Engineering Papers⌕ Search

SEARCH · Engineering Papers

Results for “small dataset”

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 181 records · Page 10

A hybrid data–model approach to map soil thickness in mountain hillslopes

Abstract. Soil thickness plays a central role in the interactions between vegetation, soils, and topography, where it controls the retention and release of water, carbon, nitrogen, and metals. However, mapping soil thickness, here defined as the mobile regolith layer, at high spatial resolution remains challenging. Here, we develop a hybrid model that combines a process-based model and empirical relationships to estimate the spatial heterogeneity of soil thickness with fine spatial resolution (0.5 m). We apply this model to two aspects of hillslopes (southwest- and northeast-facing, respectively) in the East River watershed in Colorado. Two independent measurement methods – auger and cone penetrometer – are used to sample soil thickness at 78 locations to calibrate the local value of unconstrained parameters within the hybrid model. Sensitivity analysis using the hybrid model reveals that the diffusion coefficient used in hillslope diffusion modeling has the largest sensitivity among all input parameters. In addition, our results from both sampling and modeling show that, in general, the northeast-facing hillslope has a deeper soil layer than the southwest-facing hillslope. By comparing the soil thickness estimated between a machine-learning approach and this hybrid model, the hybrid model provides higher accuracy and requires less sampling data. Modeling results further reveal that the southwest-facing hillslope has a slightly faster surface soil erosion rate and soil production rate than the northeast-facing hillslope, which suggests that the relatively less dense vegetation cover and drier surface soils on the southwest-facing slopes influence soil properties. With seven parameters in total for calibration, this hybrid model can provide a realistic soil thickness map with a relatively small amount of sampling dataset comparing to machine-learning approach. Integrating process-based modeling and statistical analysis not only provides a thorough understanding of the fundamental mechanisms for soil thickness prediction but also integrates the strengths of both statistical approaches and process-based modeling approaches.

58 GEOSCIENCES↗

GeneLab for High Schools: Data Mining for the Next Generation

Modern biological sciences have become increasingly based on molecular biology and high-throughput molecular techniques, such as genomics, transcriptomics, and proteomics. NASA Scientists and the NASA Space Biology Program have aimed to examine the fundamental building blocks of life (RNA, DNA and protein) in order to understand the response of living organisms to space and aid in fundamental research discoveries on Earth. In an effort to enable NASA funded science to be available to everyone, NASA has collected the data from omics studies and curated them in a data system called GeneLab. Whilst most college-level interns, academics and other scientists have had some interaction with omics data sets and analysis tools, high school students often have not. Therefore, the Space Biology Program is implementing a new Summer Program for high-school students that aims to inspire the next generation of scientists to learn about and get involved in space research using GeneLabs Data System. The program consists of three main components core learning modules, focused on developing students knowledge on the Space Biology Program and Space Biology research, Genelab and the data system, and previous research conducted on model organisms in space; networking and team work, enabling students to interact with guest lecturers from local universities and their fellow peers, and also enabling them to visit local universities and genomics centers around the Bay area; and finally an independent learning project, whereby students will be required to form small groups, analyze a dataset on the Genelab platform, generate a hypothesis and develop a research plan to test their hypothesis. This program will not only help inspire high-school students to become involved in space-based research but will also help them develop key critical thinking and bioinformatics skills required for most college degrees and furthermore, will enable them to establish networks with their peers and connections with university Professors that may help them achieve their educational goals.

