Engineering Papers⌕ Search

SEARCH · Engineering Papers

Results for “supervised training”

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 325 records · Page 18

Weakly supervised anomaly detection with event-level variables

We introduce a new topology for weakly supervised anomaly detection searches, diobject plus X. In this topology, one looks for a resonance decaying to two standard model particles produced in association with other anomalous event activity (X). This additional activity is used for classification. We demonstrate how anomaly detection techniques which have been developed for dijet searches focusing on jet substructure anomalies can be applied to event-level anomaly detection in this topology. To robustly capture event-level features of multiparticle kinematics, we employ new physically motivated variables derived from the geometric structure of a collision’s phase space manifold. As a proof of concept, we explore the application of this approach to several benchmark signals in the di-𝜏 and di-𝜇 plus X final states. We demonstrate that our anomaly detection approach can reach discovery-level significances for signals that would be missed in a conventional bump-hunt approach.

72 PHYSICS OF ELEMENTARY PARTICLES AND FIELDS↗

MAD: Self-Supervised Masked Anomaly Detection Task for Multivariate Time Series

In this paper, we introduce Masked Anomaly Detection (MAD), a general self-supervised learning task for multivariate time series anomaly detection. With the increasing availability of sensor data from industrial systems, being able to detecting anomalies from streams of multivariate time series data is of significant importance. Given the scarcity of anomalies in real-world applications, the majority of literature has been focusing on modeling normality. The learned normal representations can empower anomaly detection as the model has learned to capture certain key underlying data regularities. A typical formulation is to learn a predictive model, i.e., use a window of time series data to predict future data values. In this paper, we propose an alternative self-supervised learning task. By randomly masking a portion of the inputs and training a model to estimate them using the remaining ones, MAD is an improvement over the traditional left-to-right next step prediction (NSP) task. Our experimental results demonstrate that MAD can achieve better anomaly detection rates over traditional NSP approaches when using exactly the same neural network (NN) base models, and can be modified to run as fast as NSP models during test time on the same hardware, thus making it an ideal upgrade for many existing NSP-based NN anomaly detection models.

97 MATHEMATICS AND COMPUTING↗

Machine Learning Toolkit for Extreme Scale

Support Vector Machines (SVM) is a popular machine learning technique, which has been applied to a wide range of domains such as science, finance, and social networks for supervised learning. MaTEx undertakes the challenge of designing a scalable parallel SVM training algorithm for large scale systems, which includes commodity multi-core machines, tightly connected supercomputers and cloud computing systems. Several techniques are proposed for improved speed and memory space usage including adaptive and aggressive elimination of samples for faster convergence , and sparse format representation of data samples. Several heuristics for earliest possible to lazy elimination of non-contributing samples are considered in MaTEx. In many cases, where an early sample elimination might result in a false positive, low overhead mechanisms for reconstruction of key data structures are proposed. The proposed algorithm and heuristics are implemented and evaluated on various publicly available datasets

Source record↗

Automated Threat Recognition For Aviation Security Applications

We have developed a framework for automated threat recognition (ATR) of explosive threat materials for both single-energy and dual-energy X-ray CT systems. Under this framework, two different types of ATR have been developed. The first type of ATR employs supervised machine learning with statistical characterization of target materials for threat identification training. The reliance only on statistical characterization information for threat training uniquely enables this style of ATR to adapt quickly to evolving threats. The second type of ATR employs deep learning through convolutional neural networks. Convolutional neural networks are attractive due to their human-like capacity for learning and strong ability to identify trends and patterns. Although this method is more powerful than the first type, it requires large amounts of training data and therefore is less agile. Each ATR performs threat characterization at the voxel neighborhood level. This approach avoids the use of threat shape as a detection criterion, which is prohibited by DHS and TSA guidelines.

