Engineering Papers⌕ Search

SEARCH · Engineering Papers

Results for “generalizability”

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 73 records · Page 4
All sourcesNASA NTRSDOE OSTI
DOE OSTI2024

Implications of stop-and-go traffic on training learning-based car-following control

Learning-based car-following control (LCC) of connected and autonomous vehicles (CAVs) is gaining significant attention with the advancement of computing power and data accessibility. While the flexibility and large model capacity of model-free architecture enable LCC to potentially outperform the model-based car-following (CF) model in improving traffic efficiency and mitigating congestion, the generalizability of LCC for traffic conditions different from the training environment/dataset is not well-understood. Herein, this study seeks to explore the impact of stop-and-go traffic in the training dataset on the generalizability of LCC. It uses the characteristics of lead vehicle trajectories to describe stop-and-go traffic, and links the theory of identifiability (i.e., obtaining a unique parameter estimation result using sensor measurements) to the generalizability of behavior cloning (BC) and policy-based deep reinforcement learning (DRL). Correspondingly, the study shows theoretically that: (i) stop-and-go traffic can enable the property of identifiability and enhance the control performance of BC-based LCC in different traffic conditions; (ii) stop-and-go traffic is not necessary for DRL-based LCC to generalize to different traffic conditions; (iii) DRL-based LCC trained with only constant-speed lead vehicle trajectories (not sufficient to ensure identifiability) can be generalized to different traffic conditions; and (iv) stop-and-go traffic increases variance in the training dataset, which improves the convergence of parameter estimation while negatively impacting the convergence of DRL to the optimal control policy. Numerical experiments validate the above findings, illustrating that BC-based LCC entails comprehensive training datasets for generalizing to different traffic conditions, while DRL-based LCC can achieve generalization with simple free-flow traffic training environments. This further suggests DRL as a more promising and cost-effective LCC approach to reduce operational costs, mitigate traffic congestion, and enhance safety and mobility, which can accelerate the deployment and acceptance of CAVs.

33 ADVANCED PROPULSION SYSTEMS↗
DOE OSTI2023

Say where you sample: Increasing site selection transparency in urban ecology

Urban ecological studies have the potential to extend our understanding of socio-ecological systems beyond that of an individual city or region. Cross-comparative empirical work and synthesis are imperative to develop a general urban ecological theory. This can be achieved only if studies are replicable and generalizable. Transparency in methods reporting facilitates generalizability and replicability by documenting the decisions scientists make during the various steps of research design; this is particularly true for sampling design and selection because of their impact on both internal and external validity and the potential to unintentionally introduce bias. Three interdependent aspects of sample design are study sample selection (e.g., specific organisms, soils, or water), sample specification (measurement of specific variable of interest), and site selection (locations sampled). Of these, documentation of site selection—the where component of sample design—is underrepresented in the urban ecology literature. Using a stratified random sample of 158 papers from 12 major urban ecology journals, we investigated how researchers selected study sites in urban ecosystems and evaluated whether their site selection methods were transparent. We extracted data from these papers using a 50-question, theory-based questionnaire and a multiple-reviewer approach. Our sample represented almost 45 years of urban ecology research across 40 different countries. We found that more than 80% of the papers read were not transparent in their site selection methodology. We do not believe site selection methods are replicable for 70% of the papers read. Key weaknesses include incomplete descriptions of populations and sampling frames, urban gradients, sample selection methods, and property access. Low transparency in reporting the where methodology limits urban ecologists’ ability to assess the internal and external validity of studies’ findings and to replicate published studies; it also limits the generalizability of existing studies. The challenges of low transparency are particularly relevant in urban ecology, a field where standard protocols for site selection and delineation are still being developed. These limitations interfere with the fields’ ability to build theory and inform policy. We conclude by offering a set of recommendations to increase transparency, replicability, and generalizability.

