Engineering Papers⌕ Search

SEARCH · Engineering Papers

Results for “Multiple Errors”

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 1,099 records · Page 61

Reconfigurable Microwave Phase Delay Element for Frequency Reference and Phase-Shifter Applications

A technique was developed to provide a reconfigurable high-precision micro - wave electrical phase delay for resonators and phase shifters. The invention employs multiple branches of transmission lines with open-ended or ground-ended terminations as configurable bits or digits. This technique minimizes the errors due to limited precision of switching devices. In addition, the proposed linear analytical approach significantly produces a much simpler design than that of other prior inventions at the time of this reporting. Microwave components such as filters, phase delay elements, or resonators require a method that can accurately adjust their frequency responses. Most tuning techniques offer very wide frequency tuning range; however, it is often difficult and expensive to tune their response in a very narrow operating frequency, especially when the tuning element reaches its minimum discrete step due to fabrication tolerances. The problem becomes worse as the operating frequency is in mm-wave frequency range (>26 GHz). The electrical tuning sensitivity of a microwave line is dependent on the position of the tuning element with respect to the reference termination. By placing this tuning element away from this reference with the main transmission line connecting the two elements together the sensitivity of the tuning element can change significantly. This concept can be used in the system that requires multiple tuning sensitivities. In this case, multiple tuning branches are superimposed in the main transmission line. The proposed invention allows the transmission-line electrical length to be accurately programmed using switching elements that have limited accuracy. The invention consists of multiple branches of transmission lines connected to discrete switching devices with open-ended terminations. They are used as discrete tuning elements. These elements are connected to the main microwave transmission line and are separated by a well-defined electrical degree spacing. Each branch is programmed to have different electrical degree sensitivity, such as a combination of discrete steps in each branch, which results in a reflective line with a unique effective phase response. To reduce the number of switching devices, it is desirable to program the devices in binary configuration where each branch represents one bit in the base-2 number system. This invention allows the transmission line electrical length to be tuned precisely with customizable sensitivity based on the known sensitivity of the base tuning circuit. The tuning resolution is dependent on the distance among tuning branches.

Hsieh, Wen-Ting↗

Development of an open-source regional data assimilation system in PEcAn v. 1.7.2: application to carbon cycle reanalysis across the contiguous US using SIPNET

Abstract. The ability to monitor, understand, and predict the dynamics of the terrestrial carbon cycle requires the capacity to robustly and coherently synthesize multiple streams of information that each provide partial information about different pools and fluxes. In this study, we introduce a new terrestrial carbon cycle data assimilation system, built on the PEcAn model–data eco-informatics system, and its application for the development of a proof-of-concept carbon “reanalysis” product that harmonizes carbon pools (leaf, wood, soil) and fluxes (GPP, Ra, Rh, NEE) across the contiguous United States from 1986–2019. We first calibrated this system against plant trait and flux tower net ecosystem exchange (NEE) using a novel emulated hierarchical Bayesian approach. Next, we extended the Tobit–Wishart ensemble filter (TWEnF) state data assimilation (SDA) framework, a generalization of the common ensemble Kalman filter which accounts for censored data and provides a fully Bayesian estimate of model process error, to a regional-scale system with a calibrated localization. Combined with additional workflows for propagating parameter, initial condition, and driver uncertainty, this represents the most complete and robust uncertainty accounting available for terrestrial carbon models. Our initial reanalysis was run on an irregular grid of ∼ 500 points selected using a stratified sampling method to efficiently capture environmental heterogeneity. Remotely sensed observations of aboveground biomass (Landsat LandTrendr) and leaf area index (LAI) (MODIS MOD15) were sequentially assimilated into the SIPNET model. Reanalysis soil carbon, which was indirectly constrained based on modeled covariances, showed general agreement with SoilGrids, an independent soil carbon data product. Reanalysis NEE, which was constrained based on posterior ensemble weights, also showed good agreement with eddy flux tower NEE and reduced root mean square error (RMSE) compared to the calibrated forecast. Ultimately, PEcAn's new open-source regional data assimilation framework provides a scalable workflow for harmonizing multiple data constraints and providing a uniform synthetic platform for carbon monitoring, reporting, and verification (MRV) as well as accelerating terrestrial carbon cycle research.

