SEARCH · Search NASA

Results for “Machine learning models”

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

SourceResearch connectionFrom yearThrough year

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

At least 271 records · Page 15

All sources NASA NTRS DOE OSTI

DOE OSTI2024

Designing complex concentrated alloys with quantum machine learning and language modeling

Designing novel complex concentrated alloys (CCAs) is an essential topic in materials science. However, due to the complicated high-dimensional component-property relationship, tuning material properties by researchers’ experience is challenging, even when guided by physical or empirical rules. Here, we adopt quantum computing (QC) technology and machine learning models to provide a proof-of-concept application of QC in physical metallurgy. We propose a quantum support vector machine (QSVM) model to predict single-phase CCAs. We show that fine-tuned quantum kernels with entanglement deliver promising performance, with a maximum accuracy of 89.4%. The QSVM model is then used to identify 1,741 lightweight CCAs jointly with a new text-mining-based method. Meanwhile, we devise a controllable approach to study the effect of noise on model performance and find that the noise level needs to be minimized for high-performance QSVM models. Finally, this study provides a practical and general approach to designing CCAs based on quantum technologies.

36 MATERIALS SCIENCE↗

DOE OSTI2025

PowerModel-AI: A First On-the-Fly Machine-Learning Predictor for AC Power Flow Solutions

The real-time creation of machine-learning models via active or on-the-fly learning has attracted considerable interest across various scientific and engineering disciplines. These algorithms enable machines to build models autonomously while remaining operational. Through a series of query strategies, the machine can evaluate whether newly encountered data fall outside the scope of the existing training set. In this study, we introduce PowerModel-AI, an end-to-end machine learning software designed to accurately predict AC power flow solutions. We present detailed justifications for our model design choices and demonstrate that selecting the right input features effectively captures load flow decoupling inherent in power flow equations. Our approach incorporates on-the-fly learning, where power flow calculations are initiated only when the machine detects a need to improve the dataset in regions where the model’s suboptimal performance is based on specific criteria. Otherwise, the existing model is used for power flow predictions. This study includes analyses of five Texas A&M synthetic power grid cases, encompassing the 14-, 30-, 37-, 200-, and 500-bus systems. The training and test datasets were generated using PowerModels.jl, an open-source power flow solver/optimizer developed at Los Alamos National Laboratory, NM, USA.

24 POWER TRANSMISSION AND DISTRIBUTION↗

DOE OSTI2026

Xanthos-Lake Dataset

The Xanthos-Lake v1.0 dataset provides the input data, trained machine-learning models, and simulation outputs needed to characterize lake water balance, snow and ice conditions, and mixing-layer temperature within the Xanthos global hydrological modeling framework. The dataset supports lake representation across a wide range of lake sizes and hydroclimatic conditions by combining xLSIM, a basin-specific machine-learning emulator of lake snow, ice, ice-cover fraction, and mixing-layer temperature, with the Xanthos-Lake water-balance model. The archive contains NetCDF datasets used to train and evaluate xLSIM, trained model weights, processed meteorological and lake-property inputs, and basin- and lake-category-specific simulation outputs. These materials are organized into four primary data groups, described below. Snowice_model_inputs: Contains the NetCDF input data used to train xLSIM. The xLSIM machine-learning framework uses three lake-based datasets. The meteorological forcing dataset provides monthly relative humidity, specific humidity, surface wind speed, maximum and minimum air temperature, downward longwave and shortwave radiation, snowfall, surface air pressure, and total precipitation. Lake surface area is included as an additional static predictor. The target-state dataset provides lake ice thickness, snow depth, snow cover, and lake mixing-layer temperature, while a companion lake-surface dataset provides the lake ice-cover fraction. Before training, ice thickness and snow depth are converted from meters to centimeters, mixing-layer temperature is converted from kelvin to degrees Celsius and constrained to nonnegative values, and ice-cover fraction is converted from a fraction to a percentage. The predictor variables are normalized using statistics calculated across the selected lakes and time steps. Snowice_model_outputs: Contains the NetCDF outputs generated by xLSIM. For each basin, xLSIM produces a file containing observed and predicted lake-state variables for the training, validation, and testing periods. The modeled variables include lake ice thickness, snow depth, snow cover, mixing-layer temperature, and lake ice-cover fraction. For basins without a sufficiently persistent snow-and-ice signal, the emulator predicts only mixing-layer temperature. The outputs also include training and validation loss histories, the selected model configuration, identifiers of the lakes used in training, and SHAP-based feature-importance information at the global, lake, and seasonal-regime levels. The trained machine-learning model weights are provided separately within the dataset archive. Together, these files support model evaluation and subsequent coupling with the Xanthos-Lake water-balance framework. XanthosLAKES: Contains the NetCDF input data used by the Xanthos-Lake framework. Monthly meteorological inputs include relative and specific humidity, downward shortwave and longwave radiation, mean, maximum, and minimum air temperature, wind speed, precipitation, snowfall, and surface air pressure. Static lake-property datasets provide lake identifiers, geographic locations, surface area, volume, mean depth, elevation, drainage area, fetch, outlet-routing information, and associated Xanthos grid-cell attributes. Separate bathymetric datasets provide the coefficients of the area–depth and volume–depth relationships for each aggregated lake unit. GLEV-based records provide observed lake surface area and evaporation data used to initialize lake states, define reference conditions, and calibrate and evaluate the model. Xanthos-Lake Outputs: Contains the basin- and lake-category-specific NetCDF outputs generated by Xanthos-Lake. Monthly variables include lake surface area, storage volume, outlet discharge, evaporation rate, evaporation volume, lake–groundwater exchange, lake inflow, ice thickness, snow depth, snow-cover fraction, ice-cover fraction, and mixing-layer temperature. The files also contain lake-specific calibration and validation statistics, including normalized root-mean-square error, mean absolute error, Nash–Sutcliffe efficiency, Kling–Gupta efficiency, and percent bias. Stored calibrated and derived parameters include the weir discharge coefficient, fractional freeboard, groundwater exchange coefficient, reference water level, corresponding reference surface area and storage volume, weir-width adjustment factor, and the fraction of routed inflow entering the lake. Basin identifiers, lake category, simulation period, calibration and validation periods, and parameter-schema information are retained as NetCDF metadata.

