SEARCH · Engineering Papers

Results for “Large Language Models (LLMs)”

Search indexed NASA NTRS and DOE OSTI research on propulsion, heat transfer, battery materials and energy systems. Follow report and document links to the original sources.

SourceResearch connectionFrom yearThrough year

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

At least 73 records · Page 4

All sources NASA NTRS DOE OSTI

DOE OSTI2025

GEPA: Reflective Prompt Evolution Can Outperform Reinforcement Learning

Large language models (LLMs) are increasingly adapted to downstream tasks via reinforcement learning (RL) methods like Group Relative Policy Optimization (GRPO), which often require thousands of rollouts to learn new tasks. We argue that the interpretable nature of language often provides a much richer learning medium for LLMs, compared to policy gradients derived from sparse, scalar rewards. To test this, we introduce GEPA (Genetic-Pareto), a prompt optimizer that thoroughly incorporates natural language reflection to learn high-level rules from trial and error. Given any AI system containing one or more LLM prompts, GEPA samples trajectories (e.g., reasoning, tool calls, and tool outputs) and reflects on them in natural language to diagnose problems, propose and test prompt updates, and combine complementary lessons from the Pareto frontier of its own attempts. As a result of GEPA's design, it can often turn even just a few rollouts into a large quality gain. Across six tasks, GEPA outperforms GRPO by 6% on average and by up to 20%, while using up to 35x fewer rollouts. GEPA also outperforms the leading prompt optimizer, MIPROv2, by over 10% (e.g., +12% accuracy on AIME-2025), and demonstrates promising results as an inference-time search strategy for code optimization. We release our code at https://github.com/gepa-ai/gepa.

97 MATHEMATICS AND COMPUTING↗

DOE OSTI2026

Scaling open-weight large language models for hydropower regulatory information extraction: A systematic analysis

Information extraction from regulatory and technical documents using large language models (LLMs) involves practical trade-offs between extraction quality and computational cost. We evaluate eight open-weight LLMs spanning 0.6B–70B parameters on hydropower licensing documents and report deployment-oriented evidence under a unified extraction schema and evaluation protocol. Across the model set, we observe clear scale-dependent trends in both baseline extraction quality and the effectiveness of reflective reasoning (self-checking) under our fixed-prompt, no-augmentation setting. Mid-scale models often provide a favorable balance of accuracy and efficiency, whereas the smallest models show limited or inconsistent gains from the reasoning variants tested. Larger models achieve the highest overall F1 scores but incur substantially greater compute and infrastructure requirements. We further find that reliability failure modes can distort conventional metrics in this domain: in particular, high recall can coincide with systematic extraction errors when models fabricate values for fields that are absent from the source text, underscoring the importance of conservative null handling and evidence-grounded evaluation. Overall, our study provides a reproducible resource–performance comparison for open-weight LLM-based extraction in hydropower regulatory documentation and offers practical guidance for model selection under different deployment constraints.

Evaluation protocol↗

DOE OSTI2024

Prompt Phrase Ordering Using Large Language Models in HPC: Evaluating Prompt Sensitivity

Large language models (LLMs) have demonstrated effective performance in domain-specific tasks, often requiring a well-designed prompt to guide their responses. However, optimizing the right prompt is challenging due to prompt sensitivity—the phenomenon where small changes in the prompt can lead to significant variations in performance. In this study, we evaluate prompt performance by examining all permutations of independent phrases to investigate prompt sensitivity and robustness. We used two datasets: the GSM8k dataset, which assesses mathematical reasoning, and a custom template prompt for summarizing database metadata. Our goal was to evaluate the performance across all permutations of a sequence of prompt phrases. The study was conducted using the llama3-instruct- 7B model hosted on Ollama, with computations parallelized in a high-performance computing environment. By comparing the average index of phrases in the best and worst-performing prompts, we found that the order of independent phrases within a prompt significantly impacts LLM performance. Additionally, we used Hamming distance to assess changes between phrase orderings, concluding that prompt modifications can dramatically affect scores, often by almost random chance. These findings support existing research on prompt sensitivity. We discuss the challenges of prompt optimization, noting that altering phrases in a successful prompt does not always result in another successful prompt.

