Search NASA⌕ Search

SEARCH · Search NASA

Results for “Interpretable Artificial Intelligence”

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.

214 records · Page 12
All sourcesNASA NTRSDOE OSTI
NASA NTRSDate not supplied

TechEdSat-13: The First Flight of an AI/ML Neuromorphic Processor

Neuromorphic processors, inspired by the wiring of the brain, permit certain classes of Artificial Intelligence/Machine Learning (AI/ML) algorithms to run far more efficiently. Ultimately, such systems will make the small- and nano-satellite platforms even more useful in terms of greatly improved power, communication and internal data management. In addition, on-board processing of images and data will help to not only rapidly interpret the information, but also reduce the amount of data that needs to be transmitted to ground stations. This initial flight experiment uses the Intel/ Loihi processor combined with a custom interface board and three communication channels to run the AI/ML scripts. These will vary with increasing length and complexity during the course of the mission. The algorithms will use the (at first) limited sensor data to ‘learn’ - with comparisons to similar architecture in comparable ground experiments. Some of the applications for successor flights in the TES-n flight series include Cognitive Communications, whereby the overall communication system is optimized per overflight – by optimizing timing and data transmission functions. Lastly, the performance of the Loihi 14nm process technology will be monitored for performance in the LEO radiation environment, thus looking for induced hardware and software errors. This information will help guide future radiation protection techniques to extend the lifetime and overall utility. The TES-13 is a 3U nanosat successfully launched by the Virgin Orbit Launcher-1 on January 13, 2022, and will presage more flights and AI/ML applications to come.

TechEdSat↗
DOE OSTI2024

Using AI Segmentation Models to Improve Foreign Body Detection and Triage from Ultrasound Images

Medical imaging can be a critical tool for triaging casualties in trauma situations. In remote or military medicine scenarios, triage is essential for identifying how to use limited resources or prioritize evacuation for the most serious cases. Ultrasound imaging, while portable and often available near the point of injury, can only be used for triage if images are properly acquired, interpreted, and objectively triage scored. Here, we detail how AI segmentation models can be used for improving image interpretation and objective triage evaluation for a medical application focused on foreign bodies embedded in tissues at variable distances from critical neurovascular features. Ultrasound images previously collected in a tissue phantom with or without neurovascular features were labeled with ground truth masks. These image sets were used to train two different segmentation AI frameworks: YOLOv7 and U-Net segmentation models. Overall, both approaches were successful in identifying shrapnel in the image set, with U-Net outperforming YOLOv7 for single-class segmentation. Both segmentation models were also evaluated with a more complex image set containing shrapnel, artery, vein, and nerve features. YOLOv7 obtained higher precision scores across multiple classes whereas U-Net achieved higher recall scores. Using each AI model, a triage distance metric was adapted to measure the proximity of shrapnel to the nearest neurovascular feature, with U-Net more closely mirroring the triage distances measured from ground truth labels. Overall, the segmentation AI models were successful in detecting shrapnel in ultrasound images and could allow for improved injury triage in emergency medicine scenarios.

47 OTHER INSTRUMENTATION↗
NASA NTRS2022

TechEdSat-13: The First Flight of a Neuromorphic Processor

Neuromorphic processors, inspired by the wiring of the brain, permit certain classes of Artificial Intelligence/Machine Learning (AI/ML) algorithms to run far more efficiently. Ultimately, such systems will make the small- and nano-satellite platforms even more useful in terms of greatly improved power, communication and internal data management. In addition, on-board processing of images and data will help to not only rapidly interpret the information, but also reduce the amount of data that needs to be transmitted to ground stations. This initial flight experiment uses the Intel/Loihi processor combined with a custom interface board and three communication channels to run the AI/ML scripts. These will vary with increasing length and complexity during the course of the mission. The algorithms will use the (at first) limited sensor data to ‘learn’ - with comparisons to similar architecture in comparable ground experiments. Some of the applications for successor flights in the TES-n flight series include Cognitive Communications, whereby the overall communication system is optimized per overflight – by optimizing timing and data transmission functions. Lastly, the performance of the Loihi 14nm process technology will be monitored for performance in the LEO radiation environment, thus looking for induced hardware and software errors. This information will help guide future radiation protection techniques to extend the lifetime and overall utility. The TES-13 is a 3U nanosat successfully launched by the Virgin Orbit Launcher-1 on January 13, 2022, and will presage more flights and AI/ML applications to come.