54 ENVIRONMENTAL SCIENCES↗
DOE OSTI2021

A cross-study analysis of drug response prediction in cancer cell lines

Abstract To enable personalized cancer treatment, machine learning models have been developed to predict drug response as a function of tumor and drug features. However, most algorithm development efforts have relied on cross-validation within a single study to assess model accuracy. While an essential first step, cross-validation within a biological data set typically provides an overly optimistic estimate of the prediction performance on independent test sets. To provide a more rigorous assessment of model generalizability between different studies, we use machine learning to analyze five publicly available cell line-based data sets: National Cancer Institute 60, ancer Therapeutics Response Portal (CTRP), Genomics of Drug Sensitivity in Cancer, Cancer Cell Line Encyclopedia and Genentech Cell Line Screening Initiative (gCSI). Based on observed experimental variability across studies, we explore estimates of prediction upper bounds. We report performance results of a variety of machine learning models, with a multitasking deep neural network achieving the best cross-study generalizability. By multiple measures, models trained on CTRP yield the most accurate predictions on the remaining testing data, and gCSI is the most predictable among the cell line data sets included in this study. With these experiments and further simulations on partial data, two lessons emerge: (1) differences in viability assays can limit model generalizability across studies and (2) drug diversity, more than tumor diversity, is crucial for raising model generalizability in preclinical screening.

59 BASIC BIOLOGICAL SCIENCES↗
DOE OSTI2024

GraMeR: Gra ph Me ta R einforcement learning for multi-objective influence maximization

Influence maximization (IM) is a combinatorial problem of identifying a subset of seed nodes in a network (graph), which when activated, provide a maximal spread of influence in the network for a given diffusion model and a budget for seed set size. IM has numerous applications such as viral marketing, epidemic control, sensor placement and other network-related tasks. However, its practical uses are limited due to the computational complexity of current algorithms. Recently, deep reinforcement learning has been leveraged to solve IM in order to ease the computational burden. However, there are serious limitations in current approaches, including narrow IM formulation that only consider influence via spread and ignore self-activation, low scalability to large graphs, and lack of generalizability across graph families leading to a large running time for every test network. In this work, we address these limitations through a unique approach that involves: (1) Formulating a generic IM problem as a Markov decision process that handles both intrinsic and influence activations; (2)incorporating generalizability via meta-learning across graph families. There are previous works that combine deep reinforcement learning with graph neural network, but this work solves a more realistic IM problem and incorporates generalizability across graphs via meta reinforcement learning. Extensive experiments are carried out in various standard networks to validate performance of the proposed Graph Meta Reinforcement learning (GraMeR) framework. Finally, the results indicate that GraMeR is multiple orders faster and generic than conventional approaches when applied on small to medium scale graphs.

97 MATHEMATICS AND COMPUTING↗
DOE OSTI2024

Chemistry Informed Machine Learning-Based Heat Capacity Prediction of Solid Mixed Oxides

Knowing heat capacity is crucial for modeling temperature changes with the absorption and release of heat and for calculating the thermal energy storage capacity of oxide mixtures with energy applications. The current prediction methods (ab initio simulations, computational thermodynamics, and the Neumann–Kopp rule) are computationally expensive, not fully generalizable, or inaccurate. Machine learning has the potential of being fast, accurate, and generalizable, but it has been scarcely used to predict mixture properties, particularly for mixed oxides. Here, we demonstrate a method for the generalizable prediction of heat capacity of solid oxide pseudobinary mixtures using heat capacity data obtained from computational thermodynamics and descriptors from ab initio databases. Further, models trained through this workflow achieved an error (mean absolute error of 0.43 J mol –1 K –1 ) lower than the uncertainty in differential scanning calorimetry measurements, and the workflow can be extended to predict other properties derived from the Gibbs free energy and for higher-order oxide mixtures.