genelab↗

Open Science for Life in Space: Data Sharing and Tools for Knowledge Discovery

The fast-growing array of space biological data, which in the past was simply archived after minimal analysis, holds great potential if it can be reorganized and formatted for Open Science. Organizing the data for such analysis is a challenge because of its diverse nature (molecular, cellular, tissue, whole organism, behavior; tabular, imagery). Open Science is the concept that the more people have access to scientifically curated data, the more knowledge will be gained. This led NASA to start the development of GeneLab in 2015. GeneLab houses spaceflight and space-analog multi-omics datasets from plant, rodent, small animal, and microbial experiments. The success and knowledge gained from GeneLab led to a new alliance of NASA “Open Science Data Repositories” (OSDR), which include the Ames Life Sciences Data Archive (ALSDA) and the NASA Biological Institutional Scientific Collection (NBISC). Both are adopting the GeneLab data system, so data are more findable, accessible, interoperable, and reusable (FAIR). OSDR systems provide users the ability to upload, download, search, share, analyze, and visualize. Open Science also needs strong confidence in the data, which is gained through building science communities. With ~400 current members, GeneLab and ALSDA formed Analysis Working Groups (AWGs) to provide feedback on processing pipelines, metadata curation standards (for ‘omics and phenotypic-physiological-behavioral assays), and to collaborate in effectively reusing data. The AWG also led to the development of the Radiation Biology Ontology (RBO), ensuring radiation metadata are efficiently captured, connected, and interoperable. Feedback from the AWG provided design input toward the new single point-of-entry data submission portal for all investigators to submit, curate, and share their research data. Space biological data is now maximally open access, collected-curated with rich metadata, and formatted for interoperability to enable systems biology, meta-analysis, knowledge graphs, machine learning, modeling, and other reuse approaches. With potential for further federation of OSDR for data mining with traditional biological and medical databases (NIH, NCI, EBI, etc.), a new era for space biology has begun to support the knowledge discovery necessary for Lunar and Martian missions.

Ryan T Scott↗

Genetic Algorithms and Classification Trees in Feature Discovery: Diabetes and the NHANES database

This paper presents a feature selection methodology that can be applied to datasets containing a mixture of continuous and categorical variables. Using a Genetic Algorithm (GA), this method explores a dataset and selects a small set of features relevant for the prediction of a binary (1/0) response. Binary classification trees and an objective function based on conditional probabilities are used to measure the fitness of a given subset of features. The method is applied to health data in order to find factors useful for the prediction of diabetes. Results show that our algorithm is capable of narrowing down the set of predictors to around 8 factors that can be validated using reputable medical and public health resources.

Heredia-Langner, Alejandro↗

Prediction Models for Dynamic Demand Response

As Smart Grids move closer to dynamic curtailment programs, Demand Response (DR) events will become necessary not only on fixed time intervals and weekdays predetermined by static policies, but also during changing decision periods and weekends to react to real-time demand signals. Unique challenges arise in this context vis-a-vis demand prediction and curtailment estimation and the transformation of such tasks into an automated, efficient dynamic demand response (D 2 R) process. While existing work has concentrated on increasing the accuracy of prediction models for DR, there is a lack of studies for prediction models for D 2 R, which we address in this paper. Our first contribution is the formal definition of D 2 R, and the description of its challenges and requirements. Our second contribution is a feasibility analysis of very-short-term prediction of electricity consumption for D 2 R over a diverse, large-scale dataset that includes both small residential customers and large buildings. Our third, and major contribution is a set of insights into the predictability of electricity consumption in the context of D 2 R. Specifically, we focus on prediction models that can operate at a very small data granularity (here 15-min intervals), for both weekdays and weekends - all conditions that characterize scenarios for D 2 R. We find that short-term time series and simple averaging models used by Independent Service Operators and utilities achieve superior prediction accuracy. We also observe that workdays are more predictable than weekends and holiday. Also, smaller customers have large variation in consumption and are less predictable than larger buildings. Key implications of our findings are that better models are required for small customers and for non-workdays, both of which are critical for D 2 R. Also, prediction models require just few days’ worth of data indicating that small amounts of historical training data can be used to make reliable predictions, simplifying the complexity of big data challenge associated with D 2 R.

Aman, Saima↗

Classification of Cloud Particle Imagery from Aircraft Platforms Using Convolutional Neural Networks

Abstract A vast amount of ice crystal imagery exists from a variety of field campaign initiatives that can be utilized for cloud microphysical research. Here, nine convolutional neural networks are used to classify particles into nine regimes on over 10 million images from the Cloud Particle Imager probe, including liquid and frozen states and particles with evidence of riming. A transfer learning approach proves that the Visual Geometry Group (VGG-16) network best classifies imagery with respect to multiple performance metrics. Classification accuracies on a validation dataset reach 97% and surpass traditional automated classification. Furthermore, after initial model training and preprocessing, 10 000 images can be classified in approximately 35 s using 20 central processing unit cores and two graphics processing units, which reaches real-time classification capabilities. Statistical analysis of the classified images indicates that a large portion (57%) of the dataset is unusable, meaning the images are too blurry or represent indistinguishable small fragments. In addition, 19% of the dataset is classified as liquid drops. After removal of fragments, blurry images, and cloud drops, 38% of the remaining ice particles are largely intersecting the image border (≥10% cutoff) and therefore are considered unusable because of the inability to properly classify and dimensionalize. After this filtering, an unprecedented database of 1 560 364 images across all campaigns is available for parameter extraction and bulk statistics on specific particle types in a wide variety of storm systems, which can act to improve the current state of microphysical parameterizations.