97 MATHEMATICS AND COMPUTING↗

Real-Time, Interactive Echocardiography Over High-Speed Networks: Feasibility and Functional Requirements

Real-time, Interactive Echocardiography Over High Speed Networks: Feasibility and Functional Requirements is an experiment in advanced telemedicine being conducted jointly by the NASA Lewis Research Center, the NASA Ames Research Center, and the Cleveland Clinic Foundation. In this project, a patient undergoes an echocardiographic examination in Cleveland while being diagnosed remotely by a cardiologist in California viewing a real-time display of echocardiographic video images transmitted over the broadband NASA Research and Education Network (NREN). The remote cardiologist interactively guides the sonographer administering the procedure through a two-way voice link between the two sites. Echocardiography is a noninvasive medical technique that applies ultrasound imaging to the heart, providing a "motion picture" of the heart in action. Normally, echocardiographic examinations are performed by a sonographer and cardiologist who are located in the same medical facility as the patient. The goal of telemedicine is to allow medical specialists to examine patients located elsewhere, typically in remote or medically underserved geographic areas. For example, a small, rural clinic might have access to an echocardiograph machine but not a cardiologist. By connecting this clinic to a major metropolitan medical facility through a communications network, a minimally trained technician would be able to carry out the procedure under the supervision and guidance of a qualified cardiologist.

Bobinsky, Eric A.↗

Machine Learning Toolkit for Extreme Scale

Support Vector Machines (SVM) is a popular machine learning technique, which has been applied to a wide range of domains such as science, finance, and social networks for supervised learning. MaTEx undertakes the challenge of designing a scalable parallel SVM training algorithm for large scale systems, which includes commodity multi-core machines, tightly connected supercomputers and cloud computing systems. Several techniques are proposed for improved speed and memory space usage including adaptive and aggressive elimination of samples for faster convergence , and sparse format representation of data samples. Several heuristics for earliest possible to lazy elimination of non-contributing samples are considered in MaTEx. In many cases, where an early sample elimination might result in a false positive, low overhead mechanisms for reconstruction of key data structures are proposed. The proposed algorithm and heuristics are implemented and evaluated on various publicly available datasets

Source record↗

Extracting Material Property Measurements from Scientific Literature with Limited Annotations

Extracting material property data from scientific text is pivotal for advancing data-driven research in chemistry and materials science; however, the extensive annotation effort required to produce training data for named entity recognition (NER) models for this task often makes it a barrier to extracting specialized data sets. Here, in this work, we present a comparative study of the conventional, supervised NER methodology to alternative few-shot learning architectures and large language model (LLM)-based approaches that mitigate the need to label large training data sets. We find that the best-performing LLM (GPT-4o) not only excels in directly extracting relevant material properties based on limited examples but also enhances supervised learning through data augmentation. We supplement our findings with error and data quality assessments to provide a nuanced understanding of factors that impact property measurement extraction.

36 MATERIALS SCIENCE↗

Infinite neural network quantum states: entanglement and training dynamics

We study infinite limits of neural network quantum states (∞-NNQS), which exhibit representation power through ensemble statistics, and also tractable gradient descent dynamics. Ensemble averages of entanglement entropies are expressed in terms of neural network correlators, and architectures that exhibit volume-law entanglement are presented. The analytic calculations of entanglement entropy bound are tractable because the ensemble statistics are simplified in the Gaussian process limit. A general framework is developed for studying the gradient descent dynamics of neural network quantum states (NNQS), using a quantum state neural tangent kernel (QS-NTK). For ∞-NNQS the training dynamics is simplified, since the QS-NTK becomes deterministic and constant. An analytic solution is derived for quantum state supervised learning, which allows an ∞-NNQS to recover any target wavefunction. Numerical experiments on finite and infinite NNQS in the transverse field Ising model and Fermi Hubbard model demonstrate excellent agreement with theory. ∞-NNQS opens up new opportunities for studying entanglement and training dynamics in other physics applications, such as in finding ground states.

quantum state supervised learning↗

Machine-learning based approach to examine ecological processes influencing the diversity of riverine dissolved organic matter composition

