Sparse Observations Research Articles

Platoon Trajectory Completion in a Mixed Traffic Environment Under Sparse Observation

Obtaining sufficient trajectory data of human-driven vehicles (HDVs) is critical for effective control of connected automated vehicles (CAVs) in mixed traffic of HDVs and CAVs. However, due to limited sensing and communication capabilities, only a fraction of HDVs’ trajectories are often observed. This paper proposes a completion method to recovery all HDVs’ trajectories in a mixed platoon based on partial observations. The trajectory completion problem is formulated as an optimization problem, aiming to minimize the error between observed and completed trajectories with car-following constraints defined by Newell’s simplified car-following model. The method also allows various model parameters of different drivers, which is known as the inter-driver heterogeneity, to reduce the completion error. Validation using empirical trajectory data shows that the proposed method greatly lowers the completion error than other typical trajectory completion methods under sparse observation (6s/sample).

Probing the innermost region of the AU Microscopii debris disc

Context. AU Mic is a young and nearby M-dwarf star harbouring a circumstellar debris disc and one recently discovered planet on an eight-day orbit. Large-scale structures within the disc were also discovered and are moving outwards at high velocity. Aims. We aim to study this system with the highest spatial resolution in order to probe the innermost regions and to search for additional low-mass companions or set detection limits. Methods. The star was observed with two different high-angular resolution techniques probing complementary spatial scales. We obtained new Ks-band sparse aperture masking observations with VLT/SPHERE, which we combined with data from VLT/NACO, VLTI/PIONIER and VLTI/GRAVITY. Results. We did not detect additional close companions within the separation range 0.02–7 au from the parent star. We determined magnitude upper limits for companions of H ~ 9.8 mag within 0.02-0.5 au, Ks ~ 11.2 mag within 0.4–2.4 au, and L ~ 10.7 mag within 0.7–7 au. Using theoretical isochrones, we converted these magnitudes into upper limits on the mass of ~17 Mjup, ~12 Mjup, and ~9 Mjup, respectively. The PIONIER observations also allowed us to determine the angular diameter of AU Mic, θLD = 0.825 ± 0.033stat ± 0.038sys mas, which converts to a linear radius R = 0.862 ± 0.052 R⊙ when combined with the Gaia parallax. Conclusions. We did not detect the newly discovered planets orbiting AU Mic (M < 0.2 Mjup), but we derived upper limit masses for the innermost region of AU Mic. We do not have any detection with a significance beyond 3σ, the most significant signal with PIONIER being 2.9σ and that with SPHERE being 1.6σ. We applied the pyMESS2 code to estimate the detection probability of companions by combining radial velocities, multi-band SPHERE imaging, and our interferometric detection maps. We show that 99% of the companions down to ~0.5 Mjup can be detected within 0.02 au or 1 Mjup down to 0.2 au. The low-mass planets orbiting at ≲0.11 au (≲11 mas) from the star will not be directly detectable with the current adaptive optics (AO) and interferometric instruments because of its close orbit and very high contrast (~10−10 K). It will also be below the angular resolution and contrast limit of the next Extremely Large Telescope Infrared (ELT IR) imaging instruments.

Investigating NUCAPS Skill in Profiling Saharan Dust for Near-Real-Time Forecasting

Dust outflows off Northwest Africa often propagate westward across the North Tropical Atlantic Basin (NTAB) into the greater Caribbean and US. From a health perspective, weather forecasters in these regions often monitor hazardous air quality associated with this dust. However, forecasters can be constrained by sparse data observations upwind over the Atlantic of the impacted populated areas. Global satellite sounding retrievals can potentially augment and enhance the operational forecasting toolkit for monitoring Saharan dust episodes. The focus of this paper was to examine the skill of the NOAA Unique Combined Atmospheric Processing System (NUCAPS) temperature and water vapor profiles within the dust and non-dust conditions during the March 2019 NOAA Aerosols and Ocean Science Expedition (AEROSE). During this time, the NOAA Ron Brown research ship launched radiosondes to coincide with satellite overpasses that served as independent ground truth data for evaluating NUCAPS. Compared to RAOBs from the Ron Brown, the SNPP and NOAA-20 NUCAPS-derived soundings showed skill in profiling atmospheric conditions supporting Saharan dust monitoring. Outside of dust regions, the NOAA-20 NUCAPS surface temperature bias peaks at 2.0 K; the surface water vapor bias is minimal (~1000 hPa), with a small cold bias that peaks at −50% between 742 and 790 hPa. Corresponding temperature RMS values are less than 2.0 K; water vapor RMS values are generally below 70%. Within the dust regions, NOAA-20 NUCAPS temperature soundings show a cold bias peak of 2.6 K at 918 hPa and 113% of a moist bias peak at the same level. Corresponding temperature RMS values maximize at 3.5 K at 945 hPa; the water vapor RMS shows a peak value of 106% at the same level. Weather forecasters can apply NUCAPS across the NTAB in issuing timely and accurate hazardous air quality warnings and visibility alerts to health officials and the general public.

