SEARCH · Search NASA
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.
This implementation of a Physics Guided Neural Network (PHYGNN) augments a traditional neural network loss function with a generic term that can be used to guide the neural network to learn physical or theoretical constraints.
Not Available
As neural networks are more frequently used to solve problems in science and engineering, the methods used to incorporate scientific knowledge into these networks are becoming increasingly complex. Here, this work breaks down these complicated techniques into a set of basic strategies which can easily be applied to diverse situations. Several novel neural networks are built using the categories laid out in this work. These networks are tested on simulated data from a continuous stirred tank reactor (CSTR) model to evaluate the advantages provided by each network. The three points demonstrated in this work are: (1) architectural hybrid models can speed up convergence and reduce the amount of data necessary to train a model; (2) adding a physics-guided loss function can improve model generalization and make models more physically consistent; (3) using physics-guided initialization and transfer learning improves accuracy and speeds up convergence, but can harm generalizability if used incorrectly.
Not Available
Abstract Advancements in computing power have recently made it possible to utilize machine learning and deep learning to push scientific computing forward in a range of disciplines, such as fluid mechanics, solid mechanics, materials science, etc. The incorporation of neural networks is particularly crucial in this hybridization process. Due to their intrinsic architecture, conventional neural networks cannot be successfully trained and scoped when data are sparse, which is the case in many scientific and engineering domains. Nonetheless, neural networks provide a solid foundation to respect physics-driven or knowledge-based constraints during training. Generally speaking, there are three distinct neural network frameworks to enforce the underlying physics: (i) physics-guided neural networks (PgNNs), (ii) physics-informed neural networks (PiNNs), and (iii) physics-encoded neural networks (PeNNs). These methods provide distinct advantages for accelerating the numerical modeling of complex multiscale multiphysics phenomena. In addition, the recent developments in neural operators (NOs) add another dimension to these new simulation paradigms, especially when the real-time prediction of complex multiphysics systems is required. All these models also come with their own unique drawbacks and limitations that call for further fundamental research. This study aims to present a review of the four neural network frameworks (i.e., PgNNs, PiNNs, PeNNs, and NOs) used in scientific computing research. The state-of-the-art architectures and their applications are reviewed, limitations are discussed, and future research opportunities are presented in terms of improving algorithms, considering causalities, expanding applications, and coupling scientific and deep learning solvers.
This submission includes example files associated with the Geothermal Operational Optimization using Machine Learning (GOOML) Big Kahuna fictional power plant, which uses synthetic data to model a fictional power plant. A forecast was produced using the GOOML data model framework and fictional input data, and a genetic optimization is included which determines optimal flash plant parameters. The inputs and outputs associated with the forecast and genetic optimization are included. The input and output files consist of data, configuration files, and plots. A link to the Physics-Guided Neural Networks (phygnn) GitHub repository is also included, which augments a traditional neural network loss function with a generic loss term that can be used to guide the neural network to learn physical or theoretical constraints. phygnn is used by the GOOML framework to help integrate its machine learning models into the relevant physics and engineering applications. Note that the data included in this submission are intended to provide a demonstration of GOOML's capabilities. Additional files that have not been released to the public are needed for users to run these models and reproduce these results. Units can be found in the readme data resource.
Addressing new methodologies in deep learning (DL), machine learning (ML) and artificial intelligence (AI), the webinar speakers will provide an overview of the literature spanning these three overlapping fields as applied to energy systems research. The audience will learn how developments in these areas have added new capabilities for pattern recognition and predictive modeling that are complementary to more traditional modeling approaches used in energy systems research. The speakers will illustrate several use cases of these new approaches in the energy space, such as physics-guided neural networks to improve ML regressions for solar data and how AI agents can be used to explore power plant operations.
With over 20 years of high-resolution surface irradiance data covering most of the western hemisphere, the National Solar Radiation Database (NSRDB) is a vital public data asset. The NSRDB uses a two-step Physical Solar Model (PSM) that explicitly considers the effects of clouds and other atmospheric variables on radiative transfer. High-quality physical and optical cloud properties derived from satellite imagery are perhaps the most important data inputs to the PSM, representing the greatest source of radiation attenuation and scattering. However, traditional methods for cloud property retrieval have their own limitations and are unable to accurately predict cloud properties outside of nominal conditions. We introduce a physics-guided neural network that can accurately predict cloud properties when traditional methods fail or are inaccurate. Using this framework, we show reductions in relative Root Mean Square Error (RMSE) for Global Horizontal Irradiance (GHI) up to 13 percentage points for timesteps that previously had missing or low-quality cloud property data. We expect that this methodology will be effective in improving the quality of cloud property and solar irradiance data in the NSRDB.