Dissolved organic matter (DOM) assemblages in freshwater rivers are formed from mixtures of simple to complex compounds that are highly variable across time and space. These mixtures largely form due to the environmental heterogeneity of river networks and the contribution of diverse allochthonous and autochthonous DOM sources. Most studies are, however, confined to local and regional scales, which precludes an understanding of how these mixtures arise at large, e.g., continental, spatial scales. The processes contributing to these mixtures are also difficult to study because of the complex interactions between various environmental factors and DOM. Here we propose the use of machine learning (ML) approaches to identify ecological processes contributing toward mixtures of DOM at a continental-scale. We related a dataset that characterized the molecular composition of DOM from river water and sediment with Fourier-transform ion cyclotron resonance mass spectrometry to explanatory physicochemical variables such as nutrient concentrations and stable water isotopes ( 2 H and 18 O). Using unsupervised ML, distinctive clusters for sediment and water samples were identified, with unique molecular compositions influenced by environmental factors like terrestrial input and microbial activity. Sediment clusters showed a higher proportion of protein-like and unclassified compounds than water clusters, while water clusters exhibited a more diversified chemical composition. We then applied a supervised ML approach, involving a two-stage use of SHapley Additive exPlanations (SHAP) values. In the first stage, SHAP values were obtained and used to identify key physicochemical variables. These parameters were employed to train models using both the default and subsequently tuned hyperparameters of the Histogram-based Gradient Boosting (HGB) algorithm. The supervised ML approach, using HGB and SHAP values, highlighted complex relationships between environmental factors and DOM diversity, in particular the existence of dams upstream, precipitation events, and other watershed characteristics were important in predicting higher chemical diversity in DOM. Our data-driven approach can now be used more generally to reveal the interplay between physical, chemical, and biological factors in determining the diversity of DOM in other ecosystems.

54 ENVIRONMENTAL SCIENCES↗

Active Learning for Anomaly Detection in Environmental data

Due to the growing amount of data from in-situ sensors in environmental monitoring, it becomes necessary to automatically detect anomalous data points. Nowadays, this is mainly performed using supervised machine learning models, which need a fully labelled data set for their training process. However, the process of labelling data is typically cumbersome and, as a result, a hindrance to the adoption of machine learning methods for automated anomaly detection. In this work, we propose to address this challenge by means of active learning. This method consists of querying the domain expert for the labels of only a selected subset of the full data set. We show that this reduces the time and costs associated to labelling while delivering the same or similar anomaly detection performances. Finally, we also show that machine learning models providing a nonlinear classification boundary are to be recommended for anomaly detection in complex environmental data sets.

54 ENVIRONMENTAL SCIENCES↗

Vehicle energy consumption estimation using large scale simulations and machine learning methods

Vehicle energy consumption and powertrain operations for future vehicle powertrain technologies are predicted through full-vehicle simulations using Autonomie, a simulation tool developed at Argonne National Laboratory. Over 1 million vehicles, with different powertrain types (conventional and numerous electrified powertrains) and component technologies (advanced engines, batteries, electrical machines, light weighting) are simulated for outputs including vehicle fuel economy and cost. Despite the ability of Autonomic to predict fuel economy given various combination of vehicle inputs, the full-vehicle simulation is very time-consuming. It takes approximately 84 h to simulate for 33,060 vehicles even using distributed computing on a cluster of 128 worker nodes. Moreover, Autonomic is unable to populate a continuous output space through simulation since it only yields outputs with discrete values instead of an input-output relation. This paper proposes a novel large-scale learning and prediction process (LSLPP) via machine learning approaches to answer the need for a continuous space of vehicle fuel economy and a more efficient simulation process. By learning in a supervised fashion to discover structures in data, the machine learning approaches seek to train a model that can explain the data with which we can perform prediction over unseen data. As such, we are able to accomplish analytic continuation on the current discrete space of simulation results. In addition, we present a random sampling -based numerosity reduction algorithm incorporated in LSLPP that reduces simulation runs and improves efficiency of data procurement. We analytically derive a theoretical bound in order to guarantee the effectiveness of our algorithm. We also compare random sampling to stratified sampling and suggest sampling strategies. A user-friendly tool is developed to support the process. Our experimental results confirm that the proposed LSLPP is able to greatly accelerate prediction and analysis of vehicle fuel economy. The output of this study is used by the US government to evaluate the impact of its R&D funding on energy security and CO 2 emission as well as for the setting of Corporate Average Fuel Economy (CAFE) standards.