TechEdSat↗
DOE OSTI2024

Jet classification using high-level features from anatomy of top jets

Recent advancements in deep learning models have significantly enhanced jet classification performance by analyzing low-level features (LLFs). However, this approach often leads to less interpretable models, emphasizing the need to understand the decision-making process and to identify the high-level features (HLFs) crucial for explaining jet classification. To address this, we consider the top jet tagging problems and introduce an analysis model (AM) that analyzes selected HLFs designed to capture important features of top jets. Our AM mainly consists of the following three modules: a relation network analyzing two-point energy correlations, mathematical morphology and Minkowski functionals for generalizing jet constituent multiplicities, and a recursive neural network analyzing subjet constituent multiplicity to enhance sensitivity to subjet color charges. We demonstrate that our AM achieves performance comparable to the Particle Transformer (ParT) while requiring fewer computational resources in a comparison of top jet tagging using jets simulated at the hadronic calorimeter angular resolution scale. Furthermore, as a more constrained architecture than ParT, the AM exhibits smaller training uncertainties because of the bias-variance tradeoff. We also compare the information content of AM and ParT by decorrelating the features already learned by AM. Lastly, we briefly comment on the results of AM with finer angular resolution inputs.

72 PHYSICS OF ELEMENTARY PARTICLES AND FIELDS↗
DOE OSTI2022

MetaText: Compositional Generalization in Deep Language Models

Compositional generalization, the ability to infer and reason using novel combinations of previously encountered entities and structures, is a trait of great utility across a variety of deep learning tasks; it implies a certain ability to reason using consistent and therefore interpretable rules, to understand the constituent parts of its input at multiple levels of granularity, to be robust to spurious differences between semantically equivalent inputs, and more. We seek to understand the degree to which existing natural language models achieve or fall short of compositional generalization across a variety of tasks, whether there are distinct types of compositional generalization in practice, and to identify avenues of intervention through which we can improve models’ compositional generalization ability. After exploring these questions using an array of models, tasks, and potential interventions we see that large, pretrained language models have encountered sufficient training data to account for a variety of fine-grained compositional behavior but still struggle to reason at the level of phrases or larger language structures. Non-intrusive, data-based interventions in the form of augmenting individual sequences with compressed versions of themselves or deriving new examples from induced grammars prove insufficient to encourage greater levels of compositional reasoning, indicating that future work might benefit most from focusing on changes to a model’s inductive bias at the architectural or loss level, or by integrating compositionality-boosting data interventions into the large-scale pretraining process itself.

97 MATHEMATICS AND COMPUTING↗
DOE OSTI2023

Using Ultrasound Image Augmentation and Ensemble Predictions to Prevent Machine-Learning Model Overfitting

Deep learning predictive models have the potential to simplify and automate medical imaging diagnostics by lowering the skill threshold for image interpretation. However, this requires predictive models that are generalized to handle subject variability as seen clinically. Here, we highlight methods to improve test accuracy of an image classifier model for shrapnel identification using tissue phantom image sets. Using a previously developed image classifier neural network—termed ShrapML—blind test accuracy was less than 70% and was variable depending on the training/test data setup, as determined by a leave one subject out (LOSO) holdout methodology. Introduction of affine transformations for image augmentation or MixUp methodologies to generate additional training sets improved model performance and overall accuracy improved to 75%. Further improvements were made by aggregating predictions across five LOSO holdouts. This was done by bagging confidences or predictions from all LOSOs or the top-3 LOSO confidence models for each image prediction. Top-3 LOSO confidence bagging performed best, with test accuracy improved to greater than 85% accuracy for two different blind tissue phantoms. This was confirmed by gradient-weighted class activation mapping to highlight that the image classifier was tracking shrapnel in the image sets. Overall, data augmentation and ensemble prediction approaches were suitable for creating more generalized predictive models for ultrasound image analysis, a critical step for real-time diagnostic deployment.

