Search NASA⌕ Search

SEARCH · Search NASA

Results for “automated image analysis”

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

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

At least 397 records · Page 22
All sourcesNASA NTRSDOE OSTI
NASA NTRSDate not supplied

SatCORPS Global Cloud Composite (GCC): the Design and Delivery of A High Quality, High Resolution, Global Cloud Product Available in Near-Real Time

The NASA Satellite ClOud and Radiation Property retrieval System (SatCORPS) supports the development of an analysis ready and cloud-optimized data transformation pipeline and geospatial service enablement of a global cloud composite (GCC) product derived from global geostationary satellite imagery. This geospatial service will be available at high temporal and spatial resolution via the SatCORPS web mapping application for visualization and analysis as well as direct ingestion to common geospatial software and custom programming. The resulting global cloud composite products from the processing pipeline can then be geospatially-service enabled as ArcGIS Image Services and Open Geospatial Consortium (OGC) Web Mapping/Coverage Services for visualization and analysis via a web mapping application and common geospatial software. Near real time global observations are created through the composition of five geostationary satellites that provides modelling and forecasting communities with the capability to provide high quality and timely information to start the projection process. The Global Cloud Composite product combines information from geostationary satellites, GOES-16, GOES-17, Himawari-8, Meteosat-11 and Meteosat-9 to create a single global composite netcdf file and images using the different products within the netcdf file. The SatCORPS team, though our Global Cloud Composite (GCC) product and web-based visualization tools including Geographical Information System (GIS) services provide near real time global cloud product information to both automated processes and traditional web users that is timely and high quality derived from geostationary satellites. The Global Cloud Composite product takes advantage of the scalable processing resources provided by the AWS batch service to provide new composites every thirty minutes. Because information from each of the low earth orbiting satellites is available on schedules tuned to the specific satellite, the processing algorithm temporally composites the final dataset as each satellite’s information becomes available. The SatCORPS team has leveraged our experience using Amazon Web Services (AWS) to build a low latency high availability tool that allows end users both human and automated to acquire high quality and high-resolution Geostationary Earth Orbiting (GEO) information at zero cost to the end user. This presentation will describe how we architected and implemented the service as well as lessons learned based on our experiences both developing and operating the system. The lessons learned include how we integrated multiple services including Amazon Batch, Amazon S3 and Amazon Lambda service to create a low cost but high-performance processing system that is capable of identifying and processing the most appropriate satellite overpass information into global cloud composites. We will also describe our web-based tools including our Geographic Information System that can be used for visualization and analysis. The products from the processing can be geospatially-service enabled as ArcGIS Image Services and Open Geospatial Consortium (OGC) Web Mapping/Coverage Services for visualization and analysis via a web mapping application and common geospatial software. The SatCORPS Global Composite Cloud product provides sophisticated global composited cloud research products with very low latency that we see that as filling a rapidly growing need in the research and modelling community with no up-front nor ongoing costs associated with downloading or using the information.

AWS AMCE SMCE GCC SATCORPS GLOBAL CLOUD COMPOSITE ↗
DOE OSTI2022

A Combined Computer Vision and Deep Learning Approach for Rapid Drone-Based Optical Characterization of Parabolic Troughs