54 ENVIRONMENTAL SCIENCES↗

Fracture Strength Determination Methods for Ceramic Materials Applied to Uranium Dioxide

The Advanced Fuels Campaign is currently focusing on development of accident-tolerant fuels that possess a range of property modifications intended to improve fuel performance during accident and transient conditions as well as extend license limits to burnups beyond 62 MWd/kgU. Both of these drivers have identified understanding and mitigating fuel cracking as a key performance criterion. Small-scale cantilever beam testing has been identified as a plausible method by which to collect fracture data for UO 2 as a function of chemical and structural evolutions introduced either during fabrication or irradiation. While this method has been found to be capable of providing data that are in reasonable agreement with literature for unirradiated UO 2 , the inherently small sample volumes that can be sampled limit its ability to capture the statistical nature of mechanisms that govern the fracture of brittle ceramics. A biaxial flexure strength test was developed to be used for unirradiated UO 2 . This method is standard in the community, but no systems presently in use at national laboratories, universities, or private companies are available to be used for nuclear fuel materials. This report describes the operation and benchmarking of this system, which has been validated for a number of common oxides. Future work will extend this system to characterization of both doped UO 2 and other relevant microstructural modifications that can also be measured using small-scale cantilever beam testing. Comparison of datasets collected using both methods will allow the overall applicability of small-scale techniques to be assessed and provide confidence or bounds on their use for irradiated fuels.

11 NUCLEAR FUEL CYCLE AND FUEL MATERIALS↗

On the Generalizability of Time-of-Flight Convolutional Neural Networks for Noninvasive Acoustic Measurements

Bulk wave acoustic time-of-flight (ToF) measurements in pipes and closed containers can be hindered by guided waves with similar arrival times propagating in the container wall, especially when a low excitation frequency is used to mitigate sound attenuation from the material. Convolutional neural networks (CNNs) have emerged as a new paradigm for obtaining accurate ToF in non-destructive evaluation (NDE) and have been demonstrated for such complicated conditions. However, the generalizability of ToF-CNNs has not been investigated. In this work, we analyze the generalizability of the ToF-CNN for broader applications, given limited training data. We first investigate the CNN performance with respect to training dataset size and different training data and test data parameters (container dimensions and material properties). Furthermore, we perform a series of tests to understand the distribution of data parameters that need to be incorporated in training for enhanced model generalizability. This is investigated by training the model on a set of small- and large-container datasets regardless of the test data. We observe that the quantity of data partitioned for training must be of a good representation of the entire sets and sufficient to span through the input space. The result of the network also shows that the learning model with the training data on small containers delivers a sufficiently stable result on different feature interactions compared to the learning model with the training data on large containers. To check the robustness of the model, we tested the trained model to predict the ToF of different sound speed mediums, which shows excellent accuracy. Furthermore, to mimic real experimental scenarios, data are augmented by adding noise. We envision that the proposed approach will extend the applications of CNNs for ToF prediction in a broader range.

47 OTHER INSTRUMENTATION↗

Hyporheic, Floodplain, and Surface Water (on Floodplain and River) Geochemical Datasets, and Shapefiles on Meander C at the East River, Colorado

The purpose of this dataset was to sample surface water and groundwater geochemistry at random locations in the East River, Colorado to provide a broad spatial dataset of geochemistry. The East River is part of the Watershed Function Scientific Focus Area (WFSFA) located in the Upper Colorado River Basin, United States. We conducted this random sampling at multiple time points including during the snowmelt flooding stage, during peak flood, during the falling limb of flood, and during baseflow. This datasets contain many datafiles. We include flat text files of surface water geochemistry including surface water on the floodplain and surface water in the river, geochemistry of floodplain pore-water samples, and geochemistry of hyporheic pore-water samples. All datasets are also included as shapefiles so location geochemistry is co-located with sampling locations and sampling depths. This datasets also has the locations of small channels, and hollows that are on top of floodplain Meander C. These small channels funnel surface water from the hillslope through the floodplain and out to the river. All text datasets can be analyzed with software such as Excel or R. All shapefiles can be viewed with freely available R packages such as Leaflet.