97 MATHEMATICS AND COMPUTING↗

DOE OSTI2025

Leveraging Large Language Models for Real-World Data Evidence: A Framework for Automated Treatment Extraction and Data Harmonization

Background: The ability to comprehensively collect treatment information from cancer patient medical records would enable studies to evaluate real-world benefits and risks tied to specific treatments. Currently, it is difficult to system- atically collect high-quality treatment information because it is often stored in unstructured text. Manually extracting and standardizing drug and regimen data is time-intensive. Recent advances in large language models (LLMs) offer a potential solution for automated extraction of structured treatment information from clinical text. Objective: This study systematically evaluates the utility of four LLMs from the Llama family for automated extraction of oncology treatment information from clinical text. This information can guide researchers using cancer registry data to provide insights into cancer care and outcomes beyond clinical trials. Methods: Four instruction-tuned Llama models with varying parameter counts (1B, 3B, 8B, and 70B) were evaluated for their ability to extract treatment information from clinical documents. A unified oncology knowledge base integrating seven major public data sources was developed to standardize and normalize extracted entities—a critical step for harmonizing data from diverse sources. Extracted treatment data were compared against expert-annotated ground truth. Model performance was assessed using accuracy metrics (Precision, Recall, F1-Score) and opera- tional feasibility metrics, including processing speed and structural compliance of the output. Results: A strong positive correlation was observed between model size and extraction accuracy. F1-score improved from 0.609 for the 1B model to 0.710 (3B), 0.807 (8B), and 0.828 (70B). While larger models demonstrated superior accuracy and compliance, they incurred higher computational costs. The modest performance difference between 8B and 70B suggests diminishing returns with increasing model size. Conclusions: LLMs represent a viable technology for automating oncology treatment extraction. The 8B-parameter model emerged as a highly effective option, balancing high accuracy and computational efficiency. Selecting an appropriate LLM for deployment in cancer registries involves a trade-off between desired accuracy and available operational resources. Harmonizing extracted entities with the oncology knowledge base facilitates standardized integration into common data models, enhancing data quality for real-world evidence analyses.

artificial intelligence↗

DOE OSTI2026

Generative large language models for predictive maintenance planning

Maintenance planning and the generation of necessary components for tasks can prove time-consuming and complex. Automating the creation of recurring or similar tasks by leveraging previous planning packages and data, while uncovering insights to automate planning package generation, presents an opportunity to conserve valuable time and resources. This work aims to harness the textual and probabilistic capabilities of large language models (LLMs) to automate the generation of planning packages. Utilizing diverse data sources ranging from raw data to handwritten text, both singular and collaborative LLMs are trained and tested. Results demonstrate their capability to generate essential planning package components, effectively replicating the statistical patterns in the data. This demonstrates the use of these tools inside a digital asset for automated planning. This work outlines a methodology for constructing datasets, a training suite, and evaluation methods for LLM-based textual and conversational planning tools utilized in an asset digital twin. Results indicate that the fine-tuned models generate estimated planning information within the statistical ranges observed in real maintenance data. The models achieve high accuracy (>90%) in document question-answering and instruction generation tasks. Furthermore, the conversational retrieval-augmented generation (RAG) assistant system achieves 100% document retrieval accuracy, while conversational information capture exceeds 98% across the majority of work-package assistant modules.

97 MATHEMATICS AND COMPUTING↗

DOE OSTI2026

A Benchmarking Framework for Evaluating Large Language Model Capabilities in Nuclear Reactor Safety Applications

Large language models (LLMs) are increasingly capable of answering technical questions, synthesizing domain knowledge, and supporting engineering workflows. For nuclear science and engineering, these capabilities require careful, domain-specific evaluation before they can be credibly incorporated into safety-related activities, regulatory review, or technical decision support. This paper presents preliminary results from benchmarking framework for evaluating LLM capabilities in nuclear contexts. The framework is organized into three evaluation categories: nuclear fundamentals, general dual-use knowledge, and plant specific knowledge. These categories are intended to distinguish general nuclear engineering competence from broader technical reasoning and more context-dependent nuclear knowledge. Initial evaluations focus on nuclear fundamentals using questions representative of the knowledge expected of a nuclear professional engineer. Results indicate that contemporary frontier models perform at a high level and substantially exceed the performance of older model generations, with some models approaching saturation of the current benchmark. These findings suggest both the rapid improvement of LLM capabilities in specialized technical domains and the need for more discriminating evaluation methods. The paper presents the benchmark structure, preliminary model-comparison results, and ongoing work. This work supports development of verifiable, responsible, and safety-conscious methods for assessing AI systems in nuclear engineering applications.