32 ENERGY CONSERVATION, CONSUMPTION, AND UTILIZATI↗

Deep Transfer Learning With Field-Based Measurements for Large Area Classification

Classification of Electro-Optical (EO) datasets using Deep Neural Networks (DNNs) has lead to high-performance supervised learning algorithms. However, DNNs require a large amount of labeled data for training which often have limited availability in ecological studies. Up-to-date classification of vegetation in sensitive Arctic ecosystems continue to be a challenge. In an ecosystem undergoing rapid change, capturing the dynamics of vegetation requires the existing maps of vegetation (such as Alaska Existing Vegetation Type (AKEVT), circa 2000) to be updated based on frequent field based observation of vegetation. A method is needed to transfer the knowledge gained from field-collected observations to a larger area using DNNs. Transfer learning is a machine learning technique where a model trained on one task is re-purposed on another similar task. This paper seeks to train DNNs using field-collected observations and apply transfer learning to apply the knowledge gained to update existing vegetation map at larger scale. We test two DNN methods, (1) a deep multilayer perceptron (MLP) model and (2) siamese MLP network that uses a structure to rank similarity between inputs and can be used for training datasets with few samples and show good performance with limited datasets (e.g. few-shot learning). The results show ∼90% accuracy (using the field observations for evaluation) when transfer learning is applied to a siamese network, compared to ∼45% accurate when a MLP is trained on the AKEVT and evaluated on the field observations. The approach show promise for improving and update the existing vegetation maps over large areas using limited field-based observations.

Langford, Zach↗

Comparison of GOES Cloud Classification Algorithms Employing Explicit and Implicit Physics

Cloud-type classification based on multispectral satellite imagery data has been widely researched and demonstrated to be useful for distinguishing a variety of classes using a wide range of methods. The research described here is a comparison of the classifier output from two very different algorithms applied to Geostationary Operational Environmental Satellite (GOES) data over the course of one year. The first algorithm employs spectral channel thresholding and additional physically based tests. The second algorithm was developed through a supervised learning method with characteristic features of expertly labeled image samples used as training data for a 1-nearest-neighbor classification. The latter's ability to identify classes is also based in physics, but those relationships are embedded implicitly within the algorithm. A pixel-to-pixel comparison analysis was done for hourly daytime scenes within a region in the northeastern Pacific Ocean. Considerable agreement was found in this analysis, with many of the mismatches or disagreements providing insight to the strengths and limitations of each classifier. Depending upon user needs, a rule-based or other postprocessing system that combines the output from the two algorithms could provide the most reliable cloud-type classification.

EXPLICIT PHYSICS ALGORITHMS↗

Aero-Engines AI - A Machine-Learning App for Aircraft Engine Concepts Assessment

Effective deployment of trained machine-learning models could drive a high level of efficiency in aircraft engine conceptual design. Aero-Engines AI is a Windows app that has been created to deploy trained machine-learning models to assess aircraft engine concepts. It was created using tkinter, a GUI (graphical user interface) module that is built into the standard Python library. Employing tkinter greatly facilitates the sharing of machine-learning application as an executable file which can be run on Windows machines (without the need to have Python or any library installed). Current version of the app focuses on the performance prediction of conventional turbofans. The app gets user input for a turbofan design, preprocesses the input data, and deploys trained machine-learning models to predict turbofan thrust specific fuel consumption (TSFC), engine weight, core size, and turbomachinery stage-counts. The machine-learning predictive models were built by employing supervised deep-learning algorithm to study patterns in an existing open-source database of production and research turbofan engines. They were trained, cross-validated, and tested in Keras, an open-source neural networks API (application programming interface) written in Python, with TensorFlow (Google open-source artificial intelligence library) serving as the backend engine. The smooth deployment of these machine-learning models using the app shows that Aero-Engines AI is an easy-to-use and a time-saving tool for aircraft engine design-space exploration during the conceptual design stage.