Optical accuracy is a primary driver of parabolic trough concentrating solar power (CSP) plant performance, but can be damaged by wind loads, gravity, error during installation, and regular plant operation. Collecting and analyzing optical measurements over an entire operating parabolic trough plant is difficult, given the large scale of typical installations. Distant Observer, a software tool developed at the National Renewable Energy Laboratory, uses images of the absorber tube reflected in the collector mirror to measure both surface slope in the parabolic mirror and offset of the absorber tube from the ideal focal point. This technology has been adapted for fast data collection using low-cost commercial drones, but until recently still required substantial human labor to process large amounts of data. A new method leveraging advanced deep learning and computer vision tools can drastically reduce the time required to process images. This new method addresses the primary analysis bottleneck, identifying featureless, reflective mirror corner points to a high degree of accuracy. Recent work has shown promising results using computer vision methods. The combined deep learning and computer vision approach presented here proved highly effective and has the potential to further automate data collection and analysis, making the tool more robust. The method presented in this paper automatically identified 74.3% of mirror corners within 2 pixels of their manually marked counterparts and 91.9% within 3 pixels. This level of accuracy is sufficient for practical Distant Observer analysis within a target uncertainty. A commercial drone collected video of over 100 parabolic trough modules at an operating CSP plant to demonstrate the deep learning and computer vision method's usefulness in processing large amounts of data. These troughs were successfully analyzed using Distant Observer, paired with the new deep learning and computer vision algorithm, and can provide plant operators and trough designers with valuable insight about plant performance, operating strategies, and plant-wide optical error trends.

computer vision↗
NASA NTRS2012

VESGEN Software for Mapping and Quantification of Vascular Regulators

VESsel GENeration (VESGEN) Analysis is an automated software that maps and quantifies effects of vascular regulators on vascular morphology by analyzing important vessel parameters. Quantification parameters include vessel diameter, length, branch points, density, and fractal dimension. For vascular trees, measurements are reported as dependent functions of vessel branching generation. VESGEN maps and quantifies vascular morphological events according to fractal-based vascular branching generation. It also relies on careful imaging of branching and networked vascular form. It was developed as a plug-in for ImageJ (National Institutes of Health, USA). VESGEN uses image-processing concepts of 8-neighbor pixel connectivity, skeleton, and distance map to analyze 2D, black-and-white (binary) images of vascular trees, networks, and tree-network composites. VESGEN maps typically 5 to 12 (or more) generations of vascular branching, starting from a single parent vessel. These generations are tracked and measured for critical vascular parameters that include vessel diameter, length, density and number, and tortuosity per branching generation. The effects of vascular therapeutics and regulators on vascular morphology and branching tested in human clinical or laboratory animal experimental studies are quantified by comparing vascular parameters with control groups. VESGEN provides a user interface to both guide and allow control over the users vascular analysis process. An option is provided to select a morphological tissue type of vascular trees, network or tree-network composites, which determines the general collections of algorithms, intermediate images, and output images and measurements that will be produced.

Parsons-Wingerter, Patricia A.↗
NASA NTRS1974

Automated analysis of video data

A software/hardware system designed and developed to perform automated video imagery analysis is described. Major elements included in the analysis system are: (1) sampling oscilloscope; (2) programmable trigger unit; (3) central computer; and (4) software processing library. The function of the scope and trigger unit, in conjunction with the control computer, is to digitize the selected portion of the video image and store as amplitude and time data in computer memory. Evaluation of the video data, analogus to interpretation of a visual image, is accomplished by the processing software. It is indicated that this technique of video image analysis is applicable to a wide variety of nonaerospace applications involving video data and image analysis.

Hafner, D. J.↗
NASA NTRS2010

Automated Detection of Oscillating Regions in the Solar Atmosphere

Recently observed oscillations in the solar atmosphere have been interpreted and modeled as magnetohydrodynamic wave modes. This has allowed for the estimation of parameters that are otherwise hard to derive, such as the coronal magnetic-field strength. This work crucially relies on the initial detection of the oscillations, which is commonly done manually. The volume of Solar Dynamics Observatory (SDO) data will make manual detection inefficient for detecting all of the oscillating regions. An algorithm is presented that automates the detection of areas of the solar atmosphere that support spatially extended oscillations. The algorithm identifies areas in the solar atmosphere whose oscillation content is described by a single, dominant oscillation within a user-defined frequency range. The method is based on Bayesian spectral analysis of time series and image filtering. A Bayesian approach sidesteps the need for an a-priori noise estimate to calculate rejection criteria for the observed signal, and it also provides estimates of oscillation frequency, amplitude, and noise, and the error in all of these quantities, in a self-consistent way. The algorithm also introduces the notion of quality measures to those regions for which a positive detection is claimed, allowing for simple post-detection discrimination by the user. The algorithm is demonstrated on two Transition Region and Coronal Explorer (TRACE) datasets, and comments regarding its suitability for oscillation detection in SDO are made.

