChatGPT plugin for interactive parsing and simulating experiments; ConstructionFileSimulator, a field-tested Java project that showcases 'Experiment' objects expressed as flat files; and C6-Tools, a JavaScript project integrated with Google Sheets via Apps Script, providing a user-friendly interface for authoring and simulation of CF. The CF specification not only standardizes and modularizes molecular biology operations but also promotes collaboration, automation, and reuse, significantly reducing potential errors. The potential integration of CF with artificial intelligence, particularly GPT-4, suggests innovative automation strategies for synthetic biology. While challenges such as token limits, data storage, and biosecurity remain, proposed solutions promise a way forward in harnessing AI for experiment design. This shift from human-driven design to AI-assisted workflows, steered by high-level objectives, charts a potential future path in synthetic biology, envisioning an environment where complexities are managed more effectively.

59 BASIC BIOLOGICAL SCIENCES↗

Replacing non-biomedical concepts improves embedding of biomedical concepts

Embeddings are semantically meaningful representations of words in a vector space, commonly used to enhance downstream machine learning applications. Traditional biomedical embedding techniques often replace all synonymous words representing biological or medical concepts with a unique token, ensuring consistent representation and improving embedding quality. However, the potential impact of replacing non-biomedical concept synonyms has received less attention. Embedding approaches often employ concept replacement to replace concepts that span multiple words, such as non-small-cell lung carcinoma, with a single concept identifier (e.g., D002289). Also, all synonyms of each concept are merged into the same identifier. Here, we additionally leveraged WordNet to identify and replace sets of non-biomedical synonyms with their most common representatives. This combined approach aimed to reduce embedding noise from non-biomedical terms while preserving the integrity of biomedical concept representations. We applied this method to 1,055 biomedical concept sets representing molecular signatures or medical categories and assessed the mean pairwise distance of embeddings with and without non-biomedical synonym replacement. A smaller mean pairwise distance was interpreted as greater intra-cluster coherence and higher embedding quality. Embeddings were generated using the Word2Vec algorithm applied to a corpus of 10 million PubMed abstracts. Our results demonstrate that the addition of non-biomedical synonym replacement reduced the mean intra-cluster distance by an average of 8%, suggesting that this complementary approach enhances embedding quality. Future work will assess its applicability to other embedding techniques and downstream tasks. Python code implementing this method is provided under an open-source license.

algorithms↗

Extracting Material Property Measurement Data from Scientific Articles

Machine learning-based prediction of material properties is often hampered by the lack of sufficiently large training datasets. The majority of such measurement data is embedded in scientific literature and the ability to automatically extract these data is essential to support the development of reliable property prediction methods. In this work, we describe a methodology for an automatic property extraction framework using material solubility as the target property. We create an annotated dataset containing tags for solubility-related entities using a combination of regular expressions and manual tagging. We then compare five entity recognition models leveraging both token-level and span-level architectures on the task of classifying solute names, solubility values, and solubility units. Additionally, we explore a novel pretraining approach that leverages automated chemical name and quantity extraction tools to generate large datasets that do not rely on intensive manual effort. Finally, we perform an analysis to identify the causes of classification errors.

Panapitiya, Gihan U.↗

Explorations in combining cognitive models of individuals and system dynamics models of groups.

This report documents a demonstration model of interacting insurgent leadership, military leadership, government leadership, and societal dynamics under a variety of interventions. The primary focus of the work is the portrayal of a token societal model that responds to leadership activities. The model also includes a linkage between leadership and society that implicitly represents the leadership subordinates as they directly interact with the population. The societal model is meant to demonstrate the efficacy and viability of using System Dynamics (SD) methods to simulate populations and that these can then connect to cognitive models depicting individuals. SD models typically focus on average behavior and thus have limited applicability to describe small groups or individuals. On the other hand, cognitive models readily describe individual behavior but can become cumbersome when used to describe populations. Realistic security situations are invariably a mix of individual and population dynamics. Therefore, the ability to tie SD models to cognitive models provides a critical capability that would be otherwise be unavailable.

99 GENERAL AND MISCELLANEOUS↗

Constrained Versions of DEDICOM for Use in Unsupervised Part-Of-Speech Tagging

This reports describes extensions of DEDICOM (DEcomposition into DIrectional COMponents) data models [3] that incorporate bound and linear constraints. The main purpose of these extensions is to investigate the use of improved data models for unsupervised part-of-speech tagging, as described by Chew et al. [2]. In that work, a single domain, two-way DEDICOM model was computed on a matrix of bigram fre- quencies of tokens in a corpus and used to identify parts-of-speech as an unsupervised approach to that problem. An open problem identi ed in that work was the com- putation of a DEDICOM model that more closely resembled the matrices used in a Hidden Markov Model (HMM), speci cally through post-processing of the DEDICOM factor matrices. The work reported here consists of the description of several models that aim to provide a direct solution to that problem and a way to t those models. The approach taken here is to incorporate the model requirements as bound and lin- ear constrains into the DEDICOM model directly and solve the data tting problem as a constrained optimization problem. This is in contrast to the typical approaches in the literature, where the DEDICOM model is t using unconstrained optimization approaches, and model requirements are satis ed as a post-processing step.

97 MATHEMATICS AND COMPUTING↗