60 APPLIED LIFE SCIENCES↗
DOE OSTI2025

Improving microstructures segmentation via pretraining with synthetic data

Image analysis of material microstructures through microscopy is an integral capability in the field of materials science. The topological and chemical information obtained through microscopy allow us to draw vital connections between material microstructures, properties, and processing. While scanning electron microscopy (SEM) is able to yield a considerable wealth of information interpretable by the intuition of experts, there has been considerable interest in using machine learning, convolutional neural networks (CNNs) in particular, for such image analysis task. Training CNNs for an image analysis task requires a large annotated dataset. However, in many materials science applications, obtaining a large annotated dataset is cost and labor intensive. In this work, we study the use of synthetic data to enlarge the available annotated experimental data of uranium oxide. We utilize a modified Potts model to simulate uranium oxide particles with morphologies similar to those observed experimentally. We then leverage an image-to-image translation model to synthesize the simulated particles as if they are acquired with SEM. Through this process, we obtain pairs of particle images and their corresponding SEM representations, which corresponds to pairs of annotations and images. Unlike previous works, we leverage synthetic data for pretraining a CNN model prior, and finetune that model further with experimental data. We experimentally demonstrate that using synthetic data as incremental learning process benefits the overall performance compared to training a model on combined synthetic and experimental data.

36 MATERIALS SCIENCE↗
DOE OSTI2025

Meeting Global Health Needs via Infectious Disease Forecasting: Development of a Reliable Data-Driven Framework

Infectious diseases (IDs) have a significant detrimental impact on global health. Timely and accurate ID forecasting can result in more informed implementation of control measures and prevention policies. To meet the operational decision-making needs of real-world circumstances, we aimed to build a standardized, reliable, and trustworthy ID forecasting pipeline and visualization dashboard that is generalizable across a wide range of modeling techniques, IDs, and global locations. We forecasted 6 diverse, zoonotic diseases (brucellosis, campylobacteriosis, Middle East respiratory syndrome, Q fever, tick-borne encephalitis, and tularemia) across 4 continents and 8 countries. We included a wide range of statistical, machine learning, and deep learning models (n=9) and trained them on a multitude of features (average n=2326) within the One Health landscape, including demography, landscape, climate, and socioeconomic factors. The pipeline and dashboard were created in consideration of crucial operational metrics—prediction accuracy, computational efficiency, spatiotemporal generalizability, uncertainty quantification, and interpretability—which are essential to strategic data-driven decisions. While no single best model was suitable for all disease, region, and country combinations, our ensemble technique selects the best-performing model for each given scenario to achieve the closest prediction. For new or emerging diseases in a region, the ensemble model can predict how the disease may behave in the new region using a pretrained model from a similar region with a history of that disease. The data visualization dashboard provides a clean interface of important analytical metrics, such as ID temporal patterns, forecasts, prediction uncertainties, and model feature importance across all geographic locations and disease combinations. As the need for real-time, operational ID forecasting capabilities increases, this standardized and automated platform for data collection, analysis, and reporting is a major step forward in enabling evidence-based public health decisions and policies for the prevention and mitigation of future ID outbreaks.

60 APPLIED LIFE SCIENCES↗
DOE OSTI2026

Artificial Intelligence for Event Reconstruction and Higgs Physics at CMS and Future Colliders