37 INORGANIC, ORGANIC, PHYSICAL, AND ANALYTICAL CH↗
DOE OSTI2021

Model fusion with physics-guided machine learning: Projection-based reduced-order modeling

The unprecedented amount of data generated from experiments, field observations, and large-scale numerical simulations at a wide range of spatiotemporal scales has enabled the rapid advancement of data-driven and especially deep learning models in the field of fluid mechanics. Although these methods are proven successful for many applications, there is a grand challenge of improving their generalizability. This is particularly essential when data-driven models are employed within outer-loop applications like optimization. In this work, we put forth a physics-guided machine learning (PGML) framework that leverages the interpretable physics-based model with a deep learning model. Leveraging a concatenated neural network design from multi-modal data sources, the PGML framework is capable of enhancing the generalizability of data-driven models and effectively protects against or inform about the inaccurate predictions resulting from extrapolation. We apply the PGML framework as a novel model fusion approach combining the physics-based Galerkin projection model and long- to short-term memory (LSTM) network for parametric model order reduction of fluid flows. We demonstrate the improved generalizability of the PGML framework against a purely data-driven approach through the injection of physics features into intermediate LSTM layers. Our quantitative analysis shows that the overall model uncertainty can be reduced through the PGML approach, especially for test data coming from a distribution different than the training data. Moreover, we demonstrate that our approach can be used as an inverse diagnostic tool providing a confidence score associated with models and observations. The proposed framework also allows for multi-fidelity computing by making use of low-fidelity models in the online deployment of quantified data-driven models.

42 ENGINEERING↗
DOE OSTI2026

Ecological Insights from Transferable Plant Biomass Mapping across the Arctic using High-resolution Structure-from-Motion and LiDAR Data

Warmer temperatures, permafrost thaw, and increased wildfire activity are driving rapid ecological change across the Arctic, significantly altering plant productivity and aboveground biomass (AGB). These rapid changes highlight the urgent need to improve monitoring of vegetation dynamics in the Earth’s northern ecosystems, where high spatiotemporal heterogeneity occurs at scales finer than those captured by traditional satellite observations. The growing use of Unoccupied Aerial Systems (UASs) presents an opportunity to overcome this limitation. Yet, the diversity of UAS platforms, sensors, and data collection and processing workflows presents challenges for developing standardized, generalizable approaches. To address this challenge, we compiled 672 AGB plots co-located with 183 UAS-based Structure-from-Motion (SfM) or Light Detection and Ranging (LiDAR) surveys collected across the Arctic. Here, we: (1) evaluated the generalizability of UAS-derived canopy structure derived from high-resolution SfM and LiDAR for estimating AGB, (2) assessed scaling errors and their sources in two recent satellite-based AGB products derived from Landsat and MODIS, and (3) demonstrated the use of high-resolution AGB maps to quantify biomass variation across tundra plant functional types (PFTs) and to monitor post-fire recovery. Our results show that both SfM and LiDAR accurately captured AGB and its variability across tundra PFTs using a Random Forest (RF) model (overall RMSE: 0.336 kg/m2), with mapping performance varying slightly by region and data source. Using UAS-derived AGB maps as a benchmark, we identified systematic biases in satellite-derived AGB products, largely attributable to the magnitude of AGB and structural heterogeneity within coarse-resolution pixels. Applying our model to repeat UAS surveys following a tundra fire on Seward Peninsula, we observed rapid AGB recovery in non-shrub patches, with biomass recovering to pre-fire levels within 2 years. In contrast, shrub patches recovered more slowly, with AGB gains continuing over 2–4 years through both in-patch growth and lateral expansion (via dispersal) into remaining burned areas. Overall, these findings demonstrate the generalizability of UAS-based SfM and LiDAR data for estimating tundra AGB and highlight the potential of our approach to be broadly applied to generate high-quality AGB data for ecological monitoring and model benchmarking across the Arctic.