A new method for characterising shared space use networks using animal trapping data

Studying the social behaviour of small or cryptic species often relies on constructing networks from sparse point-based observations of individuals (e.g. live trapping data). A common approach assumes that individuals that have been detected sequentially in the same trapping location will also be more likely to have come into indirect and/or direct contact. However, there is very little guidance on how much data are required for making robust networks from such data. In this study, we highlight that sequential trap sharing networks broadly capture shared space use (and, hence, the potential for contact) and that it may be more parsimonious to directly model shared space use. We first use empirical data to show that characteristics of how animals use space can help us to establish new ways to model the potential for individuals to come into contact. We then show that a method that explicitly models individuals’ home ranges and subsequent overlap in space among individuals (spatial overlap networks) requires fewer data for inferring observed networks that are more strongly correlated with the true shared space use network (relative to sequential trap sharing networks). Furthermore, we show that shared space use networks based on estimating spatial overlap are also more powerful for detecting biological effects. Finally, we discuss when it is appropriate to make inferences about social interactions from shared space use. Our study confirms the potential for using sparse trapping data from cryptic species to address a range of important questions in ecology and evolution.Significance statementCharacterising animal social networks requires repeated (co-)observations of individuals. Collecting sufficient data to characterise the connections among individuals represents a major challenge when studying cryptic organisms—such as small rodents. This study draws from existing spatial mark-recapture data to inspire an approach that constructs networks by estimating space use overlap (representing the potential for contact). We then use simulations to demonstrate that the method provides consistently higher correlations between inferred (or observed) networks and the true underlying network compared to current approaches and requires fewer observations to reach higher correlations. We further demonstrate that these improvements translate to greater network accuracy and to more power for statistical hypothesis testing.

Halving of Swiss glacier volume since 1931 observed from terrestrial image photogrammetry

Abstract. The monitoring of glaciers in Switzerland has a long tradition, yet glacier changes during the 20th century are only known through sparse observations. Here, we estimate a halving of Swiss glacier volumes between 1931 and 2016 by mapping historical glacier elevation changes at high resolution. Our analysis relies on a terrestrial image archive known as TerrA, which covers about 86 % of the Swiss glacierised area with 21 703 images acquired during the period 1916–1947 (with a median date of 1931). We developed a semi-automated workflow to generate digital elevation models (DEMs) from these images, resulting in a 45 % total glacier coverage. Using the geodetic method, we estimate a Swiss-wide glacier mass balance of −0.52 ± 0.09 m w.e. a−1 between 1931 and 2016. This equates to a 51.5 ± 8.0 % loss in glacier volume. We find that low-elevation, high-debris-cover, and gently sloping glacier termini are conducive to particularly high mass losses. In addition to these glacier-specific, quasi-centennial elevation changes, we present a new inventory of glacier outlines with known timestamps and complete attributes from around 1931. The fragmented spatial coverage and temporal heterogeneity of the TerrA archive are the largest sources of uncertainty in our glacier-specific estimates, reaching up to 0.50 m w.e. a−1. We suggest that the high-resolution mapping of historical surface elevations could also unlock great potential for research fields other than glaciology.