54 ENVIRONMENTAL SCIENCES↗

RINO: Renormalization Group Invariance with No Labels

A common challenge with supervised machine learning (ML) in high energy physics (HEP) is the reliance on simulations for labeled data, which can often mismodel the underlying collision or detector response. To help mitigate this problem of domain shift, we propose RINO (Renormalization Group Invariance with No Labels), a self-supervised learning approach that can instead pretrain models directly on collision data, learning embeddings invariant to renormalization group flow scales. In this work, we pretrain a transformer-based model on jets originating from quantum chromodynamic (QCD) interactions from the JetClass dataset, emulating real QCD-dominated experimental data, and then finetune on the JetNet dataset -- emulating simulations -- for the task of identifying jets originating from top quark decays. RINO demonstrates improved generalization from the JetNet training data to JetClass data compared to supervised training on JetNet from scratch, demonstrating the potential for RINO pretraining on real collision data followed by fine-tuning on small, high-quality MC datasets, to improve the robustness of ML models in HEP.

Hao, Zichun [Caltech] (ORCID:0000000256244907)↗

DEVELOPING AND EXPLOITING A UNIQUE DATASET FROM SOUTH AFRICAN GOLD MINES FOR SOURCE CHARACTERIZATION AND WAVE PROPAGATION

In this project, we are developing and exploiting a unique seismic dataset to address the characteristics of small seismic events and the associated seismic signals observed at local (< 200 km) and regional (< 2000 km) distances. The dataset is being developed using mining-induced events from three deep gold mines in South Africa recorded on in-mine networks (< 1 km) composed of tens of high-frequency sensors, a network of four broadband stations installed as part of this project at the surface around the mines (1-10 km), and a network of existing broadband seismic stations at local/regional distances (50-1000 km) from the mines. Data acquisition has now been completed and includes: (1) {approx}2 years (2007 and 2008) of continuous recording by the surface broadband array, and (2) tens of thousands of mine tremors in the -3.4 < ML < 4.4 local magnitude range. Events with positive magnitudes are generally well recorded by the surface-mine stations, while magnitudes of 3.0 and larger are seen at regional distances (up to {approx} 600 km) in high-pass filtered recordings. We have now completed the quality control of the in-mine data gathered at the three gold mines included in this project. The quality control consisted of: (1) identification and analysis of outliers among the P- and S-wave travel-time picks reported by the in-mine network operator and (2) verification of sensor orientations. The outliers have been identified through a 'Wadati filter' that searches for the largest subset of P- and S-wave travel-time picks consistent with a medium of uniform wave-speed. They have observed that outliers are generally picked at a few select stations. They have also detected that trigger times were mistakenly reported as origin times by the in-mine network operator, and corrections have been obtained from the intercept times in the Wadati diagrams. Sensor orientations have been verified through rotations into the local ray-coordinate system and, when possible, corrected by correlating waveforms obtained from theoretical and empirical rotation angles. Full moment tensor solutions have been obtained for selected events within the Savuka network volume, with moment magnitudes in the 0.5 < M{sub W} < 2.6 range. The solutions were obtained by inverting P-, SV-, and SH-spectral amplitudes measured on the theoretically rotated waveforms with visually assigned polarities. Most of the solutions have a non-zero implosive contribution (47 out of 76), while a small percentage is purely deviatoric (10 out of 76). The deviatoric moment tensors range from pure double couple to pure non-double couple mechanisms. We have also calibrated the regional stations for seismic coda-derived source spectra and moment magnitude using the envelope methodology of Mayeda et al. (2003). they tie the coda M{sub w} to independent values from waveform modeling. The resulting coda-based source spectra of shallow mining-related events show significant spectral peaking that is not seen in deeper tectonic earthquakes. This coda peaking may be an independent method of identifying shallow events and is similar to coda peaking with previously observed for Nevada explosions, where the frequency of the observed spectral peak correlates with the depth of burial (Murphy et al., 2009).