This dissertation charts a trajectory in which advances in artificial intelligence (AI) play a central role in pushing the high-energy physics frontier, complementing progress driven by higher collision energies and larger colliders. The discovery potential of the LHC and future colliders relies on accurate reconstruction of increasingly complex particle collision events. In the CMS experiment, this task is performed by the particle-flow (PF) algorithm. This dissertation presents the first implementation of a machine-learning-based particle-flow (MLPF) reconstruction in the CMS detector based on transformer architectures. In simulated top quark--antiquark pair (ttbar) events under LHC Run~3 (2023--2024) conditions, MLPF improves jet energy resolution by 10--20\% compared to standard PF for jets with transverse momentum between 30--100\GeV. Runtime performance is evaluated using simulated multijet events, with a median inference time of 20\unit{ms} per event on an NVIDIA L4 GPU, compa red to approximately 110\unit{ms} for standard PF. The MLPF algorithm is also validated on Run~3 collision data, representing the first data-validated ML-based reconstruction pipeline at any LHC experiment. We then extend MLPF toward future electron--positron colliders and introduce the first full-simulation cross-detector transfer learning workflow for PF reconstruction. The model is pre-trained on simulated events from the Compact Linear Collider detector (CLICdet) and fine-tuned on the CLIC-like detector (CLD) proposed for the Future Circular Collider (FCC). This approach achieves up to a 40\% improvement in jet energy resolution over rule-based reconstruction while reducing the required training dataset size by an order of magnitude, demonstrating the potential of AI to accelerate detector development and optimization. This dissertation also demonstrates how modern AI techniques enhance the sensitivity of LHC physics analyses. A CMS search for highly Lorentz-boosted Higgs bosons decaying to \textrm{W} boson pairs is presented, focusing on the single-lepton final state. A dedicated fine-tuning strategy for \ParT yields an approximately 70\% increase in expected sensitivity relative to the baseline model. The analysis uses proton--proton collision data at a center-of-mass energy of \ensuremath{\sqrt{s}=13\TeV} collected by CMS between 2016 and 2018, corresponding to an integrated luminosity of 138\ensuremath{\ \mathrm{fb}^{-1}}. The expected significance of the search is $1.86\sigma$, with an observed signal strength of $-0.19^{+0.48}_{-0.46}$. Finally, explainable AI techniques are applied to the MLPF and \ParticleNet algorithms using layerwise relevance propagation, showing that both models base their predictions on physically meaningful features consistent with our physics intuition. Together, these results demonstrate how advanced AI methods can enhance reconstruction, analysis sensitivity, and interpretability, shaping the next era of experimental parti cle physics.

Mokhtar, Farouk [UC, San Diego]↗
DOE OSTI2023

Deep learning integrates histopathology and proteogenomics at a pan-cancer level

We introduce a pioneering approach that integrates pathology imaging with transcriptomics and proteomics to identify predictive histology features associated with critical clinical outcomes in cancer. We utilize 2,755 H&E-stained histopathological slides from 657 patients across 6 cancer types from CPTAC. Our models effectively recapitulate distinctions readily made by human pathologists: tumor vs. normal (AUROC = 0.995) and tissue-of-origin (AUROC = 0.979). We further investigate predictive power on tasks not normally performed from H&E alone, including TP53 prediction and pathologic stage. Importantly, we describe predictive morphologies not previously utilized in a clinical setting. The incorporation of transcriptomics and proteomics identifies pathway-level signatures and cellular processes driving predictive histology features. Model generalizability and interpretability is confirmed using TCGA. We propose a classification system for these tasks, and suggest potential clinical applications for this integrated human and machine learning approach. A publicly available web-based platform implements these models.

60 APPLIED LIFE SCIENCES↗
NASA NTRSDate not supplied

A Science-Focused Artificial Intelligence (AI) Responding in Real-Time to New Information: Capability Demonstration for Ocean World Missions