21 - SPECIFIC NUCLEAR REACTORS AND ASSOCIATED PLAN↗

DOE OSTI2024

Optimizing Distributed Training on Frontier for Large Language Models

Large language models (LLMs) have demonstrated remarkable success as foundational models, benefiting various downstream applications through fine-tuning. Loss scaling studies have demonstrated the superior performance of larger LLMs compared to their smaller counterparts. Nevertheless, training LLMs with billions of parameters poses significant challenges and requires considerable computational resources. For example, training a one trillion parameter GPT-style model on 20 trillion tokens requires a staggering 120 million exaflops. This research explores efficient distributed training strategies to extract this computation from Frontier, the world's first exascale supercomputer. We enable and investigate various model and data parallel training techniques, such as tensor parallelism, pipeline parallelism, and sharded data parallelism, to facilitate training a trillion-parameter model on Frontier. We empirically assess these techniques and their associated parameters to determine their impact on memory footprint, communication latency, and GPU's computational efficiency. We analyze the complex interplay among these techniques and find a strategy to combine them to achieve high throughput through hyperparameter tuning. We have identified efficient strategies for training large LLMs of varying sizes through empirical analysis and hyperparameter tuning. For 22 Billion, 175 Billion, and 1 Trillion parameters, we achieved GPU throughputs of 38.38%, 36.14%, and 31.96%, respectively. For the training of the 175 Billion parameter model and the 1 Trillion parameter model, we achieved 100% weak scaling efficiency on 1024 and 3072 Mi250X GPUs, respectively. We also achieved strong scaling efficiencies of 89% and 87% for these two models. We trained these models only tens of iterations instead of training till completion.

Yin, Junqi↗

DOE OSTI2025

Trustworthiness and Trust: Identifying Factors that Drive Successful Human-AI Interaction in Nuclear Power Plant Applications

Emerging technologies such as artificial intelligence (AI) and machine learning (ML) are rapidly evolving and considered a promising tool for efficient and continued safe operations of the U.S. nuclear power plants (NPPs). Emerging AI techniques like large language models (LLMs) are one such technology that may support personnel at existing NPPs perform work more efficiently. For example, operators may query the current operational status of a power plant via a chat interface leveraging LLMs to access plant-related information in an interactive manner rather than manually collecting various sensor data for tasks such as surveillances or completing work orders. This is a fundamental shift in the way operators currently perform their tasks today. The literature of human-automation interaction indicates that trust is a crucial factor that drives successful interaction between a human operator and an automated system, like an AI-infused NPP application. This work presents the results of a literature review on key factors that relate to trust in AI/LLM technologies for NPP applications. The relevant literature of human factors and cognitive engineering has identified various factors related to trust including trustworthiness, performance characteristics, operator skill and perceived risk. This preliminary literature review will guide development and evaluation of models involving the identified factors influencing trust in AI and develop a framework for human-centered design for interface between humans and AI. By addressing trust, this work supports developing a technical basis for designing key characteristics of AI/LLM to support calibrated trust, which will ultimately support wide-scale adoption of AI/LLM technologies, as well as ensure safe, effective, and reliable use.

99 - GENERAL AND MISCELLANEOUS↗

DOE OSTI2024

Dynamic Retrieval Augmented Generation of Ontologies using Artificial Intelligence (DRAGON-AI)

Ontologies are fundamental components of informatics infrastructure in domains such as biomedical, environmental, and food sciences, representing consensus knowledge in an accurate and computable form. However, their construction and maintenance demand substantial resources and necessitate substantial collaboration between domain experts, curators, and ontology experts. We present Dynamic Retrieval Augmented Generation of Ontologies using AI (DRAGON-AI), an ontology generation method employing Large Language Models (LLMs) and Retrieval Augmented Generation (RAG). DRAGON-AI can generate textual and logical ontology components, drawing from existing knowledge in multiple ontologies and unstructured text sources.We assessed performance of DRAGON-AI on de novo term construction across ten diverse ontologies, making use of extensive manual evaluation of results. Our method has high precision for relationship generation, but has slightly lower precision than from logic-based reasoning. Our method is also able to generate definitions deemed acceptable by expert evaluators, but these scored worse than human-authored definitions. Notably, evaluators with the highest level of confidence in a domain were better able to discern flaws in AI-generated definitions. We also demonstrated the ability of DRAGON-AI to incorporate natural language instructions in the form of GitHub issues.These findings suggest DRAGON-AI's potential to substantially aid the manual ontology construction process. However, our results also underscore the importance of having expert curators and ontology editors drive the ontology generation process.