Ireland, J.↗
DOE OSTI2025

Robust Spectral Anomaly Detection in EELS Spectral Images via 3D Convolutional Variational Autoencoders

Abstract A 3D Convolutional Variational Autoencoder (3D‐CVAE) is introduced for automated anomaly detection in electron energy‐loss spectroscopy spectrum imaging (EELS‐SI) data. This approach leverages the full 3D structure of EELS‐SI data to detect subtle spectral anomalies while preserving both spatial and spectral correlations across the datacube. By employing cross‐entropy loss and training on bulk spectra, the model learns to reconstruct bulk features characteristic of the defect‐free material. In exploring methods for anomaly detection, both the 3D‐CVAE approach and principal component analysis (PCA) are evaluated, testing their performance using FeL‐edge ΔEpeak shifts designed to simulate material defects. These results show that 3D‐CVAE achieves superior anomaly detection and maintains consistent performance across various shift magnitudes. The method demonstrates clear bimodal separation between bulk and anomalous spectra, enabling reliable classification. Further analysis verifies that lower‐dimensional representations are robust to anomalies in the data. While performance advantages over PCA diminish with decreasing anomaly concentration, our method maintains high reconstruction quality even in challenging, noise‐dominated spectral regions. This approach provides a robust framework for unsupervised automated detection of spectral anomalies in EELS‐SI data, particularly valuable for analyzing complex material systems.

Chemistry↗
NASA NTRS1995

Techniques for Improving the Performance of Future EVA Maneuvering Systems

The Simplified Aid for EVA Rescue (SAFER) is a small propulsive backpack that was developed as an in-house effort at Johnson Space Center; it is a lightweight system which attaches to the underside of the Primary Life Support Subsystem (PLSS) backpack of the Extravehicular Mobility Unit (EMU). SAFER provides full six-axis control, as well as Automatic Attitude Hold (AAH), by means of a set of cold-gas nitrogen thrusters and a rate sensor-based control system. For compactness, a single hand controller is used, together with mode switching, to command all six axes. SAFER was successfully test-flown on the STS-64 mission in September 1994 as a Development Test Objective (DTO); development of an operational version is now proceeding. This version will be available for EVA self-rescue on the International Space Station and Mir, starting with the STS-86/Mir-7 mission in September 1997. The DTO SAFER was heavily instrumented, and produced in-flight data that was stored in a 12 MB computer memory on-board. This has allowed post-flight analysis to yield good estimates for the actual mass properties (moments and products of inertia and center of mass location) encountered on-orbit. By contrast, Manned Maneuvering Unit (MMU) post-flight results were generated mainly from analysis of video images, and so were not very accurate. The main goal of the research reported here was to use the detailed SAFER on-orbit mass properties data to optimize the design of future EVA maneuvering systems, with the aim being to improve flying qualities and/or reduce propellant consumption. The Automation, Robotics and Simulation Division Virtual Reality (VR) Laboratory proved to be a valuable research tool for such studies. A second objective of the grant was to generate an accurate dynamics model in support of the reflight of the DTO SAFER on STS-76/Mir-3. One complicating factor was the fact that a hand controller stowage box was added to the underside of SAFER on this flight; the position of this box was such that two of the SAFER jets plume it. A second complication was that the EVA astronaut will sometimes be transporting a massive experiment package. This will not only alter the overall mass properties significantly, but can itself also be plumed.