Abeshu, Guta [Pacific Northwest National Laborator↗

DOE OSTI2024

Power modeling of degraded PV systems: Case studies using a dynamically updated physical model (PV-Pro)

Power modeling, widely applied for health monitoring and power prediction, is crucial for the efficiency and reliability of Photovoltaic (PV) systems. The most common approach for power modeling uses a physical equivalent circuit model, with the core challenge being the estimation of model parameters. Traditional parameter estimation either relies on datasheet information, which does not reflect the system's current health status, especially for degraded PV systems, or requires additional I-V characterization, which is generally unavailable for large-scale PV systems. Thus, we build upon our previously developed tool, PV-Pro (originally proposed for degradation analysis), to enhance its application for power modeling of degraded PV systems. PV-Pro extracts model parameters from production data without requiring I-V characterization. This dynamic model, periodically updated, can closely capture the actual degradation status, enabling precise power modeling. PV-Pro is compared with popular power modeling techniques, including persistence, nominal physical, and various machine learning models. The results indicate that PV-Pro achieves outstanding power modeling performance, with an average nMAE of 1.4 % across four field-degraded PV systems, reducing error by 17.6 % compared to the best alternative technique. Furthermore, PV-Pro demonstrates robustness across different seasons and severities of degradation. The tool is available as a Python package at https://github.com/DuraMAT/pvpro.

14 SOLAR ENERGY↗

DOE OSTI2025

..delta..-Learning of High-Fidelity Electronic Structure Using Graph Neural Networks with Modified Node-Level Features

In this work, we present a ..delta..-learning approach for predicting the eigenvalues calculated with the hybrid functional HSE06 (..epsilon..nkHSE) for a set of metal and nitrogen doped graphene catalysts (MNCs) from Perdew-Burke-Ernzerhof (PBE) inputs. The model presented here incorporates electronic scalar features along with structural information in a graph neural network (GNN). In particular, the PBE eigenvalues for different bands and k-points and orbital-resolved projectors are combined with the applied potential as node-level features along with structural information within the Atomistic Line Graph Neural Network (ALIGNN) architecture. These features enable flexibility for systems with electrified interfaces, such as in electrocatalysts and achieves mean absolute error (MAE) of less than 0.1 eV. The machine learning model reported here achieves a strong generalization to left-out adsorbates (MAE = 0.074 eV) and leave-one-chemical-space-out (MAE = 0.08 eV) and completely left-out metals (MAE = 0.072 eV), confirming the robustness of the machine learning (ML) model in predicting ..epsilon..nkHSE.