PK-RNN-V E: A deep learning model approach to vancomycin therapeutic drug monitoring using electronic health record data

Vancomycin is a commonly used antimicrobial in hospitals, and therapeutic drug monitoring (TDM) is required to optimize its efficacy and avoid toxicities. Bayesian models are currently recommended to predict the antibiotic levels. These models, however, although using carefully designed lab observations, were often developed in limited patient populations. The increasing availability of electronic health record (EHR) data offers an opportunity to develop TDM models for real-world patient populations. Here, we present a deep learning-based pharmacokinetic prediction model for vancomycin (PK-RNN-V E) using a large EHR dataset of 5,483 patients with 55,336 vancomycin administrations. PK-RNN-V E takes the patient’s real-time sparse and irregular observations and offers dynamic predictions. Our results show that RNN-PK-V E offers a root mean squared error (RMSE) of 5.39 and outperforms the traditional Bayesian model (VTDM model) with an RMSE of 6.29. We believe that PK-RNN-V E can provide a pharmacokinetic model for vancomycin and other antimicrobials that require TDM.

A Method on Vibration Positioning of Φ-OTDR System Based on Compressed Sensing

To improve the accuracy of vibration location information of the Φ-OTDR system, a method of optical fiber sensing signal processing based on Compressed Sensing (CS) is proposed in this paper. First, a sensing matrix is constructed by sparse array and observation array, and the mathematical modeling of CS is established. Then the reconstruction algorithm is designed to complete the accurate reconstruction of the vibration signal with a small number of sampling values. Second, the time domain signal composed of all scattered light at the vibration source is analyzed to complete the measurement of vibration frequency. Finally, a heterodyne Phi-OTDR system is built, using PZT to simulate the vibration source. And 36 scattered light information is continuously collected at the frequency of 100 Hz and 500 Hz. The experimental verification of the CS method in this paper is carried out, besides, the signal-to-noise ratio (SNR) is compared with the traditional difference method, Empirical Mode Decomposition (EMD) method, and Variational Mode Decomposition (VMD) method. The experimental results show that the SNR of this method is improved to 40.41 dB and 30.62 dB, the optimal spatial resolution is 9m, and the maximum relative error of frequency measurement is 8.1%. Compared with the results of traditional I/Q demodulation, the frequency information obtained by the two methods is consistent, and there is no obvious deviation from the actual frequency of PZT. While improving the SNR of the system, this method can greatly simplify the signal processing process and improve the positioning accuracy of the system, which has a certain practical application value in the fields of infrastructure monitoring, resource exploration, and vibration detection.

Bayesian additive regression trees in spatial data analysis with sparse observations

ABSTRACT Recently, the Bayesian additive regression trees (BART) model has been one of the most widely used methods in the field of Bayesian nonparametrics. The spatially adjusted BART model was proposed in 2007; however, it has a limited number of applications in spatial data analysis. We propose a new version of the BART model for spatial data analysis tailored to a setting with sparse spatial observations and high-dimensional predictors. Specifically, an adjacency-based weight matrix in a conditional autoregressive prior is replaced by a new definition of weight matrix reflecting any type of spatial structures, and selection probabilities on predictors in the model take a new sparsity-inducing prior. The applicability of the proposed model is examined through a set of simulation studies. We apply the proposed model to predict the annual ambient particulate matter (PM) concentrations for zip code locations in Massachusetts, Connecticut and Rhode Island.

Constructing high-resolution groundwater drought at spatio-temporal scale using GRACE satellite data based on machine learning in the Indus Basin