machine learning↗

Water Across Synthetic Aperture Radar Data (WASARD): SAR Water Body Classification for the Open Data Cube

The detection of inland water bodies from Synthetic Aperture Radar (SAR) data provides a great advantage over water detection with optical data, since SAR imaging is not impeded by cloud cover. Traditional methods of detecting water from SAR data involves using thresholding methods that can be labor intensive and imprecise. This paper describes Water Across Synthetic Aperture Radar Data (WASARD): a method of water detection from SAR data which automates and simplifies the thresholding process using machine learning on training data created from Geoscience Australia’s WOFS algorithm. Of the machine learning models tested, the Linear Support Vector Machine was determined to be optimal, with the option of training using solely the VH polarization or a combination of the VH and VV polarizations. WASARD was able to identify water in the target area with a correlation of 97% with WOFS. Sentinel-1, Open Data Cube, Earth Observations, Machine Learning, Water Detection 1. INTRODUCTION Water classification is an important function of Earth imaging satellites, as accurate remote classification of land and water can assist in land use analysis, flood prediction, climate change research, as well as a variety of agricultural applications [2]. The ability to identify bodies of water remotely via satellite is immensely cheaper than contracting surveys of the areas in question, meaning that an application that can accurately use satellite data towards this function can make valuable information available to nations which would not be able to afford it otherwise. Highly reliable applications for the remote detection of water currently exist for use with optical satellite data such as that provided by LANDSAT. One such application, Geoscience Australia’s Water Observations from Space (WOFS) has already been ported for use with the Open Data Cube [6]. However, water detection using optical data from Landsat is constrained by its relatively long revisit cycle of 16 days [5], and water detection using any optical data is constrained in that it lacks the ability to make accurate classifications through cloud cover [2]. The alternative solution which solves these problems is water detection using SAR data, which images the Earth using cloud-penetrating microwaves. Because of its advantages over optical data, much research has been done into water detection using SAR data. Traditionally, this has been done using the thresholding method, which involves picking a polarization band and labeling all pixels for which this band’s value is below a certain threshold as containing water. The thresholding method works since water tends to return a much lower backscatter value to the satellite than land [1]. However, this method can be flawed since estimating the proper threshold is often imprecise, complicated, and labor intensive for the end user. Thresholding also tends to use data from only one SAR polarization, when a combination of polarizations can provide insight into whether water is present. [2] In order to alleviate these problems, this paper presents an application for the Open Data Cube to detect water from SAR data using support vector machine (SVM) classification. 2. PLATFORM WASARD is an application for the Open Data Cube, a mechanism which provides a simple yet efficient means of ingesting, storing, and retrieving remote sensing data. Data can be ingested and made analysis ready according to whatever specifications the researcher chooses, and easily resampled to artificially alter a scene’s resolution. Currently WASARD supports water detection on scenes from ESA’s Sentinel-1 and JAXA’s ALOS. When testing WASARD, Sentinel-1 was most commonly used due to its relatively high spatial resolution and its rapid 6 day revisit cycle [5]. With minor alterations to the application's code, however, it could support data from other satellites. 3. METHODOLOGY Using supervised classification, WASARD compares SAR data to a dataset pre-classified by WOFS in order to train an SVM classifier. This classifier is then used to detect water in other SAR scenes outside the training set. Accuracy was measured according to the following metrics:  Precision: a measure of what percentage of the points WASARD labels as water are truly water  Recall: a measure of what percentage of the total water cover WASARD was able to identify.  