58 GEOSCIENCES↗

Separating Physically Distinct Mechanisms in Complex Infrared Plasmonic Nanostructures via Machine Learning Enhanced Electron Energy Loss Spectroscopy

Electron energy loss spectroscopy (EELS) enables direct exploration of plasmonic phenomena at the nanometer level. To isolate individual plasmon modes, linear unmixing methods can be used to separate different physical mechanisms, but in larger and more complex systems the interpretability of the components becomes uncertain. Here, infrared plasmonic resonances in self-assembled heterogeneous monolayer films of doped-semiconductor nanoparticles are examined beyond linear unmixing techniques, and both supervised and unsupervised machine-learning-based analyses of hyperspectral EELS datasets are demonstrated. Additionally, in the supervised approach, a human operator labels a small number of pixels in the hyperspectral dataset corresponding to features of interest which are then propagated across the entire dataset. In the unsupervised approach, non-linear autoencoders are used to create a highly-reduced latent-space representation of the dataset, within which insight into the relevant physics can be gleaned from straightforward distance metrics that do not depend on operator input and bias. The advantage of these approaches is that the labeling separates physical mechanisms without altering the data, enabling robust analyses of the influence of heterogeneities in mesoscale complex systems.

36 MATERIALS SCIENCE↗

Biological Research and Space Health Enabled by Machine Learning to Support Deep Space Missions

A key science goal of the NASA “Moon to Mars” campaign is to understand how biology responds to the Lunar, Martian, and deep space environments in order to advance fundamental knowledge and support human space missions. Through artificial intelligence (AI) and machine learning (ML), a paradigm shift has begun in space biosciences and engineered astronaut health systems, to enable Earth-independence and mission operations autonomy. We describe current AI/ML methods to support 1) fundamental biology, 2) in situ analytics, 3) high performance computing, 4) automated science, 5) self-driving labs, 6) remote data management, 7) integrated mission biomonitoring, and 8) a Precision Space Health system. AI/ML approaches that can be integrated to support these domains include active learning, explainable AI, adaptive learning, causal inference, knowledge graphs, federated learning, transfer learning, and large language models. Finally, we present results from several current ML projects that are underway in the space biology field to address key challenges of small sample n, high feature count, heterogeneity, and sparse data. These include 1) connecting omics to phenotypic data using an ensemble model to infer causality of rodent liver health disruption, 2) usage of explainable ML to interrogate muscular underpinnings of muscle atrophy, 3) ML models analyzing and determining directed acyclic graphs of human health risk leveraging rodent bone datasets, 4) usage of large pre-trained models connecting biomedical knowledgebases with small spaceflight datasets to understand gene-to-gene interactions, and 5) a suite of benchmarked open science datasets enabling programmers to identify best algorithms to answer space biology questions.

space biology↗

Biological Research and Space Health Enabled by Machine Learning to Support Deep Space Missions

A key science goal of the NASA “Moon to Mars” campaign is to understand how biology responds to the Lunar, Martian, and deep space environments in order to advance fundamental knowledge, reduce risk, and support safe, productive human space missions. Through the powerful emerging computer science approaches of artificial intelligence (AI) and machine learning (ML), a paradigm shift has begun in biomedical science and engineered astronaut health systems, to enable Earth-independence and autonomy of mission operations. We present a decadal view of AI/ML architecture to support deep space mission goals, developed in concert with leaders in the field. We describe current AI/ML methods to support 1) fundamental biology, 2) in situ analytics, 3) high performance computing hardware, 4) automated science, 5) self-driving labs, 6) remote data management, 7) integrated real-time mission biomonitoring, and 8) a Precision Space Health system. Cutting-edge AI/ML approaches that can be integrated to support these domains include active learning, explainable AI, adaptive learning, causal inference, knowledge graphs, federated learning, transfer learning, and large language models. Finally, we present results from several current ML projects that are underway in the field to address key challenges of small sample n, high feature count, heterogeneity, and sparse data. These include 1) connecting omics data to phenotypic data using an ensemble model to infer causality of spaceflight rodent liver health disruption, 2) usage of explainable ML to interrogate the muscular underpinnings of spaceflight muscle atrophy, 3) ML models analyzing and determining directed acyclic graphs of human space health risk leveraging rodent bone datasets, 4) usage of large pre-trained models connecting biomedical knowledgebases with small spaceflight datasets to understand gene-to-gene interaction networks, and 5) a suite of benchmarked open science datasets (spaceflight mouse liver; radiation DNA damage) enabling programmers to identify the best ML algorithms to answer space biological science questions.