54 ENVIRONMENTAL SCIENCES↗

Machine Learning Based Multi-Physical-Model Blending for Enhancing Renewable Energy Forecast -- Improvement via Situation Dependent Error Correction

With increasing penetration of solar and wind energy to the total energy supply mix, the pressing need for accurate energy forecasting has become well-recognized. Here we report the development of a machine-learning based model blending approach for statistically combining multiple meteorological models for improving the accuracy of solar/wind power forecast. Importantly, we demonstrate that in addition to parameters to be predicted (such as solar irradiance and power), including additional atmospheric state parameters which collectively define weather situations as machine learning input provides further enhanced accuracy for the blended result. Functional analysis of variance shows that the error of individual model has substantial dependence on the weather situation. The machine-learning approach effectively reduces such situation dependent error thus produces more accurate results compared to conventional multi-model ensemble approaches based on simplistic equally or unequally weighted model averaging. Validation over an extended period of time results show over 30% improvement in solar irradiance/power forecast accuracy compared to forecasts based on the best individual model.

29 ENERGY PLANNING, POLICY, AND ECONOMY↗

Advanced Error Modeling for Prompt Yield Estimation [Slides]

Multi-Phenomenology Explosion Monitoring (MultiPEM) Toolbox provides a capability for post-detonation characterization focusing on yield estimation; Synthesize information from multiple sensor types; Leverage historical databases to inform statistical models; Allows for rapid or comprehensive analysis; Implemented in R code.

45 MILITARY TECHNOLOGY, WEAPONRY, AND NATIONAL DEF↗

An investigation of errors and data processing techniques for an RF multilateration system

The development of an RF Multilateration system to provide accurate position and velocity measurements during the approach and landing phase of Vertical Takeoff Aircraft operation is discussed. The system uses an angle-modulated ranging signal to provide both range and range rate measurements between an aircraft transponder and multiple ground stations. Range and range rate measurements are converted to coordinate measurements and the coordinate and coordinate rate information is transmitted by an integral data link to the aircraft. Data processing techniques are analyzed to show advantages and disadvantages. Error analyses are provided to permit a comparison of the various techniques.

Britt, C. L., Jr.↗

Sub-daily Statistical Downscaling of Meteorological Variables Using Neural Networks

A new open source neural network temporal downscaling model is described and tested using CRU-NCEP reanal ysis and CCSM3 climate model output. We downscaled multiple meteorological variables in tandem from monthly to sub-daily time steps while also retaining consistent correlations between variables. We found that our feed forward, error backpropagation approach produced synthetic 6 hourly meteorology with biases no greater than 0.6% across all variables and variance that was accurate within 1% for all variables except atmospheric pressure, wind speed, and precipitation. Correlations between downscaled output and the expected (original) monthly means exceeded 0.99 for all variables, which indicates that this approach would work well for generating atmospheric forcing data consistent with mass and energy conserved GCM output. Our neural network approach performed well for variables that had correlations to other variables of about 0.3 and better and its skill was increased by downscaling multiple correlated variables together. Poor replication of precipitation intensity however required further post-processing in order to obtain the expected probability distribution. The concurrence of precipitation events with expected changes in sub ordinate variables (e.g., less incident shortwave radiation during precipitation events) were nearly as consistent in the downscaled data as in the training data with probabilities that differed by no more than 6%. Our downscaling approach requires training data at the target time step and relies on a weak assumption that climate variability in the extrapolated data is similar to variability in the training data.

54 ENVIRONMENTAL SCIENCES↗

Shared Dosimetry Error in Epidemiological Dose-Response Analyses

Radiation dose reconstruction systems for large-scale epidemiological studies are sophisticated both in providing estimates of dose and in representing dosimetry uncertainty. For example, a computer program was used by the Hanford Thyroid Disease Study to provide 100 realizations of possible dose to study participants. The variation in realizations reflected the range of possible dose for each cohort member consistent with the data on dose determinates in the cohort. Another example is the Mayak Worker Dosimetry System 2013 which estimates both external and internal exposures and provides multiple realizations of "possible" dose history to workers given dose determinants. This paper takes up the problem of dealing with complex dosimetry systems that provide multiple realizations of dose in an epidemiologic analysis. In this paper we derive expected scores and the information matrix for a model used widely in radiation epidemiology, namely the linear excess relative risk (ERR) model that allows for a linear dose response (risk in relation to radiation) and distinguishes between modifiers of background rates and of the excess risk due to exposure. We show that treating the mean dose for each individual (calculated by averaging over the realizations) as if it was true dose (ignoring both shared and unshared dosimetry errors) gives asymptotically unbiased estimates (i.e. the score has expectation zero) and valid tests of the null hypothesis that the ERR slope β is zero. Although the score is unbiased the information matrix (and hence the standard errors of the estimate of β) is biased for β≠0 when ignoring errors in dose estimates, and we show how to adjust the information matrix to remove this bias, using the multiple realizations of dose. Use of these methods for several studies, including the Mayak Worker Cohort and the U.S. Atomic Veterans Study, is discussed.