Yang, Daryl [ORNL] (ORCID:0000000317057823)↗
DOE OSTI2024

Invariant Features for Accurate Predictions of Quantum Chemical UV-vis Spectra of Organic Molecules

Including invariance of global properties of a phys-ical system as an intrinsic feature in graph neural networks (GNNs) enhances the model's robustness and generalizability and reduces the amount of training data required to obtain a desired accuracy for predictions of these properties. Existing open source GNN libraries construct invariant features only for specific GNN architectures. This precludes the generalization of invariant features to arbitrary message passing neural network (MPNN) layers which, in turn, precludes the use of these libraries for new, user-specified predictive tasks. To address this limitation, we implement invariant MPNNs into the flexible and scalable HydraGNN architecture. HydraGNN enables a seamless switch between various MPNNs in a unified layer sequence and allows for a fair comparison between the predictive performance of different MPNNs. We trained this enhanced HydraGNN archi-tecture on the ultraviolet-visible (UV-vis) spectrum of GDB-9 molecules, a feature that describes the molecule's electronic exci-tation modes, computed with time-dependent density functional tight binding (TD-DFTB) and available open source through the GDB-9-Ex dataset. We assess the robustness (i.e., accuracy and generalizability) of the predictions obtained using different invariant MPNNs with respect to different values of the full width at half maximum (FWHM) for the Gaussian smearing of the theoretical peaks. Our numerical results show that incorporating invariance in the HydraGNN architecture significantly enhances both accuracy and generalizability in predicting UV-vis spectra of organic molecules.

Baker, Justin↗
NASA NTRS2019

Training for Long-Duration Space Missions: A Literature Review into Skill Retention and Generalizatility

In long-duration space missions, skill retention and generalizability become ever more important as mission length increases, for it is through these capabilities that astronaut crews achieve autonomy. A framework is proposed for future research based on three types of training commonly found in operational scenarios: initial training, refresher training, and just-in-time training. One of the key points of interest identified from the body of literature is the need for a structured and quantifiable approach to characterize skill decay. Such an approach would allow gaining a deeper understanding of the mechanisms through which skill decay operates. Because simulators are used extensively in all types of training, the effects of simulator fidelity on skill retention and generalizability are paramount to understand. Some common fidelity types are defined and their effects on the human operator are discussed. From a review of the literature on all these topics, several research gaps were identified that led to three key opportunities for future research: (1) developing skill decay functions provides theoretical insights into skill decay and allows for several practical applications, such as planning refresher training, (2) the effects of simulator fidelity on skill decay functions should be investigated empirically, and (3) investigating the generalizability of skills learned in initial training to other tasks could provide space crews with greater autonomy. A cybernetic approach might give new insights.

Pieters, Marc A.↗
DOE OSTI2024

Establishing performance metrics for quantitative non-targeted analysis: a demonstration using per- and polyfluoroalkyl substances

Abstract Non-targeted analysis (NTA) is an increasingly popular technique for characterizing undefined chemical analytes. Generating quantitative NTA (qNTA) concentration estimates requires the use of training data from calibration “surrogates,” which can yield diminished predictive performance relative to targeted analysis. To evaluate performance differences between targeted and qNTA approaches, we defined new metrics that convey predictive accuracy, uncertainty (using 95% inverse confidence intervals), and reliability (the extent to which confidence intervals contain true values). We calculated and examined these newly defined metrics across five quantitative approaches applied to a mixture of 29 per- and polyfluoroalkyl substances (PFAS). The quantitative approaches spanned a traditional targeted design using chemical-specific calibration curves to a generalizable qNTA design using bootstrap-sampled calibration values from “global” chemical surrogates. As expected, the targeted approaches performed best, with major benefits realized from matched calibration curves and internal standard correction. In comparison to the benchmark targeted approach, the most generalizable qNTA approach (using “global” surrogates) showed a decrease in accuracy by a factor of ~4, an increase in uncertainty by a factor of ~1000, and a decrease in reliability by ~5%, on average. Using “expert-selected” surrogates ( n = 3) instead of “global” surrogates ( n = 25) for qNTA yielded improvements in predictive accuracy (by ~1.5×) and uncertainty (by ~70×) but at the cost of further-reduced reliability (by ~5%). Overall, our results illustrate the utility of qNTA approaches for a subclass of emerging contaminants and present a framework on which to develop new approaches for more complex use cases. Graphical Abstract