36 MATERIALS SCIENCE↗

DOE OSTI2024

wa-hls4ml: A GNN Surrogate Model for hls4ml

Recent advancements in use of machine learning techniques on field-programmable gate arrays (FPGAs) have allowed for implementation of embedded neural networks with extremely low latency. This is invaluable for particle detectors at the Large Hadron Collider, where latency and used area must be strictly bounded. The hls4ml framework is a procedure for converting from trained machine learning model software, to a synthesis result that can be used on an FPGA. However, running the pipeline is a time-consuming procedure, and there is a strong risk of failure. In particular, it is possible that the model is unable to be converted into a synthesis result, or that the resource consumption of the model will exceed the resources of the target FPGA. To aid with this development, we introduce wa-hls4ml, a surrogate model which uses a graph neural network to emulate the structure of the source models. The goal is to estimate the chance of success and resource consumption of an arbitrary model when passed through the hls4ml procedure, without the time consumption of actually running the pipeline.

43 PARTICLE ACCELERATORS↗

DOE OSTI2024

Acoustic-based monitoring and machine learning of component status for microreactor applications

This report provides a description and assessment of recent efforts to couple acoustic-based experimental measurements and characterization with machine learning models in order to enhance structural health monitoring capabilities for nuclear microreactors. With resilient embedded sensors in development by others supported by programs funded by the US Department of Energy’s Office of Nuclear Energy, the work described herein builds upon ongoing efforts to improve non-destructive testing technology that relates measured acoustic signatures to component stresses and/or structural defects, using a combination of new experimental measurements and machine learning architectures. The experimental procedure remained similar to that developed for the previous year’s demonstration of damage detection by the authors, with the same damaged sample tested under similar applied stress conditions. Notably, a new mounting fixture was designed and implemented to improve measurement consistency and a more sophisticated laser Doppler vibrometer was employed to make high-fidelity vibration measurements. Two nominally identical sets of training data were collected for each experimental setup to better understand the repeatability of the experiment and to better test the generality of trained neural network models. Additionally, we obtained new high-quality 3D mode shapes of the damaged test article at various stress and excitation levels, providing greater insights into the physical response of the sample during testing. Previously, we demonstrated that a machine learning model based on a convolutional neural network can predict structural details of an artificially introduced interface (intact, rough cut, smooth cut), and the applied torque level. In this study, we have transitioned to graph-based neural network architectures to better develop and test a flexible framework that is more suitable to being transferred away from controlled benchtop experiments and into more applied settings where less-structured data inputs may be expected. In general, performance testing of a graph neural network on frequency-domain representations of the data indicates strong and consistent identification of test conditions for datasets recorded on damaged components. With goals of predicting damage location and other changing experimental conditions using limited datasets, predictive models using a graph neural network architecture correctly predicted the applied torque level with an accuracy of 85% using only a single measurement point and predicted within one torque level in 95% of test windows. Predictions of damage location had limited success due to the symmetry and minimal number of the damage scenarios presented during model training. Results were ambiguous as to whether the model could detect the location of the artificial damage, or if it was instead learning the location of a given measurement point on the part and subsequently detecting which points were closest to the location of the damage. This finding will be factored into upcoming planned work on damaged graphite components, where new experimental tests with a larger number and variety of damage scenarios are expected to provide improved validation of recent developments in monitoring methodology.

22 GENERAL STUDIES OF NUCLEAR REACTORS↗

DOE OSTI2025

A knowledge-informed large language model framework for U.S. nuclear power plant shutdown initiating event classification for probabilistic risk assessment

Identifying and classifying shutdown initiating events (SDIEs) is critical for developing shutdown probabilistic risk assessment for nuclear power plants. Existing computational approaches cannot achieve satisfactory performance due to the challenges of unavailable large, labeled datasets, imbalanced event types, and label noise. To address these challenges, we propose a hybrid pipeline that integrates a knowledge-informed machine learning model to prescreen non-SDIEs and a large language model (LLM) to classify SDIEs into four types. In the prescreening stage, we proposed a set of 44 SDIE text patterns that consist of the most salient keywords and phrases from six SDIE types. Text vectorization based on the SDIE patterns generates feature vectors that are highly separable by using a simple binary classifier. The second stage builds Bidirectional Encoder Representations from Transformers (BERT)-based LLM, which learns generic English language representations from self-supervised pretraining on a large dataset and adapts to SDIE classification by fine-tuning it on an SDIE dataset. The proposed approaches are evaluated on a dataset with 10,928 events using precision, recall ratio, F 1 score, and average accuracy. In conclusion, the results demonstrate that the prescreening stage can exclude more than 97% non-SDIEs, and the LLM achieves an average accuracy of 95.1% for SDIE classification.

