Engineering Papers⌕ Search

SEARCH · Engineering Papers

Results for “LLM embeddings”

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.

16 records
All sourcesNASA NTRSDOE OSTI
DOE OSTI2025

Finch: Toxicity Dose Response Curve Prediction of Chemical Compounds and Mixtures

A paradigm shift in chemical risk assessment is emphasizing mixture testing over single compound analysis, eliminating animal testing, and adopting advanced modeling approaches to understand mixture activity profiles. However, existing computational models largely focus on single chemicals, with few effective solutions for modeling complex mixtures that account for synergistic or antagonistic effects and multiple Modes of Action (MoA). Conventional methods like concentration addition (CA) and independent action (IA) are insufficient for this task as they are designed for simplistic interactions and struggle to account for the dynamic and multifaceted nature of chemical mixtures, such as overlapping MoA and non-linear interactions. Finch offers a novel approach utilizing deep learning (DL) embeddings and multi-task quantitative structure-activity relationship (QSAR) models to improve chemical exposure prediction. By leveraging molecular descriptors, physiochemical properties, and large language model (LLM) embeddings from SMILES inputs, Finch preserves critical information in a latent space thereby enhancing predictive accuracy. The multi-task learning aspect of Finch is highly advantageous, as it simultaneously optimizes multiple loss functions, leveraging all available data across tasks to develop generalized representations that effectively capture complex ingredient interactions within mixtures.

59 BASIC BIOLOGICAL SCIENCES↗
DOE OSTI2026

Knowledge Graph of RB-Tnseq Data from Fitness Browser (KP-DP1)

Motivation: Predicting microbial gene fitness across environmental conditions remains a central challenge for predictive phenomics and autonomous experimentation. Fitness assays generate large volumes of genotype–phenotype measurements difficult to integrate with experimental metadata and biological function in a form that supports mechanistic reasoning. Knowledge graphs offer a semantic framework for unifying modalities and enabling context-aware inference. Results: We build GIMME (Graph Inference for Microbial Metabolism Exploration), a semantically grounded knowledge graph that unifies gene fitness measurements spanning 10 Pseudomonas species with experimental metadata and biological context. Media are decomposed into chemical components and experiments carry structured links to natural-language descriptions. The resulting graph supports two inference modes: (1) symbolic graph traversal to surface candidate gene–environment and gene–chemical associations, and (2) learned inference using heterogeneous graph neural networks that propagate information across neighborhoods. We formulate link regression over (gene, media, experiment) triplets, combining learned gene embeddings with pretrained LLM sourced text embeddings of node descriptions to predict gene fitness. We then augment a baseline MLP with an auxiliary message-passing encoder (GraphSAGE/GAT) that propagates information over gene–protein–function and media–chemical subgraphs, and fuse the two pathways with a gated residual connection. This approach produces strong agreement with held-out fitness measurements (GraphSAGE Pearson r 0.74) while also highlighting inference challenges in extreme-fitness regimes. We aggregate GAT edge-attention weights by relation type and layer to estimate which biological and environmental relations most influence fitness predictions. Conclusion: This work explores using knowledge graphs as “context graphs” for microbial phenotype prediction. They provide a rich substrate which enables explainable retrieval of supporting evidence, and provides a natural bridge to autonomous workflows that prioritize the next experiment.

59 BASIC BIOLOGICAL SCIENCES↗
DOE OSTI2026

Leveraging large language models to address data scarcity in machine learning for graphene synthesis