Introduction: Artificial intelligence (AI) has long been considered a potential mechanism to explore increasingly challenging environments, including those with extreme temperatures and pressures, limited communication capabilities, or those with demanding terrain. We posit that missions in extreme environments could deploy an onboard AI focused on science observations and goals in order to augment a traditional concept(s) of operations (ConOps). An onboard AI capability could perform functions such as data analysis in order to make high-level decisions, including prioritized data transmission for analysis by ground-based teams or autonomously-guided follow-on analyses that maximize science return. Such a capability would empower missions to respond to scientific data of interest in real-time; a mission could make observations and perform a preliminary analysis to alert ground-based scientists to an observation of interest, enabling an informed, rapid response from Earth-based teams. Enceladus Case Study for Onboard AI: We are developing an onboard AI capability for real-time telemetry response that formulates and carries-out informed decisions in service to established mission goals, enabling increased science return of a mission. We focus our AI development for use on a constellation of SmallSats orbiting Enceladus. Our Enceladus case study tests autonomous decision-making capabilities in scenarios with complex orbital dynamics, plume ejecta, extreme cold environments, power restrictions, and a requirement to maximize science return for a potential positive detection of life, while critically evaluating the potential for false positives. Telemetry includes simulated scientific data, spacecraft onboard operational data (e.g., position, velocity, and rotation), and engineering hardware performance data. Enceladus SmallSat Constellation. Our constellation includes eight SmallSat spacecraft in an 8:35 resonant orbit-based formation, leveraging Saturn’s gravitational forces to maintain stable orbits with global coverage around Enceladus. To our knowledge, we simulate the first stable configuration of multiple spacecraft in closed orbits around Enceladus, using a full ephemeris force model (Russell and Lara, 2009). Each spacecraft’s orbit will precess, causing an eastward ground track shift (from an orbiter’s perspective) of each spacecraft for each orbit. However, all spacecraft return to their original positions relative to Enceladus after eight Enceladus revolutions around Saturn. We model communication pathways between SmallSats to understand how information would need to be transmitted across the constellation to enable AI-driven decision-making and resource allocation across the fleet. Capability Demonstration. Our simulated capability demonstration inputs position, velocity, and rotation telemetry from our Enceladus-focused constellation simulations, and mass spectrometry data collected from abiotic and biotic laboratory-analog ocean world experiments (Theiling et al., 2018; Theiling, 2021; Da Poian et al., 2023). Data from these experiments are used to simulate MS measurements and different scenarios of science observations for onboard analysis performed on each of the eight spacecraft. For these demonstrations, we integrate 24 machine learning (ML) algorithms into an onboard intelligence as a ‘knowledge base’, including algorithms evaluating data quality and those predicting (with % confidence) gas composition, ocean aqueous chemistry, and whether the sample was influenced by microbial life. The onboard AI capability is designed to use the knowledge base to come to a consensus-based decision in the interpretation of the observed data in order to request additional action outside of a pre-defined ConOps. Requested actions could include e.g., prioritized downlink to Earth (for analysis by ground-based teams) or follow-on analyses performed across the constellation. The spacecraft’s intelligent onboard planner must then determine whether sufficient resources (e.g., time, power, etc.) are available and weigh the request with mission priorities. In our simulation, the constellation is able to identify potential biosignatures using onboard ML algorithms, evaluate the confidence of that prediction, and perform follow-on analyses across the fleet to confirm the detection, in order to best prepare a transmission of these data to Earth-based teams.

astrobiology↗
DOE OSTI2021

Analysis and Integration of the Hydraulic Fracturing Test Site-2 (HFTS-2) Comprehensive Dataset