space biology↗

Biological Research and Space Health Enabled by Machine Learning to Support Deep Space Missions

A key science goal of the NASA “Moon to Mars” campaign is to understand how biology responds to the Lunar, Martian, and deep space environments in order to advance fundamental knowledge, reduce risk, and support safe, productive human space missions. Through the powerful emerging computer science approaches of artificial intelligence (AI) and machine learning (ML), a paradigm shift has begun in biomedical science and engineered astronaut health systems, to enable Earth-independence and autonomy of mission operations. We present a decadal view of AI/ML architecture to support deep space mission goals, developed in concert with leaders in the field. We describe current AI/ML methods to support 1) fundamental biology, 2) in situ analytics, 3) high performance computing hardware, 4) automated science, 5) self-driving labs, 6) remote data management, 7) integrated real-time mission biomonitoring, and 8) a Precision Space Health system. Cutting-edge AI/ML approaches that can be integrated to support these domains include active learning, explainable AI, adaptive learning, causal inference, knowledge graphs, federated learning, transfer learning, and large language models. Finally, we present results from several current ML projects that are underway in the field to address key challenges of small sample n, high feature count, heterogeneity, and sparse data. These include 1) connecting omics data to phenotypic data using an ensemble model to infer causality of spaceflight rodent liver health disruption, 2) usage of explainable ML to interrogate the muscular underpinnings of spaceflight muscle atrophy, 3) ML models analyzing and determining directed acyclic graphs of human space health risk leveraging rodent bone datasets, 4) usage of large pre-trained models connecting biomedical knowledgebases with small spaceflight datasets to understand gene-to-gene interaction networks, and 5) a suite of benchmarked open science datasets (spaceflight mouse liver; radiation DNA damage) enabling programmers to identify the best ML algorithms to answer space biological science questions.

space biology↗

Lagrangian flows within reflecting internal waves at a horizontal free-slip surface

In this paper sequel to Zhou and Diamessis [“Reflection of an internal gravity wave beam off a horizontal free-slip surface,” Phys. Fluids 25, 036601 (2013)], we consider Lagrangian flows within nonlinear internal waves (IWs) reflecting off a horizontal free-slip rigid lid, the latter being a model of the ocean surface. The problem is approached both analytically using small-amplitude approximations and numerically by tracking Lagrangian fluid particles in direct numerical simulation (DNS) datasets of the Eulerian flow. Inviscid small-amplitude analyses for both plane IWs and IW beams (IWBs) show that Eulerian mean flow due to wave-wave interaction and wave-induced Stokes drift cancels each other out completely at the second order in wave steepness A, i.e., O(A{sup 2}), implying zero Lagrangian mean flow up to that order. However, high-accuracy particle tracking in finite-Reynolds-number fully nonlinear DNS datasets from the work of Zhou and Diamessis suggests that the Euler-Stokes cancelation on O(A{sup 2}) is not complete. This partial cancelation significantly weakens the mean Lagrangian flows but does not entirely eliminate them. As a result, reflecting nonlinear IWBs produce mean Lagrangian drifts on O(A{sup 2}) and thus particle dispersion on O(A{sup 4}). The above findings can be relevant to predicting IW-driven mass transport in the oceanic surface and subsurface region which bears important observational and environmental implications, under circumstances where the effect of Earth rotation can be ignored.

71 CLASSICAL AND QUANTUM MECHANICS, GENERAL PHYSIC↗

Two-Step Hyperparameter Optimization Method: Accelerating Hyperparameter Search by Using a Fraction of a Training Dataset