96 KNOWLEDGE MANAGEMENT AND PRESERVATION↗

DOE OSTI2023

Evaluation of pre-training large language models on leadership-class supercomputers

Large language models (LLMs) have arisen rapidly to the center stage of artificial intelligence as the foundation models applicable to many downstream learning tasks. However, how to effectively build, train, and serve such models for many high-stake and first-principle-based scientific use cases are both of great interests and of great challenges. Moreover, pre-training LLMs with billions or even trillions of parameters can be prohibitively expensive not just for academic institutions, but also for well-funded industrial and government labs. Furthermore, the energy cost and the environmental impact of developing LLMs must be kept in mind. Here, in this work, we conduct a first-of-its-kind performance analysis to understand the time and energy cost of pre-training LLMs on the Department of Energy (DOE)’s leadership-class supercomputers. Employing state-of-the-art distributed training techniques, we evaluate the computational performance of various parallelization approaches at scale for a range of model sizes, and establish a projection model for the cost of full training. Our findings provide baseline results, best practices, and heuristics for pre-training such large models that should be valuable to HPC community at large. We also offer insights and optimization strategies for using the first exascale computing system, Frontier, to train models of the size of GPT-3 and beyond.

97 MATHEMATICS AND COMPUTING↗

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 OSTI2024

Powering Data Centers with Clean Energy: A Techno-Economic Case Study of Nuclear and Renewable Energy Dependability

Rising data demands from artificial intelligence (AI) and large language models (LLMs) generating images, videos, and text have prompted increased need for larger and more robust data centers in the United States. Major companies interested in these larger data centers face the choice of linking them to existing regional grids, building stand-alone power supplies onsite, or a combination of both. The request, review, and approval process for new transmission lines to grids in the United States, however, has grown in recent years to times spans rivaling those of new construction for nuclear power plants. Building an islanded power supply for each data center is therefore becoming a prominent option. In this case study, several technologies are modeled in techno-economic simulations for long-term system costs subject to fixed electricity demand from a singular data center. A 250 MWe data center is assumed with additional 50 MWe for resiliency. Techno-economic simulations are conducted using the Holistic Energy Resource Optimization Network (HERON) software, which is a part of the Framework for Optimization of Resources and Economics (FORCE) tool suite. Technologies considered include solar, wind, lithium-ion batteries, and several types of nuclear reactors: large-scale reactors, small modular reactors, and microreactors. A low- and high-cost estimate for each technology is assumed to develop a range of expected economic performance. Low-cost estimates included several clean energy production tax credits. Different combinations of renewable energy generators with nuclear reactors are considered, ranging from a fully renewable-powered data center to a fully nuclear-powered data center. Historic time series of wind and solar availability from the Texas grid are used to train a reduced order model; this model then generates unique time series with similar characteristics of the training dataset. Multiple scenarios of weather and subsequent operations are simulated for each renewable-nuclear combination to determine total costs throughout the project lifetime. Fully renewable-powered configurations required large amounts of installed capacity (GW scale) in the simulations to meet the fixed demand of the data center. This is due to some scenarios in the historical dataset which captured low-wind and low-solar days, requiring over-building of these technologies as well as batteries to compensate for the low amounts of electricity generation. Fully nuclear-powered configurations outperformed the fully renewable and mixed renewable-nuclear configurations in terms of cost, with ranges between $1B and $10B in 2023 USDs compared to $40B+ for fully renewable configurations. Of the nuclear technologies, small modular reactors performed better economically than large-scale nuclear models due to lower projected capital costs, and both performed better than the microreactor models. These results demonstrate the applicability of firm, dispatchable electricity resources from baseload generators like nuclear power plants for operating facilities that run at constant power without daily variability.

22 GENERAL STUDIES OF NUCLEAR REACTORS↗

DOE OSTI2024

Towards Next-Generation Urban Decision Support Systems through AI-Powered Construction of Scientific Ontology Using Large Language Models—A Case in Optimizing Intermodal Freight Transportation

The incorporation of Artificial Intelligence (AI) models into various optimization systems is on the rise. However, addressing complex urban and environmental management challenges often demands deep expertise in domain science and informatics. This expertise is essential for deriving data and simulation-driven insights that support informed decision-making. In this context, we investigate the potential of leveraging the pre-trained Large Language Models (LLMs) to create knowledge representations for supporting operations research. By adopting ChatGPT-4 API as the reasoning core, we outline an applied workflow that encompasses natural language processing, Methontology-based prompt tuning, and Generative Pre-trained Transformer (GPT), to automate the construction of scenario-based ontologies using existing research articles and technical manuals of urban datasets and simulations. From these ontologies, knowledge graphs can be derived using widely adopted formats and protocols, guiding various tasks towards data-informed decision support. The performance of our methodology is evaluated through a comparative analysis that contrasts our AI-generated ontology with the widely recognized pizza ontology, commonly used in tutorials for popular ontology software. We conclude with a real-world case study on optimizing the complex system of multi-modal freight transportation. Our approach advances urban decision support systems by enhancing data and metadata modeling, improving data integration and simulation coupling, and guiding the development of decision support strategies and essential software components.