Williams, Trevor W.↗
NASA NTRS2013

Metric Learning to Enhance Hyperspectral Image Segmentation

Unsupervised hyperspectral image segmentation can reveal spatial trends that show the physical structure of the scene to an analyst. They highlight borders and reveal areas of homogeneity and change. Segmentations are independently helpful for object recognition, and assist with automated production of symbolic maps. Additionally, a good segmentation can dramatically reduce the number of effective spectra in an image, enabling analyses that would otherwise be computationally prohibitive. Specifically, using an over-segmentation of the image instead of individual pixels can reduce noise and potentially improve the results of statistical post-analysis. In this innovation, a metric learning approach is presented to improve the performance of unsupervised hyperspectral image segmentation. The prototype demonstrations attempt a superpixel segmentation in which the image is conservatively over-segmented; that is, the single surface features may be split into multiple segments, but each individual segment, or superpixel, is ensured to have homogenous mineralogy.

Thompson, David R.↗
NASA NTRSDate not supplied

Machine Learning for Extravehicular Mobility Unit (EMU) Glove Inspections

The Extravehicular Mobility Unit (EMU) Glove Machine Learning Inspection project utilizes machine learning to expedite the inspection, analysis, and recommendation for continued use of space suit gloves post spacewalks. Today, ISS glove photos are individually reviewed by a team of experts to determine the conditions of space suit gloves. For this project the Microsoft Azure platform is used to perform Automated Machine Learning (AutoML) to detect issues with tagged images from previous Extravehicular Activities (EVA’s) to build a predictive model. The model analyzes a test image and deems the glove GO or NO-GO for additional EVA’s. The goal for this ML project is to decrease the time spent reviewing images by ground personnel and crewmembers in high frequency EVA locations such as the Moon and Mars. For destinations such as the Moon and Mars the goal is to give crew autonomy in determining glove conditions with limited support from Earth. This paper will outline the results to date and future work needed to expand the capability for in-situ recommendations.

EVA↗
DOE OSTI2023

Autonomous x-ray scattering

Autonomous experimentation (AE) is an emerging paradigm that seeks to automate the entire workflow of an experiment, including—crucially—the decision-making step. Beyond mere automation and efficiency, AE aims to liberate scientists to tackle more challenging and complex problems. We describe our recent progress in the application of this concept at synchrotron x-ray scattering beamlines. We automate the measurement instrument, data analysis, and decision-making, and couple them into an autonomous loop. We exploit Gaussian process modeling to compute a surrogate model and associated uncertainty for the experimental problem, and define an objective function exploiting these. We provide example applications of AE to x-ray scattering, including imaging of samples, exploration of physical spaces through combinatorial methods, and coupling to in situ processing platforms These uses demonstrate how autonomous x-ray scattering can enhance efficiency, and discover new materials.