Machine learning in experimental materials science faces significant challenges due to the scarcity of data, which are costly and time-consuming to generate, particularly when relying on in-house experiments. Literature data mining offers a potential solution but introduces issues like mixed data quality, inconsistent formats, and non-uniform reporting of synthesis parameters, resulting in partially missing and heterogeneous features across the dataset. Here, we propose data imputation and feature engineering methods that employ pre-trained large language models (LLMs) to enhance machine learning performance on scarce, heterogeneous datasets, demonstrated on graphene CVD synthesis data and the ML-HydPARK hydrogen storage dataset. GPT models perform data imputation via tailored prompting and semantic normalization of inconsistently reported features through embeddings, for example, to harmonize the complex nomenclature of CVD substrates. Beyond yielding more diverse and richer feature representations than traditional methods such as K-nearest neighbors (KNN) and Multivariate Imputation by Chained Equations (MICE), LLM-based data imputation is evaluated against dataset characteristics and prompting strategies. We vary the level of autonomy granted to the LLM, from generic prompting that leverages pre-trained knowledge for autonomous data generation to data-informed prompting that constrains outputs using target-specific information, and demonstrate which level of autonomy yields superior imputation performance across datasets and feature types. The proposed data engineering methods markedly improve downstream performance; for example, in graphene layer number classification using a support vector machine (SVM), binary accuracy increases from 39% to 65% and ternary accuracy from 52% to 72%. Fine-tuning experiments on both datasets show that combining our proposed LLM-based data imputation and feature encoding methods with numerical machine learning predictors outperforms standalone fine-tuned LLM predictors in data-scarce settings. The proposed strategies emphasize data enhancement techniques rather than refining learning architectures or regularizing loss functions, offering a broadly applicable framework for improving machine learning performance on scarce, inhomogeneous datasets.

Chemical vapor deposition↗
DOE OSTI2026

XMark: Reliable Multi-Bit Watermarking for LLM-Generated Texts

Multi-bit watermarking has emerged as a promising solution for embedding imperceptible binary messages into Large Language Model (LLM)-generated text, enabling reliable attribution and tracing of malicious usage of LLMs. Despite recent progress, existing methods still face key limitations: some become computationally infeasible for large messages, while others suffer from a poor trade-off between text quality and decoding accuracy. Moreover, the decoding accuracy of existing methods drops significantly when the number of tokens in the generated text is limited, a condition that frequently arises in practical usage. To address these challenges, we propose XMark, a novel method for encoding and decoding binary messages in LLM-generated texts. The unique design of XMark’s encoder produces a less distorted logit distribution for watermarked token generation, preserving text quality, and also enables its tailored decoder to reliably recover the encoded message with limited tokens. Extensive experiments across diverse downstream tasks show that XMark significantly improves decoding accuracy while preserving the quality of watermarked text, outperforming prior methods. The code will be made publicly available upon acceptance.

Xu, Jiahao [University of Nevada, Reno]↗
DOE OSTI2025

Explainable Synthesizability Prediction of Inorganic Crystal Polymorphs Using Large Language Models

Abstract We evaluate the ability of machine learning to predict whether a hypothetical crystal structure can be synthesized and explain those predictions to scientists. Fine‐tuned large language models (LLMs) trained on a human‐readable text description of the target crystal structure perform comparably to previous bespoke convolutional graph neural network methods, but better prediction quality can be achieved by training a positive‐unlabeled learning model on a text‐embedding representation of the structure. An LLM‐based workflow can then be used to generate human‐readable explanations for the types of factors governing synthesizability, extract the underlying physical rules, and assess the veracity of those rules. These explanations can guide chemists in modifying or optimizing non‐synthesizable hypothetical structures to make them more feasible for materials design.

Kim, Seongmin [Department of Chemical and Biologic↗
DOE OSTI2025

Explainable Synthesizability Prediction of Inorganic Crystal Polymorphs Using Large Language Models

Abstract We evaluate the ability of machine learning to predict whether a hypothetical crystal structure can be synthesized and explain those predictions to scientists. Fine‐tuned large language models (LLMs) trained on a human‐readable text description of the target crystal structure perform comparably to previous bespoke convolutional graph neural network methods, but better prediction quality can be achieved by training a positive‐unlabeled learning model on a text‐embedding representation of the structure. An LLM‐based workflow can then be used to generate human‐readable explanations for the types of factors governing synthesizability, extract the underlying physical rules, and assess the veracity of those rules. These explanations can guide chemists in modifying or optimizing non‐synthesizable hypothetical structures to make them more feasible for materials design.