Pu, Shirley (ORCID:0000000201223797)↗
DOE OSTI2023

Building thermal dynamics modeling with deep transfer learning using a large residential smart thermostat dataset

Understanding thermal dynamics and obtaining the computational model of residential buildings enable its scaled application in energy retrofits, control optimization and decarbonization. In this paper, we present a deep learning approach to model building thermal dynamics with smart thermostat data collected from residential buildings, with the goal to investigate model generalizability. In the first stage, we developed and compared different Deep Learning architectures including Convolutional Neural Networks (CNN), Long Short-Term Memory (LSTM) models and CNN-LSTM to predict indoor air temperature in a multi-step time horizon. In the second stage, we implemented a Transfer Learning (TL) process, which aims to improve the prediction performance on a new set of buildings (targets), exploiting the knowledge of related or similar buildings (sources). Different TL strategies and source model identification methods were investigated. The study showed that the CNN-LSTM performed the best among the architectures compared, with an average Mean Absolute Error (MAE) of 0.26 °C for one-hour-ahead (twelve 5-min future steps) predictions. Furthermore, the results showed that freezing the LSTM layer and fine-tuning the other layers of the CNN-LSTM achieved the best performance among four TL strategies, which further improved the performance with respect to a machine learning approach by 10%, and proving the effectiveness and generalizability of the proposed approach. A comparison of three different source model identification methods showed that randomly selecting source models constrained by similar building characteristics can provide good TL performance while retaining simplicity comparing with other quantitative source identification methods.

32 ENERGY CONSERVATION, CONSUMPTION, AND UTILIZATI↗
DOE OSTI2022

Route Optimization for Energy Efficient Airport Shuttle Operations - A Case Study from Dallas Fort Worth International Airport

Air travel and requisite surface traffic supporting passenger arrival/departure constitutes a significant portion of travel and emissions in cities with large airports. An airport trip can segment into three parts namely: i) travel from a location in the city to the airport; ii) travel from a parking lot or rental car center to the terminal (i.e., within the airport premises), and iii) travel inside the terminal. Depending on the airport access mode all or a part of these legs comprise a traveler’s journey to the airport. The priority of airport ground transport management teams is to provide passengers with a seamless travel experience within the airport, so it is understandable that within airport shuttle routes might not be optimized for minimizing energy consumption. Solutions that meet the dual objective of reducing energy consumption from airport shuttle operations without compromising on passenger travel experience are key to improving system efficiency. There is currently a dearth of research and tools that can inform airports in making such decisions. Addressing this need, this research effort puts forth an optimization model that generates optimal shuttle routes for a given set of constraints, and a discrete-event simulator that evaluates the optimal solutions in a stochastic environment to understand the tradeoffs between passenger wait times, and within airport shuttle energy consumption. The proposed set of tools are tested in the context of optimizing airport shuttles routes within the Dallas Fort Worth International Airport (DFW). In addition to shuttle spatial positioning, and passenger demand information, high-fidelity vehicle data was collected using data loggers installed on DFW shuttles. Results show that 20% energy reduction in shuttle operations is possible with a modest two-minute increase in average passenger wait times. The tools developed in this research effort are designed to be generalizable and can help optimize shuttle operations planning at any major airport.

air travel↗
DOE OSTI2021

A physics-informed and hierarchically regularized data-driven model for predicting fluid flow through porous media