77 NANOSCIENCE AND NANOTECHNOLOGY↗
DOE OSTI2020

Deep Learning for Automated Identification of Eels in Sonar Data

Freshwater eels, such as the American eel (Anguilla rostrata) present numerous challenges related to safe downstream fish passage at hydroelectric facilities. One of those challenges is effective monitoring of their abundance, movements, and behavior to facilitate design and operation of eel protection and passage facilities. A previous EPRI study documented the ability of human analysts to reliably identify American eels in data obtained with a 1100/1800 kHz, multibeam sonar. This report describes a project to develop deep learning (a subset of artificial intelligence) tools to automate the time-consuming, subjective process of eel identification in multibeam sonar data. The project exploited new data collected in the laboratory and the existing data from the prior EPRI field study to develop and test deep learning and other data analytic tools, including wavelet filtering, differencing for static object removal, and convolutional neural network analysis. The analysis of the laboratory data demonstrated feasibility of the approach, revealed object characteristics observed with the sonar that distinguish eels from similarly sized and shaped acoustic targets, and provided additional data for algorithm selection and training. Deep learning algorithms trained and tested on the laboratory data alone achieved accuracy rates of greater than 98% when classifying acoustic images of eels and similar-sized neutrally buoyant sticks. The algorithm trained and tested on the pre-existing field data alone, and yielded classification accuracy of 9.3% false positives and 13.3% false negatives when distinguishing between eels and sticks/PVC pipes based on video clips (i.e., multiple, consecutive images). This performance is comparable to the classification accuracy achieved by human analysts in the prior study. The deep learning algorithm trained on a combination of video clips obtained in the laboratory and the field and tested on video clips from the field, was able to distinguish eels from sticks and PVC pipes (a river debris analog) of similar size with 100% accuracy. Outreach to the hardware, software, and end-user communities early in the project helped to identify needs and specify the application space. Outreach to those communities at the end of the project communicated project results and opportunities for further development. The project achieved proof of concept for automated identification of eel in multibeam sonar data. Future work should focus on acquisition of additional data for more robust algorithm training and testing; modification of the software tools to accommodate multiple acoustic targets in the acoustic field at a given time; identification of additional object classes; incorporation of motion in the object identification and classification algorithms; operationalizing the software tools, including integration with other existing sonar data analysis tools; and partnering with hardware and software providers for distribution of the software tools with their commercial products.

13 HYDRO ENERGY↗
DOE OSTI2024

Introducing SpaceNet 9 - Cross-Modal Satellite Imagery Registration for Natural Disaster Responses

Computer vision algorithms are increasingly leveraged to accelerate geospatial analysis for disaster response and recovery. As the diversity of remote sensing imagery grows with optical, SAR, and other modalities, a perquisite for analytics is cross-modal image registration. There is a high potential to harness computer vision for this pre-processing requirement toward enabling downstream analytics such as heterogeneous change detection, automated feature extraction, and data fusion. Advancement in these areas has the potential to simplify data wrangling tasks and further accelerate disaster response timelines. The SpaceNet 9 challenge (launching in mid-2024) focuses on addressing the cross-modal image registration problem and demonstrating the utility of such modules on earthquake impacted scenarios. This paper describes the motivation for the SpaceNet 9 and provides a first overview of the dataset, the baseline algorithm, and implications for seeking cross-modal image registration in Earth observation. Code is available at https://github.com/SpaceNetChallenge/SpaceNet9.

Hansch, Ronny↗
NASA NTRS2020

Overview and Initial Results of SAND-E: Semi-Autonomous Navigation for Detrital Environments

Unmanned aerial systems (UAS) and automated terrain analysis for science and navigation are new technologies for planetary exploration. The Mars Helicopter will fly with the Mars2020 rover, the Dragonfly quadcopter will explore Titan, and Soil Properties and Object Classification (SPOC) software will be used for path planning and navigation on the Mars2020 rover. Using an Argo J5 rover instrumented with stereo cameras and Autonomous Soil Assessment System (ASAS) software, and an off the shelf quadcopter, SAND-E tested the use of automated terrain analysis and UAS data for science operations in a Mars-analog environment in Iceland during July of 2019. Scientifically, we sought to determine changes in the physical and chemical properties of sediments along a glacial-fluvial-aeolian transport pathway. Operationally, we tested rover mission-like scenarios that included UAS images and classified terrain images. Here, we present the initial results for both the operations and science elements of the study. Site Selection: A goal of SAND-E is examine sorting and alteration of sediments in fluvial and aeolian environments in both mineral-dominated and glass-dominated basaltic settings. During the first year of the project we focused on a mineral-dominated environment. Selection of the location was based on prior publications that indicated our selected region had a greater abundance of crystalline sediments than other areas fluvial-aeolian settings in Iceland. Other criteria included the presence of both fluvial and aeolian landforms along a transport pathway such that the sediments in transport could be linked to their source rocks. We chose the Skjaldbreidauhraun glacial outwash plain, which sits at the base of Thórisjökull glacier. The site is 30 km north of Thingvellir National Park and ~2 hours from Reykjavik. The outwash plain is fed by two small catchments that drain from the base of the glacier and cut through hyaloclastite and shield volcano bedrock. The drainage progresses from steep alluvial fans near the glacier into a low-sloping fluvial braidplain that becomes confined by the Skjaldbreidur shield volcano and creates a shallow canyon cut into lava bedrock. The fluvial system was a typical braided alluvial environment composed pebble- and cobble-bedded longitudinal bars and sandy channel beds. The river remained active and fluctuated in response to diurnal runoff cycles near the glacier before disappearing into the sandy substrate downstream. The high concentration of suspended sediment in the river was evident by the cloudy water and the silt and clay-sized sediments that draped the channel beds after abandonment and created playas in the lowest sloping areas of the catchment. The entire fluvial system was affected by the winds generated by frontal systems and katabatic flows descending the glacier. This resulted in the formation of aeolian lag deposits and a wind-deflation plain where the fluvial system was not active. Wind ripples and drifts formed in abandoned fluvial channels from aeolian reworking of the sand-sized fluvial sediments. The silt- and clay-sized sediments found in fluvial channels, bar tops, and playas generated dust plumes during high wind events. Our operation sought to capture the variability in this system by sampling from the range of fluvial and aeolian features 6.3 km (proximal), 11.3 km (medial), and 14.4 km (distal) along the river from its origin at the base of glacier.