epidemiological analysis↗

Effects of Omitting Non-confounding Predictors From General Relative-Risk Models for Binary Outcomes

The effects, in terms of bias and precision, of omitting non-confounding predictive covariates from generalized linear models have been well studied, and it is known that such omission results in attenuation bias but increased precision with logistic regression. However, many epidemiologic risk analyses utilize alternative models that are not based on a linear predictor, and the effect of omitting non-confounding predictive covariates from such models has not been characterized. We employed simulation to study the effects on risk estimation of omitting non-confounding predictive covariates from an excess relative risk (ERR) model and a general additive-multiplicative relative-risk mixture model for binary outcome data in a case-control setting. We also compared the results to the effects with ordinary logistic regression. For these commonly employed alternative relative-risk models, the bias was similar to that with logistic regression when the risk was small. More generally, the bias and standard error of the risk-parameter estimates demonstrated patterns that are similar to those with logistic regression, but with greater magnitude depending on the true value of the risk. The magnitude of bias and standard error had little relation to study size or underlying disease prevalence. Prior conclusions regarding omitted covariates in logistic regression models can be qualitatively applied to the ERR and the general additive-multiplicative relative-risk mixture model without substantial change. Quantitatively, however, these alternative models may have slightly greater omitted-covariate bias, depending on the magnitude of the true risk being estimated.

60 APPLIED LIFE SCIENCES↗

Enhanced accuracy through ensembling of randomly initialized auto-regressive models for dynamical systems

Computational mechanics simulations using traditional finite element methods (FEM) require prohibitively expensive computational resources for real-time engineering applications, design optimization, and digital twin implementations. While machine learning (ML) surrogate models offer significant computational speedups, autoregressive ML models for time-dependent mechanical systems suffer from error accumulation that compromises long-term prediction reliability - a critical concern for engineering applications where accuracy over extended time horizons is essential for safety and performance assessments. Here, we propose a deep ensemble framework specifically designed to address this challenge in computational mechanics applications, where multiple ML surrogate models with random weight initializations are trained in parallel and their predictions aggregated during inference. This approach leverages statistical diversity to maximize information gain from a fixed set of training data and to mitigate error propagation, while maintaining the computational efficiency that makes ML surrogates attractive for engineering practice. We validate the framework on three representative problems spanning critical areas of computational mechanics: stress field evolution in heterogeneous microstructures under complex loading (relevant to advanced materials design and composite analysis), planetary-scale shallow water dynamics (applicable to environmental and geotechnical engineering), and Gray-Scott reaction-diffusion systems (relevant to mass transport and chemical process engineering). Across all test cases, the ensemble approach demonstrates consistent error reduction of 15-33% compared to individual models. The codes for this work are available on GitHub (https://github.com/Graham-Brady-Research-Group/AutoregressiveEnsemble_SpatioTemporal_Evolution).

autoregressive prediction↗

Generating Essential Climate Variables from Multiple Satellite Hyperspectral Remote Sensors

Hyperspectral observations from satellite-based sensors provide high information content for the Earth’s atmospheric and surface properties. Traditionally, long-term climate products are derived by performing spatial and temporal averaging of level-2 satellite products. There are two shortcomings of this approach. First, it is a time-consuming process to generate level-2 data products since modern hyperspectral satellite sensors have millions of observations each day with thousands of spectral channels for each observation. Secondly, differences in level-2 retrieval algorithms can lead to errors in the fused multi-satellite data. We have developed a radiometrically consistent spectral fingerprinting method, which overcomes the above-mentioned shortcomings, to derive climate change signals from multiple satellite sensors using spatiotemporally averaged level-1 data. We have applied this method to data collected from Atmospheric Infrared Sounder (AIRS) on Aqua satellite and Cross-track Infrared Sounder (CrIS) on SNPP and NOAA20 and generated decade-long climate data records for atmospheric temperature, water vapor, cloud, trace gases, and surface skin temperature. A key component to this work is a set of observational-based radiative kernels produced from CrIS level-1 data using a single field of view (SFOV) optimal estimation retrieval algorithm. Only limited CrIS level-1 data (e.g., 1-2 years of data) are needed to the derive radiative kernels. Our Principal Component-based Radiative Model (PCRTM) enables us to perform SFOV retrievals under all sky conditions and provides radiative kernels (including those for clouds) needed by the spectral fingerprinting method. In this presentation, we will describe the basic methodology, the details of the algorithm, and results from NASA Aqua AIRS and Suomi-NPP CrIS data. The method can be applied to study future hyperspectral remote sensors such as CLARREO (Climate Absolute Radiance and Refractivity Observatory) Pathfinder (CPF), Tropospheric Emissions: Monitoring of Pollution (TEMPO), Surface Biology and Geology (SBG), Atmosphere Observing System (AOS).