Hydraulic Fracturing Test Site-2 (HFTS-2) is a field-based research experiment performed in the Permian (Delaware) Basin. The unique aspect of this program was the acquisition of a unique, comprehensive, diagnostic dataset. Additionally, shorter parent wells drilled three years before the child wells offered clear distinction between the stages influenced by parent-child effects and the stages without any effects. The goal of this study was to analyze and integrate this comprehensive diagnostic dataset to understand the areal and vertical extent of hydraulic fractures (HF). The paper also provides insights on the effects of parent wells’ depletion on child well HF geometry based on various monitoring methods and subsurface models. Areal and vertical coverage for all HFTS-2 wells during stimulation and depletion was estimated based on analysis and interpretation of diagnostics and advanced modeling results. HFTS-2 diagnostics included microseismic (MS), pre- and post-stimulation logs and cores, bottomhole gauges, and fiber optic (FO) data. The diagnostics results (MS, FO, image logs) were integrated and used to calibrate subsurface models. Additional field tests were designed and implemented for depletion monitoring. The tailored program for monitoring depletion included vertical and slant well pressures, interference testing, and a vertical strain depletion trial. Areal Coverage: Conventional MS (and FO MS) were used to compute HF dimensions, which were compared with diagnostics (FO strain, gauge, image logs) observations and calibrated subsurface models. A post-production interference test did not show offset well communication. Vertical Coverage: Vertical coverage during stimulation was monitored using a vertical monitoring well. The stronger mechanical strain signals showed good correlation with MS event intensities, geomechanical properties, and gauge inferences. Vertical depletion was estimated based on vertical/slant well gauges and strain depletion tests. Parent-Child Effects: Diagnostics and calibrated subsurface models show asymmetry in child well fracture geometries for stages that overlap parent wells. Child well image logs serve as a good indicator for parent Downloaded from http://onepetro.org/URTECONF/proceedings-pdf/21URTC/2-21URTC/D021S031R004/2477423/urtec-2021-5241-ms.pdf/1 by Carol Worster on 28 February 2022 URTeC 5241 well HF tracking. Child well MS events had an eastward bias, in line with pre-stimulation image logs, and was confirmed by parent well frac hits. Novel/Additive Information: The dataset presents a unique, over-constrained problem space to compare independent techniques to arrive at HF metrics (i.e., stimulation height and/or half-length), unlike a single source dataset, in which calibration is done using available data to guide predictions. Here, the asymmetry in HF geometry seen in the stages influenced by parent-child effects offers unique insights into well spacing and landing, which are key capital decisions the unconventional resources industry is seeking to optimize.

58 GEOSCIENCES↗
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 OSTI2025

Consistent performance of large language models in rare disease diagnosis across ten languages and 4917 cases

Background Large language models (LLMs) are increasingly used medicine for diverse applications including differential diagnostic support. The training data used to create LLMs such as the Generative Pretrained Transformer (GPT) predominantly consist of English-language texts, but LLMs could be used across the globe to support diagnostics if language barriers could be overcome. Initial pilot studies on the utility of LLMs for differential diagnosis in languages other than English have shown promise, but a large-scale assessment on the relative performance of these models in a variety of European and non-European languages on a comprehensive corpus of challenging rare-disease cases is lacking. Methods We created 4917 clinical vignettes using structured data captured with Human Phenotype Ontology (HPO) terms with the Global Alliance for Genomics and Health (GA4GH) Phenopacket Schema. These clinical vignettes span a total of 360 distinct genetic diseases with 2525 associated phenotypic features. We used translations of the Human Phenotype Ontology together with language-specific templates to generate prompts in English, Chinese, Czech, Dutch, French, German, Italian, Japanese, Spanish, and Turkish. We applied GPT-4o, version gpt-4o-2024-08-06, and the medically fine-tuned Meditron3-70B to the task of delivering a ranked differential diagnosis using a zero-shot prompt. An ontology-based approach with the Mondo disease ontology was used to map synonyms and to map disease subtypes to clinical diagnoses in order to automate evaluation of LLM responses. Findings For English, GPT-4o placed the correct diagnosis at the first rank 19.9% and within the top-3 ranks 27.0% of the time. In comparison, for the nine non-English languages tested here the correct diagnosis was placed at rank 1 between 16.9% and 20.6%, within top-3 between 25.4% and 28.6% of cases. The Meditron3 model placed the correct diagnosis within the first 3 ranks for 20.9% of cases in English and between 19.9% and 24.0% for the other nine languages. Interpretation The differential diagnostic performance of LLMs across a comprehensive corpus of rare-disease cases was largely consistent across the ten languages tested. This suggests that the utility of LLMs in clinical settings may extend to non-English clinical settings.