Ewing, R. C.↗
DOE OSTI2023

Crack detection in fuel cell electrodes using a spatial filtering technique for overcoming noisy backgrounds

Image processing is a powerful tool that allows for rapid and automated data parsing in settings that occupy large variable spaces and require large data sets. Feature detection on difficultly discerned backgrounds is a subset of image processing that facilitates the extraction of quantitative metrics from otherwise subjective data. Crack detection and quantification is an important capability in polymer electrolyte membrane fuel cell quality control, failure analysis, and optimization. This work presents a technique to perform crack detection and quantification which overcomes challenges faced by commonly used image segmentation techniques. We demonstrate the use of a geometrically filtered noise‐level detection technique to select a binary threshold value from which we then quantify how cracked a sample is. Furthermore, we demonstrate the accuracy of our technique using programmatically generated test images of known crack amounts and their performance on real‐world fuel cell catalyst layer samples.

30 DIRECT ENERGY CONVERSION↗
NASA NTRS1991

Data simulation for the Lightning Imaging Sensor (LIS)

This project aims to build a data analysis system that will utilize existing video tape scenes of lightning as viewed from space. The resultant data will be used for the design and development of the Lightning Imaging Sensor (LIS) software and algorithm analysis. The desire for statistically significant metrics implies that a large data set needs to be analyzed. Before 1990 the quality and quantity of video was insufficient to build a usable data set. At this point in time, there is usable data from missions STS-34, STS-32, STS-31, STS-41, STS-37, and STS-39. During the summer of 1990, a manual analysis system was developed to demonstrate that the video analysis is feasible and to identify techniques to deduce information that was not directly available. Because the closed circuit television system used on the space shuttle was intended for documentary TV, the current value of the camera focal length and pointing orientation, which are needed for photoanalysis, are not included in the system data. A large effort was needed to discover ancillary data sources as well as develop indirect methods to estimate the necessary parameters. Any data system coping with full motion video faces an enormous bottleneck produced by the large data production rate and the need to move and store the digitized images. The manual system bypassed the video digitizing bottleneck by using a genlock to superimpose pixel coordinates on full motion video. Because the data set had to be obtained point by point by a human operating a computer mouse, the data output rate was small. The loan and subsequent acquisition of a Abekas digital frame store with a real time digitizer moved the bottleneck from data acquisition to a problem of data transfer and storage. The semi-automated analysis procedure was developed using existing equipment and is described. A fully automated system is described in the hope that the components may come on the market at reasonable prices in the next few years.