Kim, Seongmin [Department of Chemical and Biologic↗
DOE OSTI2026

CodeScribe Agent

SF-26-086 CodeScribe introduces a structured, multi-stage pipeline that combines deterministic program analysis with LLM-powered translation to enable incremental, testable Fortran-to-C++ migration. First, `code-scribe index` traverses the project directory tree and produces `scribe.yaml` metadata files recording all modules, subroutines, and functions at each level, giving the LLM accurate structural context instead of a hallucinated codebase model. Second, `code-scribe draft` performs the deterministic portion of translation — converting Fortran types to C++ equivalents, replacing `use` statements with `#include` and `using namespace` directives, and detecting constructs requiring special handling — while embedding`scribe-prompt` annotations that guide the LLM through non-trivial cases such as statement-function-to-lambda conversions and `extern "C"` wrapper generation. Third, `code-scribe translate` applies project-specific TOML-based few-shot prompt templates and submits the composed prompt to a pluggable LLM backend (OpenAI, Anthropic, Argonne ARGO, any OpenAI-compatible endpoint, or local Hugging Face checkpoints), producing a C++ source file, a header, and a Fortran-C++ interface file for each translated routine so the codebase compiles and runs correctly throughout the migration. Beyond translation, CodeScribe includes a tool-using coding agent (`code-scribe agent`) with read, bash, edit, and write capabilities, and a bounded loop mode (`code-scribe loop`) that runs repeated stateless agent sessions over a task file with restricted tool access — enabling sustained, auditable software development workflows for broader scientific computing tasks.

Dhruv, Akash [Argonne National Laboratory (ANL), A↗
DOE OSTI2026

Dispersion Loss Counteracts Embedding Condensation and Improves Generalization in Small Language Models

Large language models (LLMs) achieve remarkable performance through ever-increasing parameter counts, but scaling incurs steep computational costs. To better understand LLM scaling, we study representational differences between LLMs and their smaller counterparts, with the goal of replicating the representational qualities of larger models in smaller models. We observe a geometric phenomenon which we term embedding condensation, where token embeddings collapse into a narrow cone-like subspace in some language models. Through systematic analyses across multiple Transformer families, we show that small models such as GPT2 and Qwen3-0.6B exhibit severe condensation, whereas larger models such as GPT2-x1 and Qwen3-32B are more resistant to this phenomenon. Additional observations show that embedding condensation is not reliably mitigated by knowledge distillation from larger models. To fight against it, we formulate a dispersion loss that explicitly encourages embedding dispersion during training. Experiments demonstrate that it mitigates condensation, recovers dispersion patterns seen in larger models, and yields performance gains across 10 benchmarks. We believe this work offers a principled path toward improving smaller Transformers without additional parameters.

Xiao, Xi [ORNL] (ORCID:0009000009316982)↗
DOE OSTI2025

CAFE AU LAIT: Compute-Aware Federated Augmented Low-Rank AI Training

Federated finetuning is crucial for unlocking the knowledge embedded in pretrained Large Language Models (LLMs) when data are geographically distributed across clients. Unlike finetuning with data from a single institution, federated finetuning allows collaboration across multiple institutions, enabling the utilization of diverse and decentralized datasets while preserving data privacy. Given the high computing costs of LLM training and the emphasis on energy efficiency in Federated Learning (FL), Low-Rank Adaptation (LoRA) has emerged as a widely adopted algorithm due to its significantly reduced number of trainable parameters. However, this assumes that all data silos have the necessary computing resources to compute local updates of LLMs. Nevertheless, in practice, the computing resources across clients are highly heterogeneous: while some may have access to hundreds of GPUs, others might have limited or no GPU access. Recently, federated finetuning using synthetic data has been proposed, allowing clients to participate in a collaborative training run without training LLMs locally. However, our experimental results reveal a performance gap between models trained using synthetic data and those trained using local updates. Motivated by the observed heterogeneity in computing resources and the performance gap, we propose a novel two-stage algorithm that leverages the storage and computing capabilities of a strong server. In the first stage, under the coordination of the strong server, clients with limited computing resources collaborate to generate synthetic data, which is transferred to and stored on the strong server. In the second stage, the strong server uses this synthetic data on behalf of the resource-constrained clients to perform federated LoRA finetuning alongside clients with sufficient computing resources. This approach ensures that all clients can participate in the finetuning process. Experimental results demonstrate that incorporating local updates from even a small fraction of clients improves performance compared to using synthetic data for all clients. Furthermore, we incorporate the Gaussian mechanism in both stages to guarantee client-level differential privacy.