Xu Liu↗

HPC Network Simulation Tuning via Automatic Extraction of Hardware Parameters

Popular HPC network interconnection simulators such as SST/macro provide a variety of configurable parameters to explore the design space of hardware components such as network interface cards (NIC), switches, and links among them. While such knobs provide flexibility to explore design trade-offs for novel hardware, manually configuring simulations for matching configurations of the existing hardware to focus on topology exploration can be cumbersome and error-prone, leading to widely inaccurate simulations. This challenge is compounded when specifications of various (proprietary) technologies are not readily available or intentionally omitted. In this work, we propose a framework to autotune the multiple network models’ simulation configurations within SST/macro using Tree-structured Parzen Estimator-based Bayesian optimization to observe the effect on simulation accuracy across different message regimes. These regimes consist of small to large message sizes and latency to bandwidth-bound messages. We provide a detailed analysis of the simulation error for four representative HPC systems. Our Bayesian optimization based autotuning framework for network models achieves a maximum of 5x improvement in accuracy over best-effort manual configurations based on available hardware specifications.

Simulation, autotuning↗

Chemometric modeling of power conversion efficiency of organic dyes in dye sensitized solar cells for the future renewable energy

The present study reports chemometric modeling of power conversion efficiency (PCE) of dye sensitized solar cells (DSSCs) using the biggest available data set till date which comprises around 1200 dyes covering 7 chemical classes. To extract the best structural features required for higher PCE, we have developed multiple partial least squares (PLS) quantitative structure-property relationship (QSPR) models for the Triphenylamine, Phenothiazine, Indoline, Porphyrin, Coumarin, Carbazole and Diphenylamine chemical classes using descriptors derived from the best subset selection method followed by selection of best five models in each dataset based on the Mean Absolute Error (MAE) values. The models were validated both internally and externally followed by the consensus predictions employing “Intelligent Consensus Predictor” tool to examine whether the quality of predictions can be improved with the “intelligent” selection of multiple PLS models. The quality of predictions for the respective external sets showed that the consensus models (CM) are better than the individual models (IM) in most of the cases. From the insights of the developed models, we concluded that attributes like a packed structure toward higher conductivity of electrons, auxiliary donor fragment of aromatic tertiary amines, number of thiophenes inducing the bathochromic shift and augmenting the absorption, presence of additional electron donors, enhancement of electron-donating abilities, number of non-aromatic conjugated C(sp 2 ) which helps as conjugation extension units to broaden the absorption and highly conjugated $π$-systems exert positive contributions to the PCE. On the contrary, features negatively contributing to PCE are the followings: fragments which lower the tendency of localized $π–π*$ transition, fragments related to larger volume and surface area of dyes along with hydrophobicity resulting in poor adhesion, fragment RC = N causing dye hydrolysis, steric hindrance for π electronic mobility, fragments enhancing polarity, etc. Finally, the identified features from the best QSPR model of the coumarin dataset was employed in designing of ten more efficient coumarin dyes (predicted %PCE ranging from 8.93 to 10.62) than the existing ones.

14 SOLAR ENERGY↗

High-throughput characterization, correlation, and mapping of leaf photosynthetic and functional traits in the soybean ( Glycine max ) nested association mapping population