With over 20 years of high-resolution surface irradiance data covering most of the western hemisphere, the National Solar Radiation Database (NSRDB) is a vital public data asset. The NSRDB uses a two-step Physical Solar Model (PSM) that explicitly considers the effects of clouds and other atmospheric variables on radiative transfer. High-quality physical and optical cloud properties derived from satellite imagery are perhaps the most important data input to the PSM, representing the greatest source of radiation attenuation and scattering. However, traditional methods for cloud property retrieval have their own limitations and are unable to accurately predict cloud properties outside of nominal conditions. We introduce a physics-guided neural network that can accurately predict cloud properties when traditional methods fail or are inaccurate. Using this framework, we show reductions in relative Root Mean Square Error (RMSE) for Global Horizontal Irradiance (GHI) up to 13 percentage points for timesteps that previously had missing or low-quality cloud property data. We expect that this methodology will be effective in improving the quality of cloud property and solar irradiance data in the NSRDB.
Machine learning (ML) provides powerful pathways for predicting spectroscopic observables from atomic structures, but its broader impact depends on making model predictions interpretable in terms of physical and chemical principles. Here, we introduce a physics-guided graph neural network (GNN) model that predicts Zn K-edge X-ray spectroscopy (XAS) spectra of aqueous ZnCl 2 solutions. Training data are generated from ab initio XAS calculations on molecular dynamics snapshots obtained using a machine learning interatomic potential. The GNN reproduces experimental spectra across concentrations from dilute (<0.1 m) to highly concentrated (30 m, “water-in-salt”) regimes and scales efficiently to large, disordered liquid systems beyond the reach of conventional ab initio approaches. Gradient-based attribution analysis reveals that the model learns physically meaningful structure-spectrum relationships. Ligand-specific attributions reflect orbital hybridization patterns and the origin of the excitations derived from the density functional theory. Bond-length attributions recover spectral shifts consistent with multiple-scattering theory. Finally, this work bridges data-driven prediction with electronic-structure theory, establishing a general paradigm for interpretable ML that links atomic structure, electronic structure, and spectroscopic observables.
Abstract. Rivers and river habitats around the world are under sustained pressure from human activities and the changing global environment. Our ability to quantify and manage the river states in a timely manner is critical for protecting the public safety and natural resources. In recent years, vector-based river network models have enabled modeling of large river basins at increasingly fine resolutions, but are computationally demanding. This work presents a multistage, physics-guided, graph neural network (GNN) approach for basin-scale river network learning and streamflow forecasting. During training, we train a GNN model to approximate outputs of a high-resolution vector-based river network model; we then fine-tune the pretrained GNN model with streamflow observations. We further apply a graph-based, data-fusion step to correct prediction biases. The GNN-based framework is first demonstrated over a snow-dominated watershed in the western United States. A series of experiments are performed to test different training and imputation strategies. Results show that the trained GNN model can effectively serve as a surrogate of the process-based model with high accuracy, with median Kling–Gupta efficiency (KGE) greater than 0.97. Application of the graph-based data fusion further reduces mismatch between the GNN model and observations, with as much as 50 % KGE improvement over some cross-validation gages. To improve scalability, a graph-coarsening procedure is introduced and is demonstrated over a much larger basin. Results show that graph coarsening achieves comparable prediction skills at only a fraction of training cost, thus providing important insights into the degree of physical realism needed for developing large-scale GNN-based river network models.
The National Solar Radiation Database (NSRDB) provides high-resolution spatiotemporal solar irradiance data for the entire globe. The NSRDB uses a two-step Physical Solar Model (PSM) to compute the effects of clouds and other atmospheric variables on the solar radiation reaching the surface of the Earth. Physical and optical cloud properties are fundamental inputs to the PSM and are derived from the National Oceanic and Atmospheric Administration's Geostationary Operational Environmental Satellites. This paper describes recent improvements to the NSRDB driven by physics-guided machine learning methods for cloud property retrieval. The impacts of these new methods on the NSRDB irradiance data are validated using an extensive set of ground measurement sites, showing significant improvement for all sites. We report on average, the mean absolute percentage error for global horizontal irradiance and direct normal irradiance show reductions of 2.16 and 3.95 percentage points respectively for all daylight conditions, 5.92 and 17.39 percentage points respectively for cloudy conditions, and 9.00 and 22.59 percentage points respectively for gap-filled cloudy conditions. These new methods will help improve the quality and accuracy of the irradiance and cloud data in the NSRDB.
One of the biggest challenges to achieve the goal of producing fusion energy in tokamak devices is the necessity of avoiding disruptions of the plasma current due to instabilities. The disruption event characterization and forecasting (DECAF) framework has been developed in this purpose, integrating physics models of many causal events that can lead to a disruption. Two different machine learning approaches are proposed to improve the ideal magnetohydrodynamic (MHD) no-wall limit component of the kinetic stability model included in DECAF. First, a random forest regressor (RFR), was adopted to reproduce the DCON computed change in plasma potential energy without wall effects for a large database of equilibria from the national spherical torus experiment (NSTX). This tree-based method provides an analysis of the importance of each input feature, giving an insight into the underlying physics phenomena. Secondly, a fully-connected neural network has been trained on sets of calculations with the DCON code, to get an improved closed form equation of the no-wall β limit as a function of the relevant plasma parameters indicated by the RFR. The neural network has been guided by physics theory of ideal MHD in its extension outside the domain of the NSTX experimental data. The estimated value has been incorporated into the DECAF kinetic stability model and tested against a set of experimentally stable and unstable discharges. Moreover, the neural network results were used to simulate a real-time stability assessment using only quantities available in real-time. Finally, the portability of the model was investigated, showing encouraging results by testing the NSTX-trained algorithm on the mega ampere spherical tokamak (MAST).