99 - GENERAL AND MISCELLANEOUS↗

DOE OSTI2024

Modeling with uncertainty quantification reveals the essentials of a non-canonical algal carbon-concentrating mechanism

The thermoacidophilic red alga Cyanidioschyzon merolae survives its challenging environment likely in part by operating a carbon-concentrating mechanism (CCM). Here, we demonstrated that C. merolae 's cellular affinity for CO 2 is stronger than the affinity of its rubisco for CO 2 . This finding provided additional evidence that C. merolae operates a CCM while lacking the structures and functions characteristic of CCMs in other organisms. To test how such a CCM could function, we created a mathematical compartmental model of a simple CCM, distinct from those we have seen previously described in detail. The results of our modeling supported the feasibility of this proposed minimal and non-canonical CCM in C. merolae . To facilitate the robust modeling of this process, we measured and incorporated physiological and enzymatic parameters into the model. Additionally, we trained a surrogate machine-learning model to emulate the mechanistic model and characterized the effects of model parameters on key outputs. This parameter exploration enabled us to identify model features that influenced whether the model met the experimentally derived criteria for functional carbon concentration and efficient energy usage. Such parameters included cytosolic pH, bicarbonate pumping cost and kinetics, cell radius, carboxylation velocity, number of thylakoid membranes, and CO 2 membrane permeability. Our exploration thus suggested that a non-canonical CCM could exist in C. merolae and illuminated the essential features generally necessary for CCMs to function.

Steensma, Anne K. [Michigan State Univ., East Lans↗

DOE OSTI2024

The need for carbon-emissions-driven climate projections in CMIP7

Abstract. Previous phases of the Coupled Model Intercomparison Project (CMIP) have primarily focused on simulations driven by atmospheric concentrations of greenhouse gases (GHGs), for both idealized model experiments and climate projections of different emissions scenarios. We argue that although this approach was practical to allow parallel development of Earth system model simulations and detailed socioeconomic futures, carbon cycle uncertainty as represented by diverse, process-resolving Earth system models (ESMs) is not manifested in the scenario outcomes, thus omitting a dominant source of uncertainty in meeting the Paris Agreement. Mitigation policy is defined in terms of human activity (including emissions), with strategies varying in their timing of net-zero emissions, the balance of mitigation effort between short-lived and long-lived climate forcers, their reliance on land use strategy, and the extent and timing of carbon removals. To explore the response to these drivers, ESMs need to explicitly represent complete cycles of major GHGs, including natural processes and anthropogenic influences. Carbon removal and sequestration strategies, which rely on proposed human management of natural systems, are currently calculated in integrated assessment models (IAMs) during scenario development with only the net carbon emissions passed to the ESM. However, proper accounting of the coupled system impacts of and feedback on such interventions requires explicit process representation in ESMs to build self-consistent physical representations of their potential effectiveness and risks under climate change. We propose that CMIP7 efforts prioritize simulations driven by CO2 emissions from fossil fuel use and projected deployment of carbon dioxide removal technologies, as well as land use and management, using the process resolution allowed by state-of-the-art ESMs to resolve carbon–climate feedbacks. Post-CMIP7 ambitions should aim to incorporate modeling of non-CO2 GHGs (in particular, sources and sinks of methane and nitrous oxide) and process-based representation of carbon removal options. These developments will allow three primary benefits: (1) resources to be allocated to policy-relevant climate projections and better real-time information related to the detectability and verification of emissions reductions and their relationship to expected near-term climate impacts, (2) scenario modeling of the range of possible future climate states including Earth system processes and feedbacks that are increasingly well-represented in ESMs, and (3) optimal utilization of the strengths of ESMs in the wider context of climate modeling infrastructure (which includes simple climate models, machine learning approaches and kilometer-scale climate models).