F1 Score: a harmonic average of the precision and recall scores Both precision and recall are calculated at the end of the training phase, when the trained classifier is compared to a testing dataset. Because the WOFS algorithm’s classifications are used as the truth values when training a WASARD classifier, when precision and recall are mentioned in this paper, they are always with respect to the values produced by WOFS on a similar scene of Landsat data, which themselves have a classification accuracy of 97% [6]. Visual representations of water identified by WASARD in this paper were produced using the function wasard_plot(), which is included in WASARD. 3.1 Algorithm Selection The machine learning model used by WASARD is the Linear Support Vector Machine (SVM). This model uses a supervised learning algorithm to develop a classifier, meaning it creates a vector which can be multiplied by the vector formed by the relevant data bands to determine whether a pixel in a SAR scene contains water. This classifier is trained by comparing data points from selected bands in a SAR scene to their respective labels, which in this case are “water” or “not water” as given by the WOFS algorithm. The SVM was selected over the Random Forest model, which outperformed the SVM in training speed, but had a greater classification time and lower accuracy, and the Multilayer Perceptron Artificial Neural Network, which had a slightly higher average accuracy than the SVM, but much greater training and classification times. Figure 1: Visual representation of the SVM Classifier. Each white point represents a pixel in a SAR scene. In Figure 1, the diagonal line separating pixels determined to be water from those determined not to be water represents the actual classification vector produced by the SVM. It is worth noting that once the model has been trained, classification of pixels is done in a similar manner as in the thresholding method. This is especially true if only one band was used to train the model. 3.1 Feature Selection Sentinel-1 collects data from two bands: the Vertical/Vertical polarization (VV) and the Vertical/Horizontal polarization (VH). When 100 SVM classifiers were created for each polarization individually, and for the combination of the two, the following results were achieved: Figure 2: Accuracy of classifiers trained using different polarization bands. Precision and Recall were measured with respect to the values produced by WOFS. Figure 2 demonstrates that using both the VV and VH bands trades slightly lower recall for significantly greater precision when compared with the VH band alone, and that using the VV band alone is inferior in both metrics. WASARD therefore defaults to using both the VV and VH bands, and includes the option to use solely the VH band. The VV polarization’s lower precision compared to the VH polarization is in contrast to results from previous research and may merit further analysis [4]. 3.2 Training a Classifier The steps in training a classifier with WASARD are 1. Selecting two scenes (one SAR, one optical) with the same spatial extents, and acquired close to each other in time, with a preference that the scenes are taken on the same day. 2. Using the WOFS algorithm to produce an array of the detected water in the scene of optical data, to be used as the labels during supervised learning 3. Data points from the selected bands from the SAR acquisition are bundled together into an array with the corresponding labels gathered from WOFS. A random sample with an equal number of points labeled “Water” and “Not Water” is selected to be partitioned into a training and a testing dataset 4. Using Scikit-Learn’s LinearSVC object, the training dataset is used to produce a classifier, which is then tested against the testing dataset to determine its precision and recall The result is a wasard_classifier object, which has the following attributes: 1. f1, recall, and precision: 3 metrics used to determine the classifier’s accuracy 2. Coefficient: Vector which the SVM uses to make its predictions. The classifier detects water when the dot product of the coefficient and the vector formed by the SAR bands is positive 3. Save(): allows a user to save a classifier to the disk in order to use it without retraining 4. wasard_classify(): Classifies an entire xarray of SAR data using the SVM classifier All of the above steps are performed automatically when the user creates a wasard_classifier object. 3.3 Classifying a Dataset Once the classifier has been created, it can be used to detect water in an xarray of SAR data using wasard_classify(). By taking the dot product of the classifier’s coefficients and the vector formed by the selected bands of SAR data, an array of predictions is constructed. A classifier can effectively be used on the same spatial extents as the ones where it was trained, or on any area with a similar landscape. While