The complicated phenomenon induced by inadequate precipitation is a drought that impacts water resources and human life. Traditional methods to assess groundwater drought events are hindered due to sparse groundwater observations on a spatio-temporal scale. These groundwater drought events are not well studied in the study area of the Indus Basin Irrigation System (IBIS) holistically. This study applied four machine learning models to the training datasets of Gravity Recovery and Climate Experiment (GRACE) Terrestrial Water Storage (TWS) and Groundwater Storage (GWS) data to improve resolution to 0.25° from 1°. The Extreme Gradient Boosting (XGBoost) model outperformed the four models and results showed Pearson correlation (R) (0.99), Nash Sutcliff Efficiency (NSE) (0.99), Root Mean Square Error (RMSE) (5.22 mm), and Mean Absolute Error (MAE) (2.75 mm). The GRACE Groundwater Drought Index (GGDI) was calculated by normalizing XGBoost-downscaled GWS. The trend characteristics, the temporal evolution, and spatial distribution of GGDI were analyzed across the IBIS from 2003 to 2016. The wavelet coherence approach was used to evaluate the relationship between teleconnection factors and GGDI. The XGBoost downscaling model can accurately reproduce local groundwater behavior, with the acceptable correlation of coefficient values for validation (ranging from 0.02 to 0.84). The accumulated Standardized Precipitation Evapotranspiration Index (SPEI) with the time of 1, 3, and 6 months, and self-calibrated Palmer Drought Severity Index (sc-PDSI) were used to validate GGDI. The findings have demonstrated that GGDI has comparable drought patterns to SPEI-3 and SPEI-6 and sc-PDSI. The teleconnection factors have a significant impact on the GGDI shown by the wavelet coherence technique. The impact of the sea surface temperature index (namely, NINO3.4) on GGDI was observed significantly high among other teleconnection factors in the IBIS. The proposed framework can serve as a useful tool for drought monitoring and a better understanding of extreme hydroclimatic conditions in the IBIS and other similar climatic regions.

Modelling glacier mass balance and climate sensitivity in the context of sparse observations: application to Saskatchewan Glacier, western Canada

Abstract. Glacier mass balance models are needed at sites with scarce long-term observations to reconstruct past glacier mass balance and assess its sensitivity to future climate change. In this study, North American Regional Reanalysis (NARR) data were used to force a physically based, distributed glacier mass balance model of Saskatchewan Glacier for the historical period 1979–2016 and assess its sensitivity to climate change. A 2-year record (2014–2016) from an on-glacier automatic weather station (AWS) and historical precipitation records from nearby permanent weather stations were used to downscale air temperature, relative humidity, wind speed, incoming solar radiation and precipitation from the NARR to the station sites. The model was run with fixed (1979, 2010) and time-varying (dynamic) geometry using a multitemporal digital elevation model dataset. The model showed a good performance against recent (2012–2016) direct glaciological mass balance observations as well as with cumulative geodetic mass balance estimates. The simulated mass balance was not very sensitive to the NARR spatial interpolation method, as long as station data were used for bias correction. The simulated mass balance was however sensitive to the biases in NARR precipitation and air temperature, as well as to the prescribed precipitation lapse rate and ice aerodynamic roughness lengths, showing the importance of constraining these two parameters with ancillary data. The glacier-wide simulated energy balance regime showed a large contribution (57 %) of turbulent (sensible and latent) heat fluxes to melting in summer, higher than typical mid-latitude glaciers in continental climates, which reflects the local humid “icefield weather” of the Columbia Icefield. The static mass balance sensitivity to climate was assessed for prescribed changes in regional mean air temperature between 0 and 7 ∘C and precipitation between −20 % and +20 %, which comprise the spread of ensemble Representative Concentration Pathway (RCP) climate scenarios for the mid (2041–2070) and late (2071–2100) 21st century. The climate sensitivity experiments showed that future changes in precipitation would have a small impact on glacier mass balance, while the temperature sensitivity increases with warming, from −0.65 to −0.93 m w.e. a−1 ∘C−1. The mass balance response to warming was driven by a positive albedo feedback (44 %), followed by direct atmospheric warming impacts (24 %), a positive air humidity feedback (22 %) and a positive precipitation phase feedback (10 %). Our study underlines the key role of albedo and air humidity in modulating the response of winter-accumulation type mountain glaciers and upland icefield-outlet glacier settings to climate.