Boeck, William L.↗
NASA NTRS2015

Analyses of Magnetic Resonance Imaging of Cerebrospinal Fluid Dynamics Pre and Post Short and Long-Duration Space Flights

Preliminary results are based on analyses of data from 17 crewmembers. The initial analysis compares pre to post-flight changes in total cerebral blood flow (CBF) and craniospinal CSF flow volume. Total CBF is obtained by summation of the mean flow rates through the 4 blood vessels supplying the brain (right and left internal carotid and vertebral arteries). Volumetric flow rates were obtained using an automated lumen segmentation technique shown to have 3-4-fold improved reproducibility and accuracy over manual lumen segmentation (6). Two cohorts, 5 short-duration and 8 long-duration crewmembers, who were scanned within 3 to 8 days post landing were included (4 short-duration crewmembers with MRI scans occurring beyond 10 days post flight were excluded). The VIIP Clinical Practice Guideline (CPG) classification is being used initially as a measure for VIIP syndrome severity. Median CPG scores of the short and long-duration cohorts were similar, 2. Mean preflight total CBF for the short and long-duration cohorts were similar, 863+/-144 and 747+/-119 mL/min, respectively. Percentage CBF changes for all short duration crewmembers were 11% or lower, within the range of normal physiological fluctuations in healthy individuals. In contrast, in 4 of the 8 long-duration crewmembers, the change in CBF exceeded the range of normal physiological fluctuation. In 3 of the 4 subjects an increase in CBF was measured. Large pre to post-flight changes in the craniospinal CSF flow volume were found in 6 of the 8 long-duration crewmembers. Box-Whisker plots of the CPG and the percent CBF and CSF flow changes for the two cohorts are shown in Figure 4. Examples of CSF flow waveforms for a short and two long-duration (CPG 0 and 3) are shown in Figure 5. Changes in CBF and CSF flow dynamics larger than normal physiological fluctuations were observed in the long-duration crewmembers. Changes in CSF flow were more pronounced than changes in CBF. Decreased CSF flow dynamics were observed in a subject with VIIP signs. Study limitations include a slightly longer landing-to-MRI scan period for the short-duration cohort and limited sensitivity of the subjective discrete ordinal CPG scale. This limitation can be overcome by using imaging based parametric measures of VIIP severity such as globe deformation measures.

Alperin, Noam↗
DOE OSTI2022

Unified Language Frontend for Physic-Informed AI/ML

Artificial intelligence and machine learning (AI/ML) are becoming important tools for scientific modeling and simulation as in several other fields such as image analysis and natural language processing. ML techniques can leverage the computing power available in modern systems and reduce the human effort needed to configure experiments, interpret and visualize results, draw conclusions from huge quantities of raw data, and build surrogates for physics based models. Domain scientists in fields like fluid dynamics, microelectronics and chemistry can automate many of their most difficult and repetitive tasks or improve the design times by use of the faster ML-surrogates. However, modern ML and traditional scientific highperformance computing (HPC) tend to use completely different software ecosystems. While ML frameworks like PyTorch and TensorFlow provide Python APIs, most HPC applications and libraries are written in C++. Direct interoperability between the two languages is possible but is tedious and error-prone. In this work, we show that a compiler-based approach can bridge the gap between ML frameworks and scientific software with less developer effort and better efficiency. We use the MLIR (multi-level intermediate representation) ecosystem to compile a pre-trained convolutional neural network (CNN) in PyTorch to freestanding C++ source code in the Kokkos programming model. Kokkos is a programming model widely used in HPC to write portable, shared-memory parallel code that can natively target a variety of CPU and GPU architectures. Our compiler-generated source code can be directly integrated into any Kokkosbased application with no dependencies on Python or cross-language interfaces.

71 CLASSICAL AND QUANTUM MECHANICS, GENERAL PHYSIC↗