Artificial intelligence↗
DOE OSTI2024

An Investigation of LES Wall Modeling for Rayleigh–Bénard Convection via Interpretable and Physics-Aware Feedforward Neural Networks with DNS

Abstract The traditional approach of using the Monin–Obukhov similarity theory (MOST) to model near-surface processes in large-eddy simulations (LESs) can lead to significant errors in natural convection. In this study, we propose an alternative approach based on feedforward neural networks (FNNs) trained on output from direct numerical simulation (DNS). To evaluate the performance, we conduct both a priori and a posteriori tests. In the a priori (offline) tests, we compare the statistics of the surface shear stress and heat flux, computed from filtered DNS input variables, to the stress and flux obtained from the filtered DNS. Additionally, we investigate the importance of various input features using the Shapley additive explanations value and the conditional average of the filter grid cells. In the a posteriori (online) tests, we implement the trained models in the System for Atmospheric Modeling (SAM) LES and compare the LES-generated surface shear stress and heat flux with those in the DNS. Our findings reveal that vertical velocity, a traditionally overlooked flow quantity, is one of the most important input features for determining the wall fluxes. Increasing the number of input features improves the a priori test results but does not always improve the model performance in the a posteriori tests because of the differences in input variables between the LES and DNS. Last, we show that physics-aware FNN models trained with logarithmic and scaled parameters can well extrapolate to more intense convection scenarios than in the training dataset, whereas those trained with primitive flow quantities cannot. Significance Statement The traditional near-surface turbulence model, based on a shear-dominated boundary layer flow, does not represent near-surface turbulence in natural convection. Using a feedforward neural network (FNN), we can construct a more accurate model that better represents the near-surface turbulence in various flows and reveals previously overlooked controlling factors and process interactions. Our study shows that the FNN-generated models outperform the traditional model and highlight the importance of the near-surface vertical velocity. Furthermore, the physics-aware FNN models exhibit the potential to extrapolate to convective flows of various intensities beyond the range of the training dataset, suggesting their broader applicability for more accurate modeling of near-surface turbulence.

54 ENVIRONMENTAL SCIENCES↗
DOE OSTI2024

GeoThermalCloud: Cloud Fusion of Big Data and Multi-Physics Models using Machine Learning for Discovery, Exploration, and Development of Hidden Geothermal Resources