Wang, Jiayi [ORNL]↗
DOE OSTI2025

Leveraging large language models to automate the identification of healthcare access barriers for veterans

Objective: To develop and evaluate an automated system for identifying healthcare barriers focusing on transportation issues in veterans’ clinical notes using large language models (LLMs) and to assess the impact of different prompting strategies on classification performance and explanation consistency. Methods: We developed a hybrid system combining pattern matching for templated notes with LLM analysis for free-text notes. Using 2000 manually annotated clinical notes, we compared four prompting strategies (dual-role short, dual-role long, analysis-first, analysis-only) across Mistral-7B and Llama-3.1 models. We evaluated classification performance using standard metrics and assessed explanation consistency through embedding similarity analysis. Results: The analysis-first strategy achieved superior performance, with Mistral-7B reaching an F1 score of 0.914, outperforming traditional machine learning approaches (GBM: 0.786, BERT: 0.811). LLMs demonstrated higher explanation consistency within models (mean cosine similarity 0.887–0.908) compared to cross-model similarities (0.767–0.872). Pattern matching successfully handled 6.7% of templated notes deterministically. Mistral-7B showed greater internal consistency but higher abstention rates compared to Llama-3.1. Conclusion: Requiring LLMs to analyze evidence before classification improves both accuracy and explanation consistency for identifying transportation barriers in clinical notes. This approach enables automated barrier detection at scale while providing clinically relevant explanations, supporting both population-level healthcare planning and individual patient care decisions.

Healthcare access barriers↗
DOE OSTI2026

An AI-driven framework for evaluating local and state authorities’ permitting processes

The demand for new energy infrastructure is increasing across the United States, but heterogenous permitting processes and embedded requirements across different local jurisdictions can cause project delays, increase “soft costs,” and hinder developer expansion. This study analyzes the variability in local permitting requirements across the U.S. and develops a quantitative approach to describe their clarity and effectiveness in enabling infrastructure project development. By using an Energy Language Model (ELM), a large language model (LLM) for energy technologies, we systematically gathered permitting information from nearly 300 state-, county-, and city-level documents, creating a structured dataset of requirements and procedures on an unprecedented scale and speed. Our analysis revealed that local (city and county) permitting requirement documents are underrepresented compared to state-level guidance documents, which can impede timely and cost-effective installation of new electric infrastructure. Our validation process showed that the final database has an accuracy of approximately 95%. We, further, created a new quantitative method to score permitting requirements for clarity and efficiency, with electric vehicle supply equipment as an initial use case. The average local permitting document scored a 1.8 out of 5, which we interpret as meaning that half of the requirements developers face when installing electric infrastructure are ambiguous, increasing both cost and time. We also created a “Generalized Permit Process”, highlighting common procedural steps and identifying specific opportunities for municipalities to improve their documentation. This research establishes a systematic and scalable framework for evaluating the complexities of local infrastructure permitting processes by combining LLM-powered data collection and quantitative scoring. The framework enables policymakers and developers to identify and mitigate procedural bottlenecks, with the expectation that these improvements can accelerate application review and approval, reduce project costs, and expedite connection to utility distribution grids. As a foundational approach for streamlining local project development processes, this study’s methods are intended to be extended to a wide range of energy applications.