Deep Learning Augmented Data Assimilation: Reconstructing Missing Information with Convolutional Autoencoders

Abstract Remote sensing data play a critical role in improving numerical weather prediction (NWP). However, the physical principles of radiation dictate that data voids frequently exist in physical space (e.g., subcloud area for satellite infrared radiance or no-precipitation region for radar reflectivity). Such data gaps impair the accuracy of initial conditions derived from data assimilation (DA), which has a negative impact on NWP. We use the barotropic vorticity equation to demonstrate the potential of deep learning augmented data assimilation (DDA), which involves reconstructing spatially complete pseudo-observation fields from incomplete observations and using them for DA. By training a convolutional autoencoder (CAE) with a long simulation at a coarse “forecast” resolution (T63), we obtained a deep learning approximation of the “reconstruction operator,” which maps spatially incomplete observations to a model state with full spatial coverage and resolution. The CAE was applied to an incomplete streamfunction observation (∼30% missing) from a high-resolution benchmark simulation and demonstrated satisfactory reconstruction performance, even when only very sparse (1/16 of T63 grid density) observations were used as input. When only spatially incomplete observations are used, the analysis fields obtained from ensemble square root filter (EnSRF) assimilation exhibit significant error. However, in DDA, when EnSRF takes in the combined data from the incomplete observations and CAE reconstruction, analysis error reduces significantly. Such gains are more pronounced with sparse observation and small ensemble size because the DDA analysis is much less sensitive to observation density and ensemble size than the conventional DA analysis, which is based solely on incomplete observations. Significance Statement Data assimilation plays a critical role in improving the skills of modern numerical weather prediction by establishing accurate initial conditions. However, unobservable regions are common in observation data, particularly those derived from remote sensing. The nonlinear relationship between data from observable regions and the physical state of unobservable regions may impede DA efficiency. As a result, we propose that deep learning be used to improve data assimilation in such cases by reconstructing a spatially complete first guess of the physical state with deep learning and then applying data assimilation to the reconstructed field. Such deep learning augmentation is found effective in improving the accuracy of data assimilation, especially for sparse observation and small ensemble size.

Optimized Diffusion Imaging for Brain Structural Connectome Analysis.

High angular resolution diffusion imaging (HARDI) is a type of diffusion magnetic resonance imaging (dMRI) that measures diffusion signals on a sphere in q-space. It has been widely used in data acquisition for human brain structural connectome analysis. To more accurately estimate the structural connectome, dense samples in q-space are often acquired, potentially resulting in long scanning times and logistical challenges. This paper proposes a statistical method to select q-space directions optimally and estimate the local diffusion function from sparse observations. The proposed approach leverages relevant historical dMRI data to calculate a prior distribution to characterize local diffusion variability in each voxel in a template space. For a new subject to be scanned, the priors are mapped into the subject-specific coordinate and used to help select the best q-space samples. Simulation studies demonstrate big advantages over the existing HARDI sampling and analysis framework. We also applied the proposed method to the Human Connectome Project data and a dataset of aging adults with mild cognitive impairment. The results indicate that with very few q-space samples (e.g., 15 or 20), we can recover structural brain networks comparable to the ones estimated from 60 or more diffusion directions with the existing methods.

Calibration of COSMO model in the Central-Eastern Mediterranean area adjusted over the domains of Greece and Israel

The goal of this work is to optimize five parameters of expected high impact on COSMO model performance in non-hydrostatic regime using a calibration method based on the so called Meta-Model. The domain of consideration covers the Central-Eastern Mediterranean area with a horizontal mesh of 0.030 (~3.3 km). The optimization technique considers observations at twenty-two Greek and sixty-four Israeli meteorological stations of 24 h accumulated precipitation, minimum and maximum 2-m dry and dew-point temperatures at sixty dates spread over the year 2019, five for every month, in order to comply with the seasonal variability. The optimum model parameters found have been evaluated by application on an independent set of sixty days chosen from the same year-period and in the same manner. An overall model performance score improvement of order 5% was obtained for both sets of test cases. Better verification scores were found regarding almost all meteorological fields considered with respect to average values, mean absolute and standard deviation errors, as well as threat scores regarding precipitation. The use of relatively sparse local observations, instead of the gridded data sets available to places with dense meteorological station networks where the methodology has been successfully applied in numerical weather prediction and climatology, makes this approach, upon its successful realization, suitable to a significantly larger community of model users with modest observation resources. This approach is also expected to be of value for domains with a large marine area fraction, like the domain chosen, where observations are sporadic, if available, by nature as well as to regions of low population density.