This paper presents a new deep learning data-driven model for predicting structure dependent pore-fluid velocity fields in rock. The model is based on a Convolutional Auto-Encoder (CAE) artificial neural network capable of learning from image data generated by direct numerical simulations of fluid flow through pore-structures, such as by Lattice Boltzmann or molecular dynamics methods. The main novelty of the model in comparison to previous CAE-based data-driven approaches consists of three parts. The first is a methodology for decomposing the full-domain of the porous media into sub-regions, or “sub-domains”, in order to reduce the overall size of the CAE, batch process the sub-domains in parallel, and enable the CAE to learn local and generalizable nonlinear mappings of pore-fluid velocities. The second consists of embedding the finite difference solutions of the incompressible Navier-Stokes and continuity equations into convolutional layers prior to the CAE in order to provide the CAE with knowledge of fluid dynamics physics (PhyFlow). The third main novelty is that the training of the CAE is regularized with a hierarchical loss function that encourages the learning of fluid flow patterns (in a way similar to ranked modes in principal component analysis), ranking from most to least important. This is shown to increase the stability in learning, reduce over-fitting, and promote interpretability of the CAE neural network layers (HierCAE). The comprehensive new data-driven model, which we call the PhyFlow-HierCAE model, is shown to exhibit improved accuracy and generalizability of flow field predictions over conventional CAE models, attributable to the embedded physical knowledge and the hierarchical regularization, as well as realize orders of magnitude speed-ups in computation times as a surrogate for the direct numerical simulations. Examples of training and forward predictions on unseen pore-structures are provided and evaluated for data from Lattice Boltzmann and molecular dynamics simulations of pore-fluid flow. The model is shown to be a fast and accurate emulator (or “surrogate”) for predicting effective permeability of unseen pore-structures based on learning from relatively small direct numerical simulation datasets.

71 CLASSICAL AND QUANTUM MECHANICS, GENERAL PHYSIC↗
DOE OSTI2022

A weighted state redistribution algorithm for embedded boundary grids

State redistribution is an algorithm that stabilizes cut cells for embedded boundary grid methods. This work extends the earlier algorithm in several important ways. First, state redistribution is extended to three spatial dimensions. Second, we discuss several algorithmic changes and improvements motivated by the more complicated cut cell geometries that can occur in higher dimensions. In particular, we introduce a weighted version with less dissipation in an easily generalizable framework. Third, we demonstrate that state redistribution can also stabilize a solution update that includes both advective and diffusive contributions. Notably, the stabilization algorithm is shown to be effective for incompressible as well as compressible reacting flows. Finally, we discuss the implementation of the algorithm for several exascale-ready simulation codes based on AMReX, demonstrating ease of use in combination with domain decomposition, hybrid parallelism and complex physics.

97 MATHEMATICS AND COMPUTING↗
DOE OSTI2023

Mapping foliar photosynthetic capacity in sub-tropical and tropical forests with UAS-based imaging spectroscopy: Scaling from leaf to canopy