96 KNOWLEDGE MANAGEMENT AND PRESERVATION↗

DOE OSTI2025

LLM Benchmarking with LLaMA2: Evaluating Code Development Performance Across Multiple Programming Languages

The rapid evolution of large language models (LLMs) has opened new possibilities for automating various tasks in software development. This paper evaluates the capabilities of the LLaMA 2-70B model in automating these tasks for scientific applications written in commonly used programming languages. Using representative test problems, we assess the model's capacity to generate code, documentation, and unit tests, as well as its ability to translate existing code between commonly used programming languages. Our comprehensive analysis evaluates the compilation, runtime behavior, and correctness of the generated and translated code. Additionally, we assess the quality of automatically generated code, documentation, and unit tests. Here, our results indicate that while LLaMA 2-70B frequently generates syntactically correct and functional code for simpler numerical tasks, it encounters substantial difficulties with more complex, parallelized, or distributed computations, requiring considerable manual corrections. We identify key limitations and suggest areas for future improvements to better leverage AI-driven automation in scientific computing workflows.

97 MATHEMATICS AND COMPUTING↗

DOE OSTI2026

Agentic framework for programmatic crystal structure generation using a fine-tuned worker–supervisor large language model

Platinum group metals (PGMs) underpin many catalytic technologies but face severe supply constraints, motivating the search for alternative materials and computational methods to accelerate discovery. While atomistic simulation tools such as Pymatgen and ASE have streamlined structure manipulation, they require detailed inputs, limiting accessibility for experimentalists and slowing early-stage exploration. Here, in this study, we present an AI-driven agentic framework that orchestrates worker–supervisor large language models (LLMs). The worker translates natural-language prompts of varying abstraction into valid crystallographic structures using a compact LLM fine-tuned with low-rank adaptation on a curated text–code–CIF dataset, emphasizing energy-efficient training. Benchmarking against the baseline CodeGen-350M-mono model shows that fine-tuning reduces hallucination rates from 100% to as low as 5% and improves structural match accuracy to up to 82% for fully specified inputs. Accuracy declines with decreasing prompt detail but remains nontrivial even when only stoichiometry and space group are provided, underscoring the LLM’s capacity for crystallographic inference. The supervisor Claude LLM evaluates the outputs and triggers iterative refinement through the worker’s built-in structure manipulation capabilities (e.g., supercell scaling, strain, vacancy, and substitution operations). We further demonstrate use cases for technologically relevant catalysts, including IrO 2 , pyrochlore Pb 2 Ir 2 O 7 , Ni 2 FeO 4 , and Ni 3 Mo, where the framework generates physically consistent structures that can be refined via geometry optimization. This work introduces a low-energy, language-driven pathway for integrating human and machine intelligence in materials design, paving the way for AI-assisted synthesis planning and high-throughput screening of complex oxides.

AI agent↗

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

Kim, Seongmin [Department of Chemical and Biologic↗

DOE OSTI2024

chatHPC: Empowering HPC users with large language models

The ever-growing number of pre-trained large language models (LLMs) across scientific domains presents a challenge for application developers. While these models offer vast potential, fine-tuning them with custom data, aligning them for specific tasks, and evaluating their performance remain crucial steps for effective utilization. However, applying these techniques to models with tens of billions of parameters can take days or even weeks on modern workstations, making the cumulative cost of model comparison and evaluation a significant barrier to LLM-based application development. To address this challenge, we introduce an end-to-end pipeline specifically designed for building conversational and programmable AI agents on high performance computing (HPC) platforms. Our comprehensive pipeline encompasses: model pre-training, fine-tuning, web and API service deployment, along with crucial evaluations for lexical coherence, semantic accuracy, hallucination detection, and privacy considerations. Here, we demonstrate our pipeline through the development of chatHPC, a chatbot for HPC question answering and script generation. Leveraging our scalable pipeline, we achieve end-to-end LLM alignment in under an hour on the Frontier supercomputer. We propose a novel self-improved, self-instruction method for instruction set generation, investigate scaling and fine-tuning strategies, and conduct a systematic evaluation of model performance. The established practices within chatHPC will serve as a valuable guidance for future LLM-based application development on HPC platforms.

97 MATHEMATICS AND COMPUTING↗