54 ENVIRONMENTAL SCIENCES↗

DOE OSTI2026

Particle trajectory representation learning with masked point modeling

Liquid argon time projection chambers (LArTPCs) offer millimeter-scale 3D images of particle trajectories, enabling precision studies of neutrino oscillation, detection of supernova and solar neutrinos, searches for exotic dark matter, and proton decay. Current approaches utilize supervised machine learning models, requiring extensive simulations of particle physics and detector response that can introduce bias. Self-supervised learning (SSL), a machine learning approach that learns useful representations of unlabeled data from the data itself, has significantly advanced how large datasets are utilized for representation learning; however, its potential for applications to sensory data in high precision particle physics experiments remains largely unexplored. We introduce the Point-based liquid argon masked autoencoder (PoLAr-MAE), a self-supervised framework that learns physically meaningful representations directly from unlabeled LArTPC images. PoLAr-MAE achieves remarkable data efficiency for a point-level segmentation task, outperforming fully supervised methods in low data regimes. Linear classifiers on model outputs demonstrate robust performance across multiple downstream tasks. Our results position sensor-level SSL as a practical foundation model strategy for LArTPCs.

Young, Samuel [Stanford Univ., CA (United States)]↗

DOE OSTI2024

A tutorial review of machine learning-based model predictive control methods

Abstract This tutorial review provides a comprehensive overview of machine learning (ML)-based model predictive control (MPC) methods, covering both theoretical and practical aspects. It provides a theoretical analysis of closed-loop stability based on the generalization error of ML models and addresses practical challenges such as data scarcity, data quality, the curse of dimensionality, model uncertainty, computational efficiency, and safety from both modeling and control perspectives. The application of these methods is demonstrated using a nonlinear chemical process example, with open-source code available on GitHub. The paper concludes with a discussion on future research directions in ML-based MPC.

Wu, Zhe [Department of Chemical and Biomolecular E↗

DOE OSTI2024

Applying Machine‐Learning Methods to Laser Acceleration of Protons: Lessons Learned From Synthetic Data

ABSTRACT In this study, we consider three different machine‐learning methods—a three‐hidden‐layer neural network, support vector regression, and Gaussian process regression—and compare how well they can learn from a synthetic data set for proton acceleration in the Target Normal Sheath Acceleration regime. The synthetic data set was generated from a previously published theoretical model by Fuchs et al. 2005 that we modified. Once trained, these machine‐learning methods can assist with efforts to maximize the peak proton energy, or with the more general problem of configuring the laser system to produce a proton energy spectrum with desired characteristics. In our study, we focus on both the accuracy of the machine‐learning methods and the performance on one GPU including memory consumption. Although it is arguably the least sophisticated machine‐learning model we considered, support vector regression performed very well in our tests.

Desai, Ronak↗

DOE OSTI2026

Machine learning-enhanced hybrid modeling approach for better identification of a building thermal network model and improved prediction

The gray-box modeling approach, which uses a semi-physical thermal network model, has been widely used in building prediction applications, such as model predictive control (MPC). However, unmeasured disturbances, such as occupants, lighting, and in/exfiltration loads, make it challenging to apply this approach to practical buildings. In this word, we propose a hybrid modeling approach that integrates the gray-box model with a model for unmeasured disturbance. After reviewing several system identification approaches, we systematically designed the unmeasured disturbance model with a model selection process based on statistical tests to make it robust. We generated data based on the building model calibrated by real operational data and then trained the hybrid model for two different weather conditions. The hybrid model approach demonstrates an RMSE reduction of approximately 0.2–0.9 °C and 0.3–2 °C on 1-day ahead temperature prediction compared to the Conventional approach for mild (Berkeley, CA) and cold (Chicago, IL) climates, respectively. In addition, this approach was applied to experimental data obtained from the laboratory building to be used for the MPC application, showing superior prediction performances.

32 ENERGY CONSERVATION, CONSUMPTION, AND UTILIZATI↗

DOE OSTI2024

Observational benchmarks inform representation of soil organic carbon dynamics in land surface models