Accurate understanding of the variability in foliar physiological traits across landscapes is critical to improve parameterization and evaluation of terrestrial biosphere models (TBMs) that seek to represent the response of terrestrial ecosystems to a changing climate. Numerous studies suggest imaging spectroscopy can characterize foliar biochemical and morphological traits at the canopy scale, but there is only limited evidence for retrieving canopy photosynthetic capacity (e.g., maximum carboxylation rate, V c,max and maximum electron transport rate, J max ). Moreover, the effect of canopy structure within forest communities on scaling up spectra-trait relationships from leaf to canopy level is not well known. To advance the spectra-trait approach and enable the estimation of key traits using remote sensing, we collected imaging spectroscopy data from an Unoccupied Aerial System (UAS) platform over two forest sites in China (a subtropical forest in Mt. Dinghu and a tropical rainforest in Xishuangbanna). At these sites, we also collected ground measurements of leaf spectra and traits, including biochemical (leaf nitrogen, phosphorus, chlorophyll, and water content), morphological (leaf mass per area, LMA) and physiological (V c,max25 and J max25 ) traits (n=135 tree-crowns from 42 species across two sites). Using a partial least-squares regression (PLSR) approach, we built and tested spectra-trait models with repeated cross-validation. The spectral models developed with leaf spectra were directly transferred to canopy spectra to evaluate the effect of canopy structure. Here we further applied canopy spectral models to map these traits at individual tree-crown scale. The results demonstrate that (1) UAS-based canopy spectra can be used to estimate V c,max (R 2 =0.55, nRMSE=11.79%), Jmax (R 2 =0.54, nRMSE=12.34%), and five additional foliar traits (R 2 =0.38-0.60, nRMSE=10.11-13.56%) at the tree-crown scale with demonstrated generalizability across two sites; (2) canopy structure strongly affects the spectratrait relationships from leaf to canopy level, but the effects vary considerably across foliar traits and cannot be well captured by the 4SAIL canopy radiative transfer model. UAS-based imaging spectroscopy maps large variability in all foliar traits (including physiological traits) with spatially explicit information, reproducing the field-observed inter- and intra-specific variations. These results demonstrate the capability of using UAS-based imaging spectroscopy for characterizing the variability of foliar physiological traits at individual tree-crown scale over forest landscapes and highlight the similar generalizability but different biophysical mechanisms underlying spectra-trait relationships at leaf and canopy levels.

54 ENVIRONMENTAL SCIENCES↗
DOE OSTI2021

Assessing Decadal Predictability in an Earth-System Model Using Explainable Neural Networks

We show that explainable neural networks can identify regions of oceanic variability that contribute predictability on decadal timescales in a fully coupled Earth-system model. The neural networks learn to use sea-surface temperature anomalies to predict future continental surface temperature anomalies. We then use a neural-network explainability method called layerwise relevance propagation to infer which oceanic patterns lead to accurate predictions made by the neural networks. In particular, regions within the North Atlantic Ocean and North Pacific Ocean lend the most predictability for surface temperature across continental North America. We apply the proposed methodology to decadal variability, although the concept is generalizable to other timescales of predictability. Furthermore, while our approach focuses on predictable patterns of internal variability within climate models, it should be generalizable to observational data as well. Our study contributes to the growing evidence that explainable neural networks are important tools for advancing geoscientific knowledge.

58 GEOSCIENCES↗
DOE OSTI2023

Stream Temperature Prediction in a Shifting Environment: Explaining the Influence of Deep Learning Architecture

Stream temperature is a fundamental control on ecosystem health. Recent efforts incorporating process guidance into deep learning models for predicting stream temperature have been shown to outperform existing statistical and physical models. This performance is in part because deep learning architectures can actively learn spatiotemporal relationships that govern how water and energy propagate through a river network. However, exploration of how spatiotemporal awareness and process guidance influence a model's generalizability under shifting environmental conditions such as climate change is limited. Here, we use Explainable Artificial Intelligence (XAI) to interrogate how differing deep learning architectures affect a model's learned spatial and temporal dependencies, and how those learned dependencies affect a model's ability to maintain high accuracy when applied to unseen environmental conditions. Using the Delaware River Basin in the northeastern United States as a test case, we compare two spatiotemporally aware process–guided deep learning models for predicting stream temperature (a recurrent graph convolution network—RGCN, and a temporal convolution graph model—Graph WaveNet). Both models achieve equally high predictive performance when testing data are well represented in the training data (test root mean squared errors of 1.64°C and 1.65°C); however, Graph WaveNet significantly outperforms RGCN in 4 out of 5 experiments where test partitions represent different types of unseen environmental conditions. XAI results show that the architecture of Graph WaveNet leads to learned spatial relationships with greater fidelity to physical processes, and that this fidelity improves the generalizability of the model when applied to shifting and/or unseen environmental conditions.

54 ENVIRONMENTAL SCIENCES↗