Kreiser, Zachary↗

Evaluating the Computational Requirements of using SVM software to train Data-Intensive Problems

Support vector machine (SVM) technology is a growing class of computational algorithms for solving classification problems using supervised learning. There are many freely available SVM codes with implementations of varying granularity in the core optimization task. The performance of SVM implementations is related to four key elements; (1) the size and dimension of the data set on which they are operating, (2) the granularity of their core SVM optimization implementation, (3) the kernel transformation applied to the data, and the (4) underlying hardware on which the implementation is running. To assess the performance of different SVM implementations, several freely available codes representing the spectrum of optimization granularity were built on a variety of hardware—two linux clusters, and a shared memory machine. Binary classifiers were trained on datasets of varying size containing anywhere from a few hundred to over 50,000 vectors. Performance of the methods was measured in terms of wall-clock time for training, statistical quality of the resulting classifier, hardware performance (memory footprint, memory bandwidth, floating point unit usage, I/O bandwidth, and instruction stalls), robustness and portability.

32 ENERGY CONSERVATION, CONSUMPTION, AND UTILIZATI↗

Deep learning to estimate permeability using geophysical data

Time-lapse electrical resistivity tomography (ERT) is a popular geophysical method to estimate three-dimensional (3D) permeability fields from electrical potential difference measurements. Traditional inversion and data assimilation methods are used to ingest this ERT data into hydrogeophysical models to estimate permeability. Due to ill-posedness and the curse of dimensionality, existing inversion strategies provide poor estimates and low resolution of the 3D permeability field. Recent advances in deep learning provide us with powerful algorithms to overcome this challenge. This paper presents a deep learning (DL) framework to estimate the 3D subsurface permeability from time-lapse ERT data. To test the feasibility of the proposed framework, we train DL-enabled inverse models on simulation data. Each measurement in both synthetic and field data is standardized by removing the mean and scaling the time-series to unit variance. This pre-processing step is necessary to bring simulation data closer to field observations. Subsurface process models based on hydrogeophysics are used to generate this synthetic data. Training performed on limited simulation data resulted in the DL model over-fitting. An advanced data augmentation based on mixup is implemented to generate additional training samples to overcome this issue. This mixup technique creates weakly labeled (low-fidelity) samples from strongly labeled (high-fidelity) data. The weakly labeled training data is then used to develop DL-enabled inverse models and reduce over-fitting. As both time-lapse ERT (1133048 features/realization) and 3D permeability (585453 features/realization) data samples are from a high-dimensional space, principal component analysis (PCA) is employed to reduce dimensionality. Encoded ERT and encoded permeability are generated using the trained PCA estimators. A deep neural network is then trained to map the encoded ERT to encoded permeability. This mixup training and unsupervised learning allowed us to build a fast and reasonably accurate DL-based inverse model under limited simulation data. Results show that proposed weak supervised learning can capture salient spatial features in the 3D permeability field. Quantitatively, the average mean squared error (in terms of the natural log) on the strongly labeled training, validation, and test datasets is less than 0.5. The R 2 -score (global metric) is greater than 0.75, and the percent error in each cell (local metric) is less than 10%. Finally, an added benefit in terms of computational cost is that the proposed DL-based inverse model is at least O(10 4 ) times faster than running a forward model once it is trained. Data generation, DL model training, and hyperparameter tuning to identify optimal neural network architectures utilized high-performance computing resources while the DL inference is performed on a standard laptop. Approximately, O(10 5 ) processor hours are used for generating data and DL tuning and training. We acknowledge that the data generation and DL model development are expensive. But once a DL model is trained, it can be re-used for inversion rapidly for the given system, with set physics and domain. Note that traditional inversion may require multiple forward model simulations (e.g., in the order of 10 to 1000), which are very expensive. This computational savings ≈ O(10 5 ) – O(10 7 )) makes the proposed DL-based inverse model attractive for subsurface imaging and real-time ERT monitoring applications due to fast and yet reasonably accurate estimations of permeability field.

58 GEOSCIENCES↗

Hashed Binary Search Sampling for Convolutional Network Training with Large Overhead Image Patches

Very large overhead imagery associated with ground truth maps has the potential to generate billions of training image patches for machine learning algorithms. However, random sampling selection criteria often leads to redundant and noisy-image patches for model training. With minimal research efforts behind this challenge, the current status spells missed opportunities to develop supervised learning algorithms that generalize over wide geographical scenes. In addition, much of the computational cycles for large scale machine learning are poorly spent crunching through noisy and redundant image patches. We demonstrate a potential framework to address these challenges specifically, while evaluating a human settlement detection task. A novel binary search tree sampling scheme is fused with a kernel based hashing procedure that maps image patches into hash-buckets using binary codes generated from image content. The framework exploits inherent redundancy within billions of image patches to promote mostly high variance preserving samples for accelerating algorithmic training and increasing model generalization.

Lunga, Dalton↗