We report photosynthesis is a key target to improve crop production in many species including soybean [Glycine max (L.) Merr.]. A challenge is that phenotyping photosynthetic traits by traditional approaches is slow and destructive. There is proof-of-concept for leaf hyperspectral reflectance as a rapid method to model photosynthetic traits. However, the crucial step of demonstrating that hyperspectral approaches can be used to advance understanding of the genetic architecture of photosynthetic traits is untested. To address this challenge, we used full-range (500–2,400 nm) leaf reflectance spectroscopy to build partial least squares regression models to estimate leaf traits, including the rate-limiting processes of photosynthesis, maximum Rubisco carboxylation rate, and maximum electron transport. In total, 11 models were produced from a diverse population of soybean sampled over multiple field seasons to estimate photosynthetic parameters, chlorophyll content, leaf carbon and leaf nitrogen percentage, and specific leaf area (with R 2 from 0.56 to 0.96 and root mean square error approximately <10% of the range of calibration data). We explore the utility of these models by applying them to the soybean nested association mapping population, which showed variability in photosynthetic and leaf traits. Genetic mapping provided insights into the underlying genetic architecture of photosynthetic traits and potential improvement in soybean. Notably, the maximum Rubisco carboxylation rate mapped to a region of chromosome 19 containing genes encoding multiple small subunits of Rubisco. We also mapped the maximum electron transport rate to a region of chromosome 10 containing a fructose 1,6-bisphosphatase gene, encoding an important enzyme in the regeneration of ribulose 1,5-bisphosphate and the sucrose biosynthetic pathway. The estimated rate-limiting steps of photosynthesis were low or negatively correlated with yield suggesting that these traits are not influenced by the same genetic mechanisms and are not limiting yield in the soybean NAM population. Leaf carbon percentage, leaf nitrogen percentage, and specific leaf area showed strong correlations with yield and may be of interest in breeding programs as a proxy for yield. This work is among the first to use hyperspectral reflectance to model and map the genetic architecture of the rate-limiting steps of photosynthesis.

54 ENVIRONMENTAL SCIENCES↗

Correction of confidence intervals in excess relative risk models using Monte Carlo dosimetry systems with shared errors

Introduction: Measurement errors are ubiquitous in epidemiological studies, especially for doses arising from environmental and occupational exposures, which are difficult to assess. Correction methods for independent measurement errors, have been comprehensively described elsewhere, including regression calibration, simulation extrapolation (SIMEX), etc. However, when exposure estimates are constructed using complicated physical and biological models, uncertainties in the dosimetry system can become very complex. In such systems, uncertainties in some parameters may affect a large group of study participants simultaneously, and hence are “shared” as opposed to independent dosimetric uncertainties. Representing the structure of the uncertainty in the dosimetry by providing multiple realizations of possible dose has been described for several studies. The goal of this paper is to implement, and explore the properties of, a generalized estimating equation-based approach to shared uncertainties based on principles described in an earlier paper. Methods: The effects of shared and unshared dosimetry error on parameter estimates and confidence intervals in a generalized excess relative risk model are examined. A post-hoc correction method is implemented that constructs corrected confidence intervals based on an underlying variance structure described in an earlier paper. In this correction method it is assumed that the estimated slope parameter in a linear excess relative risk model is distributed as a mixture of normal and lognormal components. A simulation study of the behavior of this correction was developed based on multiple realizations provided by the Monte Carlo dosimetry system used for the plutonium exposures in the Mayak Worker Cohort and a risk model with close similarity to that used in analyses of lung cancer data in that cohort. Results: In the simulation study we found that shared dose errors severely degraded the coverage probabilities of confidence limits for the slope parameter in the linear excess relative risk model relating uncertain exposure to lung cancer risk. Using our correction method to adjust the confidence intervals provided much improved coverage probabilities for this parameter. A smaller amount of degradation in confidence interval coverage for other parameters (e.g. those in the baseline and dose modification terms) was noted, and the proposed correction method also improved coverage for these parameters as well. Discussion: Compared to other recently described methods our proposed correction method relies on relatively familiar software to fit dose response models. The basic modeling philosophy described here first uses existing Poisson regression methods to estimate parameters in the excess relative risk model; in this first stage mean dose (computed from the multiple realizations) is used as the independent variable in the regression. Once the model is fit we then correct (widen) the standard errors for all parameters using the proposed correction method. The simulation studies indicate that the methods described here are an important contribution to the available tools for dealing with shared measurement errors in complex risk analysis.