The primary goals of this project are exploring hidden geothermal resources in the U.S.A. and designing profitable enhanced geothermal systems (EGS). Many processes and parameters control geothermal exploration and energy production from geothermal fields. Diverse datasets (e.g., geology, geochemistry, geophysics, satellite, airborne geophysics) are available to help characterize subsurface geothermal conditions. Sparse and multi-scale characteristics of these datasets prohibit properly leveraging these datasets for geothermal exploration and profitable EGS design. Recent advancements in machine learning (ML) promise to resolve these issues. The tremendous challenges and risks of geothermal exploration and production bring the demand for novel ML methods and tools that can (1) analyze large field datasets, (2) assimilate model simulations (large inputs and outputs), (3) process sparse datasets, (4) perform transfer learning (between sites with different exploratory levels), (5) extract hidden geothermal signatures in the field and simulation data, (6) label geothermal resources and processes, (7) identify high-value data acquisition targets, and (8) guide geothermal exploration and production by selecting optimal exploration, production, and drilling strategies. To address these necessities, ML-based geothermal resources exploration and enhanced geothermal systems (EGS) design tools have been developed. The exploration tool is called GeoThermalCloud and EGS design tool is called GeoDT-ML. GeoThermalCloud (https://github.com/SmartTensors/GeoThermalCloud.jl) utilizes a LANL unsupervised ML platform called SmartTensors (https://tensors.lanl.gov/) to automate data analyses and interpretations by extracting hidden signatures to identify geothermal prospects. Also, it enables the identification of critical measurements needed to identify geothermal resource signatures. Alternatively, GeoDT-ML (https://github.com/SmartTensors/GeoThermalCloud.jl/tree/master/EGS) is an ML-based alternative to GeoDT (https://github.com/GeoDesignTool/GeoDT.git), a fast, simplified multi-physics solver to evaluate EGS project designs in uncertain geologic systems. GeoDT-ML leverages recent advances in deep learning and high-performance computing. It is a faster and simpler version of GeoDT. To make this project a success, we used capabilities of LANL, PNNL, Google, Stanford, and Julia Computing. We analyzed eight datasets of the U.S.A. using GeothermalCloud and demonstrated potential highly prospective geothermal resources and identified key factors defining highly prospective sites. The first data set includes 44 locations in southwest New Mexico and 18 geological, hydrogeological, geophysical, geothermal, geochemical attributes. We defined low- and medium-temperature hydrothermal systems and discovered a new highly prospective site. The second data set analyzed 18 shallow water chemistry attributes at 14,342 locations in the Great Basin. It demarcated modestly, moderately, and highly prospective sites including key attributes for each type of prospectivity. The third data set analyzed Utah FORGE data including satellite (InSAR), geophysical (gravity, seismic), geochemical, and geothermal attributes. Here, we performed prospectivity analysis to identify future drilling locations using geological, geochemical, and geophysical attributes. Maps of temperature at depth and heat flow are constructed based on the available data. Prospectivity maps were generated, and drilling locations were proposed for future geothermal field exploration. The fourth data set analyzed 21 attributes at 120 locations in Tularosa Basin, New Mexico; data comes from past play fairway analyses in this region. ML analyses identified geothermal signatures associated with modestly, moderately, and highly hydrothermal systems. We also defined dominant attributes and spatial distribution of the geothermal signatures. The fifth, sixth, seventh, and eighth datasets include Tohatchi Springs, New Mexico, Hawaii, Brady site, Nevada, and EGS Collab, respectively. Moreover, we coupled GeothermalCloud and magnetotellurics data to pinpoint drilling locations for developing geothermal projects in the Tularosa Basin, New Mexico. GeothermalCloud found potential prospective locations for geothermal resources near White Sands Missile Range and McGregor Range at Fort Bliss. Magnetotellurics data determined the potential depth (~1800m) of geothermal prospects at McGregor Range based on apparent resistivity structures/layers in the subsurface. The McGregor Range consists of three resistivity layers and two resistivity structures. Magnetotellurics data also helps identify that the western portion of the McGregor Range has thick and low-resistivity earth materials. The low resistivity to the west is most likely for a fault system. Assuming temperature is consistent with a geothermal reservoir, the west-central part of the McGregor Range has the highest geothermal potential because of the increase in porosity and associated permeability attributed to the interpreted fault system. Also, we devised a coupling strategy between a process model and GeothermalCloud to characterize hydrogeological conditions and geothermal conditions, respectively. The process model characterizes hydrogeological and geothermal conditions on highly prospective geothermal sites provided by GeothermalCloud. We developed a physics-informed neural network (PINN) version of the Burns equation that can be easily coupled with GeothermalCloud. Furthermore, we performed an optimal design decision maximizing the economic value of an EGS power plant. This study optimized the range of well spacing between injection and production wells maximizing net present value in dollars (NPV). For this task, we used the GeoDT to simulate the Utah FORGE EGS development cycle from the initial well design to the end of production. Next, we accomplished another crucial task, which is predicting permeability of geothermal reservoirs. Predicting permeability of geothermal reservoirs is a non-trivial task because of huge computational runtime of simulation and lack of measurements. To avoid these limitations, we used easy-to-measure chemical concentrations in the subsurface as measurement data and convolutional neural network based ML model of a high-fidelity model. Next, we predicted permeability using Markov chain Monte Carlo simulation. We found that Markov chain Monte Carlo simulation predicts permeability with a high certainty if the prediction zone in the simulation area has chemical concentration data. Finally, we analyzed the DOE funded INGENIOUS and GeoDAWN projects data. For discovering hidden geothermal systems in the Great Basin, the INGENIOUS project accumulated old data, collected new data, and released them in 2022. The dataset includes a total of 24 geological, geophysical, and geochemical attributes. Data resolution and scale significantly vary prohibiting an appropriate usage. To avoid such limitations, we brought all data in the same resolution and scale by applying the inverse distance weighting interpolation technique for predicting data in unsampled locations. Subsequently, we analyzed LiDAR data of the GeoDAWN project. We received data in tiles format. The DOE’s overarching goal is to use ML on LiDAR data for finding favorable geological structures (e.g., step up faults in Brady, Nevada). To serve the purpose, we need to label favorable geologic structures that correspond to LiDAR data. We wrote an algorithm to label the LiDAR data with the favorable geologic structures.

15 GEOTHERMAL ENERGY↗