Artificial neural networks offer a highly nonlinear and adaptive model for predicting complex interactions between input-output parameters. However, these networks require large datasets which often exceed practical considerations in modeling experimental results. To alleviate the dataset size requirement, a method known as physics guided machine learning has been applied to construct several neural networks for predicting propeller tonal noise in the time domain over a broad range of flight conditions. Three space-filling designs, namely, Latin-Hypercube, Sphere-Packing, and Grid-Space, were used to distribute points throughout the input parameter space encompassing nondimensional flight conditions and observer geometry. Each neural network’s performance was validated by conditions outside of the training set and compared to the Propeller Analysis System tool from the NASA Aircraft Noise Prediction Program. Compared to the Grid-Space input design, the Latin-Hypercube and the Sphere-Packing designs provided a better representation of the domain for training. Regarding the network archetype, a fully connected perceptron was found to outperform the partially connected perceptron in their ability to predict tonal noise for small datasets. The black-box nature of these neural networks was also explored to understand how the networks constructed the waveform and understand why some network designs produce better models.
Observations of atmospheric 3D winds are critical for improving short-range and long-range forecasting. Such advancement in forecasting directly applies to research in a number of areas including convective processes, wildfire plumes and tornado prediction. Atmospheric Motion Vectors (AMVs) provide a passively sensed approach to quantifying horizontal motion and cloud heights, which are typically sourced from geostationary sensors due to the availability of high frequency observations. Recent work has shown that estimating AMVs by tracking individual pixels with dense optical flow is a promising new direction. In this work, we compare state-of-the-art convolutional neural networks for optical flow in a physics-guided deep learning framework for predicting AMVs. The approach is semi-supervised and uses physically informed wind vectors from high-resolution numerical simulations for supervised learning followed by fine-tuning though warping and reconstruction of full-disk geostationary images (GOES-16). In the vertical direction, we use labels from the CALIPSO low-earth orbit satellite to predict cloud height from 16-band geostationary images with a neural network. We present results for both tasks on held-out time periods and secondary datasets.
Atmospheric 3D winds in the horizontal and vertical directions are critical for improving short-range and long-range forecasting. Such advancement in forecasting directly applies to research in a number of areas including convective processes, wildfire plumes and tornado prediction. Atmospheric Motion Vectors (AMVs) provide a passively sensed approach to quantifying horizontal motion and cloud heights, which are typically sourced from geostationary sensors due to the availability of high frequency observations. Recent work has shown that estimating AMVs by tracking individual pixels with dense optical flow is a promising new direction. In this work, we use a state-of-the-art convolutional neural network for optical flow (FlowNetS) in a physics-guided deep learning framework for predicting AMVs in the horizontal direction. The approach is semi-supervised and uses physically informed wind vectors from high-resolution numerical simulations (DYAMOND) for supervised learning followed by fine-tuning though warping and reconstruction of full-disk geostationary images (GOES-16). In the vertical direction, we use labels from the CALIPSO low-earth orbit satellite to predict cloud height from 16-band geostationary images with a neural network. We present results for both tasks on held-out time periods and secondary datasets.
In the past decade, dramatic progress has been made in the field of machine learning. This paper explores the possibility of applying deep learning in power system state estimation. Traditionally, physics-based models are used including weighted least square (WLS) or weighted least absolute value (WLAV). These models typically consider a single snapshot of the system without capturing temporal correlations of system states. In this paper, a physics-guided deep learning (PGDL) method is proposed. Specifically, inspired by autoencoders, deep neural networks (DNNs) are used to learn the temporal correlations. The estimated system states from DNNs are then checked against physics laws by running through a set of power flow equations. Hence, the proposed PGDL is both data-driven and physics-guided. The accuracy and robustness of the proposed PGDL method are compared with traditional methods in standard IEEE cases. Simulations show promising results and the applicability is further discussed.
Physics informed neural networks (PINNs) are deep learning based techniques for solving partial differential equations (PDEs) encountered in computational science and engineering. Guided by data and physical laws, PINNs find a neural network that approximates the solution to a system of PDEs. Such a neural network is obtained by minimizing a loss function in which any prior knowledge of PDEs and data are encoded. Despite its remarkable empirical success in one, two or three dimensional problems, there is little theoretical justification for PINNs. As the number of data grows, PINNs generate a sequence of minimizers which correspond to a sequence of neural networks. We want to answer the question: Does the sequence of minimizers converge to the solution to the PDE? We consider two classes of PDEs: linear second-order elliptic and parabolic. By adapting the Schauder approach and the maximum principle, we show that the sequence of minimizers strongly converges to the PDE solution in C 0 . Furthermore, we show that if each minimizer satisfies the initial/boundary conditions, the convergence mode becomes H 1 . Computational examples are provided to illustrate our theoretical findings. To the best of our knowledge, this is the first theoretical work that shows the consistency of PINNs.