61 RADIATION PROTECTION AND DOSIMETRY↗

NuSTAR AND Swift Observations of the Very High State in GX 339-4: Weighing the Black Hole With X-Rays

We present results from spectral fitting of the very high state of GX339-4 with Nuclear Spectroscopic Telescope Array (NuSTAR) and Swift. We use relativistic reflection modeling to measure the spin of the black hole and inclination of the inner disk and find a spin of a = 0.95+0.08/-0.02 and inclination of 30deg +/- 1deg (statistical errors). These values agree well with previous results from reflection modeling. With the exceptional sensitivity of NuSTAR at the high-energy side of the disk spectrum, we are able to constrain multiple physical parameters simultaneously using continuum fitting. By using the constraints from reflection as input for the continuum fitting method, we invert the conventional fitting procedure to estimate the mass and distance of GX 339-4 using just the X-ray spectrum, finding a mass of 9.0+1.6/-1.2 Stellar Mass and distance of 8.4 +/- 0.9 kpc (statistical errors).

accretion↗

Decay curve analysis for data error quantification in time-domain induced polarization imaging

In recent years, the time-domain induced polarization (TDIP) imaging technique has emerged as a suitable method for the characterization and the monitoring of hydrogeologic and biogeochemical processes. However, one of the major challenges refers to the resolution of the electrical images. Hence, various studies have stressed the importance of data processing, error characterization, and the deployment of adequate inversion schemes. A widely accepted method to assess data error in electrical imaging relies on the analysis of the discrepancy between normal and reciprocal measurements. Nevertheless, the collection of reciprocals doubles the acquisition time and is only viable for a limited subset of commonly used electrode configurations (e.g., dipole-dipole [DD]). To overcome these limitations, we have developed a new methodology to quantify the data error in TDIP imaging, which is entirely based on the analysis of the recorded IP decay curve and does not require recollection of data (e.g., reciprocals). The first two steps of the methodology assess the general characteristics of the decay curves and the spatial consistency of the measurements for the detection and removal of outliers. In the third and fourth steps, we quantify the deviation of the measured decay curves from a smooth model for the estimation of random error of the total chargeability and transfer resistance measurement. The error models and imaging results obtained from this methodology - in the following referred to as "decay curve analysis" - are compared with those obtained following a conventional normal-reciprocal analysis revealing consistent results. As a result, we determine the applicability of our methodology with real field data collected at the floodplain scale (approximately 12 ha) using multiple gradient and DD configurations.

58 GEOSCIENCES↗

Application of Modified Meshgraphnets for Subsurface Prediction during CO2 Sequestration

In the face of the increasingly dire consequences of anthropogenic climate change, capturing and storing carbon dioxide is paramount. However, several impediments exist to the safe and effective subsurface storage of CO2, such as cost of transport, identification of suitable sites for subsurface storage, and assessment of long-term risk from storage in subsurface aquifers. Accurate subsurface modeling is necessary to ensure that CO2 storage is both safe and effective. Still, such modeling has traditionally required either substantial time and computational power (numerical simulation) or a substantial amount of pre-existing data for training (machine learning models). Additionally, these models lack flexibility in dealing with both changes in discretization of the input data and generalizability beyond the data on which they are trained. In order to address these issues, this research applies graph neural networks (GNNs) to predict subsurface saturation and pressure during CO₂ injection in a model of the Illinois Basin-Decatur Project (IBDP). GNNs provide a flexible, intuitive method for representing and manipulating complex unstructured data, which is often found in many practical domain problems such as fluid flow and subsurface characterization. These unstructured grids are easily represented in GNNs by representing spatially-localized features such as permeability, porosity, saturation, and pressure as nodes in a graph and relationships between these properties as edges connecting these nodes. This research applies a specific GNN model called MeshGraphNets (MGN) to model the change in CO2 saturation and pressure over a 50-month time period (36 months of injection, 14 months post-injection). The MGN model leverages a message passing process that allows the network to learn both the spatial and temporal dynamics of this system simultaneously. Additionally, training on a limited dataset (64 realizations, 20 time points each) resulted in a high degree of accuracy in saturation prediction both within the same timeframe as the training (20 months, 0.039 average RMSE) and when projecting out to the end of injection (36 months, 0.053 average RMSE). Temporal predictions such as those generated by MGNs and other similar models are prone to accumulated error over time; in order to address this, a multi-step rollout (MSR) training process was applied to calculate training loss. This method mimics the forward prediction during inference by “rolling out” multiple time points in a single training step using the previous prediction as input to the MGN model. By calculating the loss several time steps forward from the current prediction, the model is forced to find a more stable state over time. Application of MSR to the MGN model resulted in an average 15% reduction in inference error over time during forward prediction. This study showcases the immense potential of GNNs as a game-changing methodology for predicting pressure and saturation evolution in CCS projects, ultimately paving the way for more sustainable and effective carbon storage solutions. Presentation prepared for the 2024 AiChE Annual Meeting, October 27 to November 1 2024, San Diego, CA.

Holcomb, Paul↗