SHAKING: Adjusted spherical harmonics adding KrigING method for near real-time ionospheric modeling with multi-GNSS observations

Precise positioning based on Global Navigation Satellite System (GNSS) technique requires high accuracy ionospheric total electron content (TEC) correction models to account for the ionospheric path delay errors. We present an adjusted Spherical Harmonics Adding KrigING method (SHAKING) approach for regional ionospheric vertical TEC (VTEC) modeling in real time. In the proposed SHAKING method, the VTEC information over the sparse observation data area is extrapolated by the Adjusted Spherical Harmonic (ASH) function, and the boundary distortion in regional VTEC modeling is corrected by the stochastic VTEC estimated using Kriging interpolation. Using real-time GPS, GLONASS and BDS-2/3 data streams of the Crust Movement Observation Network of China (CMONOC), the SHAKING-based regional ionospheric VTEC maps are re-constructed over China and its boundary regions. Compared to GNSS VTECs derived from the independent stations, the quality of SHAKING solution improves by 13–31% and 6–33% with respect to the ASH-only solution during high and low geomagnetic periods, respectively. Compared to the inverse distance weighting (IDW) generated result, significant quality improved of SHAKING-based VTEC maps is also observed, especially over the edge areas with an improvement of 60–80%. Overall, the proposed SHAKING method exhibits notable advantage over the existing regional VTEC modeling techniques, which can be used for regional TEC modeling and associated high-precision positioning applications.

Physics-informed PointNet: A deep learning solver for steady-state incompressible flows and thermal fields on multiple sets of irregular geometries

We present a novel physics-informed deep learning framework for solving steady-state incompressible flow on multiple sets of irregular geometries by incorporating two main elements: using a point-cloud based neural network to capture geometric features of computational domains, and using the mean squared residuals of the governing partial differential equations, boundary conditions, and sparse observations as the loss function of the network to capture the physics. While the solution of the continuity and Navier-Stokes equations is a function of the geometry of the computational domain, current versions of physics-informed neural networks have no mechanism to express this functionally in their outputs, and thus are restricted to obtain the solutions only for one computational domain with each training procedure. Using the proposed framework, three new facilities become available. First, the governing equations are solvable on a set of computational domains containing irregular geometries with high variations with respect to each other but requiring training only once. Second, after training the introduced framework on the set, it is now able to predict the solutions on domains with unseen geometries from seen and unseen categories as well. The former and the latter both lead to savings in computational costs. Finally, all the advantages of the point-cloud based neural network for irregular geometries, already used for supervised learning, are transferred to the proposed physics-informed framework. The effectiveness of our framework is shown through the method of manufactured solutions and thermally-driven flow for forward and inverse problems.

Post Processing Sparse And Instantaneous 2D Velocity Fields Using Physics-Informed Neural Networks

This work tackles the problem of resolving high-resolution velocity fields from a set of sparse off-grid observations. We follow the framework of Physics-Informed Neural Networks where simple Multi-layer Perceptor (MLP) are trained to solve partial differential equations (PDEs). In contrast with other state-of-the-art methods based of Convolutional Neural Networks, these models can be applied to super-resolve sparse Lagrangian velocity measurements. Moreover, such a framework can be easily extended to output divergence-free quantities and offer simple implementation of prior physical as regularization terms. In particular, we employ a sub-grid model based on structure-functions to improve the accuracy of the super-resolved velocity fields of turbulent flows. Numerical experimentation on synthetic data shows that the proposed approach can accurately reconstruct dense Eulerian velocity fields from sparse Lagrangian velocity measurements.