Abstract. Representing soil organic carbon (SOC) dynamics in Earth system models (ESMs) is a key source of uncertainty in predicting carbon–climate feedbacks. Machine learning models can help identify dominant environmental controllers and establish their functional relationships with SOC stocks. The resulting knowledge can be integrated into ESMs to reduce uncertainty and improve predictions of SOC dynamics over space and time. In this study, we used a large number of SOC field observations (n=54 000), geospatial datasets of environmental factors (n=46), and two machine learning approaches (namely random forest, RF, and generalized additive modeling, GAM) to (1) identify dominant environmental controllers of global and biome-specific SOC stocks, (2) derive functional relationships between environmental controllers and SOC stocks, and (3) compare the identified environmental controllers and predictive relationships with those in models used in Phase 6 of the Coupled Model Intercomparison Project (CMIP6). Our results showed that the diurnal temperature, drought index, cation exchange capacity, and precipitation were important observed environmental predictors of global SOC stocks. While the RF model identified 14 environmental factors that describe climatic, vegetation, and edaphic conditions as important predictors of global SOC stocks (R2=0.61, RMSE = 0.46 kg m−2), current ESMs oversimplify the relationships between environmental factors and SOC, with precipitation, temperature, and net primary productivity explaining > 96 % of the variability in ESM-modeled SOC stocks. Further, our study revealed notable disparities among the functional relationships between environmental factors and SOC stocks simulated by ESMs compared with observed relationships. To improve SOC representations in ESMs, it is imperative to incorporate additional environmental controls, such as the cation exchange capacity, and refine the functional relationships to align more closely with observations.

54 ENVIRONMENTAL SCIENCES↗

DOE OSTI2025

Atomate2: modular workflows for materials science

High-throughput density functional theory (DFT) calculations have become a vital element of computational materials science, enabling materials screening, property database generation, and training of “universal” machine learning models. While several software frameworks have emerged to support these computational efforts, new developments such as machine learned force fields have increased demands for more flexible and programmable workflow solutions. This manuscript introduces atomate2, a comprehensive evolution of our original atomate framework, designed to address existing limitations in computational materials research infrastructure. Key features include the support for multiple electronic structure packages and interoperability between them, along with generalizable workflows that can be written in an abstract form irrespective of the DFT package or machine learning force field used within them. Our hope is that atomate2's improved usability and extensibility can reduce technical barriers for high-throughput research workflows and facilitate the rapid adoption of emerging methods in computational material science.

97 MATHEMATICS AND COMPUTING↗

DOE OSTI2025

Using feature importance as an exploratory data analysis tool on Earth system models

Abstract. Machine learning (ML) models are commonly used to generate predictions, but these models can also support the discovery of new science. Generating accurate predictions necessitates that a model captures the structure of the underlying data. If the structure is properly extracted, ML could be a useful exploratory and evidential tool. In this paper, we present a case study that demonstrates the use of ML for exploratory data analysis (EDA) in the climate space. We apply the ML explainability method of spatiotemporal zeroed feature importance (stZFI) to understand how climate-variable associations evolve over space and time. Our analyses focus on data from ensembles of Earth system models (ESMs) which provide data on different climate states and conditions. We elect to work with ESM ensembles since they allow us to compare feature importance across alternative scenarios not available with observed data. The ensembles also account for natural variability so that we can distinguish between signal and noise due to natural climate variability when computing feature importance. The use of perturbed initial condition ensembles introduces variability mimicking the natural variability in the atmosphere; thus the signals emerging using feature importance (FI) can be evaluated against the natural variability in the climate system. For our analyses, we consider the 1991 volcanic eruption of Mount Pinatubo, which was a large stratospheric aerosol injection. We explore the climate pathway associated with the eruption from aerosols to radiation to temperature at both the near-surface and stratospheric levels. In addition to applying the method to data generated from two different ESMs, we apply stZFI to reanalysis data to compare the associations identified by stZFI. We show how stZFI tracks the importance of aerosol optical depth over time on forecasting temperatures. This case study illustrates usefulness of an ML tool (stZFI) for EDA on a well-studied climate exemplar.

Ries, Daniel (ORCID:0000000250294647)↗

DOE OSTI2025

Self-Supervised and Interpretable Anomaly Detection Using Network Transformers

Machine learning and deep neural networks (DNNs) have been proposed as a tool to identify anomalies in computer network communications. However, due the obfuscated nature of off-the-shelf machine learning models, their output often does not provide enough information to isolate the source of the anomaly to take corrective measures. In this article, we introduce the network transformer (NeT), a DNN model for anomaly detection that incorporates the graph structure of the communication network in order to improve interpretability. Further, the presented approach has the following advantages: first, enhanced interpretability by incorporating the graph structure of computer networks; second, provides a hierarchical set of features that enables analysis at different levels of granularity; second, self-supervised training that does not require labeled data. The NeT model was evaluated on a set of anomalous scenarios executed in a real industrial control system. The presented approach successfully identified the anomalies, the devices affected, and the specific connections causing the anomalies, providing a data-driven hierarchical approach to analyze the behavior of a cyber network.

97 MATHEMATICS AND COMPUTING↗