29 ENERGY PLANNING, POLICY, AND ECONOMY↗
DOE OSTI2025

Comparative Analysis of DNA LLM Classification Techniques Using Intra-Layer Feature Extraction with Autoencoder Stacks [Poster]

This project conducts a comparative analysis of DNA LLM classification techniques using Evo2, Grover, and UTRML, focusing on intra-layer feature extraction in Evo2. By extracting features from multiple layers of Evo2 and integrating them into an autoencoder stack with a binary classification head, we evaluate its effectiveness in classifying genomic sequences compared to smaller DNA language models. My findings demonstrate that Evo2 outperforms Grover and UTRML in classification accuracy on a dataset provided by department 08625, CAO2021, while UTRML offers competitive performance with lower computational costs. This study highlights the potential of advanced embedding techniques in enhancing genomic data analysis and informs future research in bioinformatics.

59 BASIC BIOLOGICAL SCIENCES↗
DOE OSTI2025

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↗
DOE OSTI2026

The ballad of LLM agents: philosophical reasoning for chemistry

Large language models (LLMs) show remarkable potential for scientific reasoning but often produce unreliable or scientifically unactionable outputs when faced with multi-step logic, domain grounding, and interpretability challenges, especially in complex fields like chemistry and materials science. Here, we introduce a framework of philosophical reasoning agents, inspired by canonical thinkers such as Socrates, Descartes, Kant, and Hume, to guide LLM behavior via structured prompt engineering. These agents embody distinct reasoning paradigms (dialectical inquiry, deductive logic, rule-based judgment, and empirical validation) and are evaluated across multiple chemistry subdomains, physical, analytical, general, inorganic, and organic chemistry, using the ChemBench benchmark. Our agentic prompting approach yields substantial accuracy gains on open-ended numerical chemistry questions, with gains of +11.5 percentage points for GPT-4o with Hume, +4.5 percentage points for GPT-5 with Kant, and +21.8 percentage points for GPT-5.1 with Socrates at the strict 1% error threshold, relative to the corresponding base models. Beyond accuracy, we observe benchmark-level model–agent performance patterns, suggesting that different prompting styles interact differently with each base model. These findings demonstrate that embedding philosophy-of-science principles into multi-agent frameworks can improve and produce interpretable, adaptive, and domain-aligned scientific LLMs.

Harb, Hassan [Argonne National Laboratory (ANL), A↗
DOE OSTI2026

Opportunities in AI/ML for the Rubin LSST Dark Energy Science Collaboration

The Vera C. Rubin Observatory's Legacy Survey of Space and Time (LSST) will produce unprecedented volumes of heterogeneous astronomical data (images, catalogs, and alerts) that challenge traditional analysis pipelines. The LSST Dark Energy Science Collaboration (DESC) aims to derive robust constraints on dark energy and dark matter from these data, requiring methods that are statistically powerful, scalable, and operationally reliable. Artificial intelligence and machine learning (AI/ML) are already embedded across DESC science workflows, from photometric redshifts and transient classification to weak lensing inference and cosmological simulations. Yet their utility for precision cosmology hinges on trustworthy uncertainty quantification, robustness to covariate shift and model misspecification, and reproducible integration within scientific pipelines. This white paper surveys the current landscape of AI/ML across DESC's primary cosmological probes and cross-cutting analyses, revealing that the same core methodologies and fundamental challenges recur across disparate science cases. Since progress on these cross-cutting challenges would benefit multiple probes simultaneously, we identify key methodological research priorities, including Bayesian inference at scale, physics-informed methods, validation frameworks, and active learning for discovery. With an eye on emerging techniques, we also explore the potential of the latest foundation model methodologies and LLM-driven agentic AI systems to reshape DESC workflows, provided their deployment is coupled with rigorous evaluation and governance. Finally, we discuss critical software, computing, data infrastructure, and human capital requirements for the successful deployment of these new methodologies, and consider associated risks and opportunities for broader coordination with external actors.

Aubourg, Eric [APC, Paris] (ORCID:000000025592023X↗