Topographic Hotspots of Southern Ocean Eddy Upwelling

The upwelling of cold water from the depths of the Southern Ocean to its surface closes the global overturning circulation and facilitates uptake of anthropogenic heat and carbon. Upwelling is often conceptualised in a zonally averaged framework as the result of isopycnal flattening via baroclinic eddies. However, upwelling is zonally non-uniform and occurs in discrete hotspots near topographic features. The mechanisms that facilitate topographically confined eddy upwelling remain poorly understood and thus limit the accuracy of parameterisations in coarse-resolution climate models.Using a high-resolution global ocean sea-ice model, we calculate spatial distributions of upwelling transport and energy conversions associated with barotropic and baroclinic instability, derived from a thickness-weighted energetics framework. We find that five major topographic hotspots of upwelling, covering less than 30% of the circumpolar longitude range, account for up to 76% of the southward eddy upwelling transport. The conversion of energy into eddies via baroclinic instability is highly spatially correlated with upwelling transport, unlike the barotropic energy conversion, which is also an order of magnitude smaller than the baroclinic conversion. This result suggests that eddy parameterisations that quantify baroclinic energy conversions could be used to improve the simulation of upwelling hotspots in climate models. We also find that eddy kinetic energy maxima are found on average 110 km downstream of upwelling hotspots in accordance with sparse observations. Our findings demonstrate the importance of localised mechanisms to Southern Ocean dynamics.

A generalized framework for unsupervised learning and data recovery in computational fluid dynamics using discretized loss functions

The authors present generalized finite-volume-based discretized loss functions integrated into pressure-linked algorithms for physics-based unsupervised training of neural networks (NNs). In contrast to automatic differentiation-based counterparts, discretized loss functions leverage well-developed numerical schemes of computational fluid dynamics (CFD) for tailoring NN training specific to the flow problems. For validation, neural network-based solvers (NN solvers) are trained by posing equations such as the Poisson equation, energy equation, and Spalart–Allmaras model as loss functions. The predictions from the trained NNs agree well with the solutions from CFD solvers while also providing solution time speed-ups of up to seven times. Another application of unsupervised learning is the novel hybrid loss functions presented in this study. Hybrid learning combines the information from sparse or partial observations with a physics-based loss to train the NNs accurately and provides training speed-ups of up to five times compared with a fully unsupervised method. Also, to properly utilize the potential of discretized loss functions, they are formulated in a machine learning (ML) framework (TensorFlow) integrated with a CFD solver (OpenFOAM). The ML-CFD framework created here infuses versatility into the training by giving loss functions access to the different numerical schemes of the OpenFOAM. In addition, this integration allows for offloading the CFD programming to OpenFOAM, circumventing bottlenecks from manually coding new flow conditions in a solely ML-based framework like TensorFlow.

Quantifying Spatio-Temporal Boundary Condition Uncertainty for the North American Deglaciation

Ice sheet models are used to study the deglaciation of North America at the end of the last ice age (past 21,000 years), so that we might understand whether and how existing ice sheets may reduce or disappear under climate change. Though ice sheet models have a few parameters controlling physical behaviour of the ice mass, they also require boundary conditions for climate (spatio-temporal fields of temperature and precipitation, typically on regular grids and at monthly intervals). The behaviour of the ice sheet is highly sensitive to these fields, and there is relatively little data from geological records to constrain them as the land was covered with ice. We develop a methodology for generating a range of plausible boundary conditions, using a low-dimensional basis representation of the spatio-temporal input. We derive this basis by combining key patterns, extracted from a small ensemble of climate model simulations of the deglaciation, with sparse spatio-temporal observations. By jointly varying the ice sheet parameters and basis vector coefficients, we run ensembles of the Glimmer ice sheet model that simultaneously explore both climate and ice sheet model uncertainties. We use these to calibrate the ice sheet physics and boundary conditions for Glimmer, by ruling out regions of the joint coefficient and parameter space via history matching. We use binary ice/no ice observations from reconstructions of past ice sheet margin position to constrain this space by introducing a novel metric for history matching to binary data.

Differential Assessment of Black-Box AI Agents

Much of the research on learning symbolic models of AI agents focuses on agents with stationary models. This assumption fails to hold in settings where the agent's capabilities may change as a result of learning, adaptation, or other post-deployment modifications. Efficient assessment of agents in such settings is critical for learning the true capabilities of an AI system and for ensuring its safe usage. In this work, we propose a novel approach to differentially assess black-box AI agents that have drifted from their previously known models. As a starting point, we consider the fully observable and deterministic setting. We leverage sparse observations of the drifted agent's current behavior and knowledge of its initial model to generate an active querying policy that selectively queries the agent and computes an updated model of its functionality. Empirical evaluation shows that our approach is much more efficient than re-learning the agent model from scratch. We also show that the cost of differential assessment using our method is proportional to the amount of drift in the agent's functionality.