Abstract Hyperparameter optimization (HPO) is an important step in machine learning (ML) model development, but common practices are archaic—primarily relying on manual or grid searches. This is partly because adopting advanced HPO algorithms introduces added complexity to the workflow, leading to longer computation times. This poses a notable challenge to ML applications, as suboptimal hyperparameter selections curtail the potential of ML model performance, ultimately obstructing the full exploitation of ML techniques. In this article, we present a two-step HPO method as a strategic solution to curbing computational demands and wait times, gleaned from practical experiences in applied ML parameterization work. The initial phase involves a preliminary evaluation of hyperparameters on a small subset of the training dataset, followed by a reevaluation of the top-performing candidate models postretraining with the entire training dataset. This two-step HPO method is universally applicable across HPO search algorithms, and we argue it has attractive efficiency gains. As a case study, we present our recent application of the two-step HPO method to the development of neural network emulators for aerosol activation. Although our primary use case is a data-rich limit with many millions of samples, we also find that using up to 0.0025% of the data—a few thousand samples—in the initial step is sufficient to find optimal hyperparameter configurations from much more extensive sampling, achieving up to 135× speedup. The benefits of this method materialize through an assessment of hyperparameters and model performance, revealing the minimal model complexity required to achieve the best performance. The assortment of top-performing models harvested from the HPO process allows us to choose a high-performing model with a low inference cost for efficient use in global climate models (GCMs).

97 MATHEMATICS AND COMPUTING↗

Anomalies in the topology of the temperature fluctuations in the cosmic microwave background: An analysis of the NPIPE and FFP10 data releases

We present a topological analysis of the temperature fluctuation maps from the Planck 2020 Data Release 4 NPIPE dataset and the Planck 2018 Data Release 3 FFP10 dataset. We performed a multiscale analysis in terms of the homology characteristics of the maps, invoking relative homology to account for the analysis in the presence of masks. We performed our analysis for a range of smoothing scales spanning sub- and super-horizon scales corresponding to a full width at half maximum (FWHM) of 5',10',20',40',80',160',320', and 640', and employed simulations based on the standard model for comparison, which assumes the initial fluctuation field to be an isotropic and homogeneous Gaussian random field. Examining the behavior of topological components, represented by the 0D homology group, we find the observations to be approximately 2σ or less deviant from the simulations for all resolutions and scales for the NPIPE dataset. For the FFP10 dataset, we detect a 2.96σ deviation between the observations and simulations at N = 128, FWHM = 80'. For the topological loops, represented by the first homology group, the simulations and observations are consistent within 2σ for most resolutions and scales for both the datasets. However, for the NPIPE dataset, we observe a high deviation between the observation and simulations in the number of loops at FWHM = 320', but at a low dimensionless threshold ν = –2.5. Under a Gaussian assumption, this would amount to a deviation of ~4σ. However, the distribution in this bin is manifestly non-Gaussian and does not obey Poisson statistics either. In the absence of a true theoretical understanding, we simply note that the significance is higher than what may be resolved by 600 simulations, yielding an empirical p-value of at most 0.0016. Specifically in this case, our tests indicate that the numbers arise from a statistically stable regime, despite being based on small numbers. For the FFP10 dataset, the differences are not as strong as for the NPIPE dataset, indicating a 2.77σ deviation at this resolution and threshold. The Euler characteristic, which is the alternating sum of the ranks of relative homology groups, reflects the deviations in the components and loops. To assess the significance of combined levels for a given scale, we employed the empirical and theoretical versions of the χ 2 test as well as the nonparametric Tukey depth test. Although all statistics exhibit a stable distribution, we favor the empirical version of the χ 2 test in the final interpretation, as it indicates the most conservative differences. For the NPIPE dataset, we find that the components and loops differ at more than 95%, but agree within the 99% confidence level with respect to the base model at N = 32, FWHM = 320'. The Euler characteristic at this resolution displays a per mil deviation. In contrast, the FFP10 dataset shows that the observations are consistent with the base model within the 95% confidence level, at this and smaller scales. This is consistent with the observations of the Planck analysis pipeline via Minkowski functionals. For the largest smoothing scale, N = 16, FWHM = 640', both datasets exhibit an anomalous behavior of the loops, where FFP10 data exhibit a deviation that is larger by an order of magnitude than that of the NPIPE dataset. In contrast, the values for the topological components and the Euler characteristic agree between observations and model to within a confidence level of 99%. However, for the largest scales, the statistics are based on low numbers and may have to be regarded with caution. Even though both datasets exhibit mild to significant discrepancies, they also exhibit contrasting behaviors at various instances. Therefore, we do not find it feasible to convincingly accept or reject the null hypothesis. Disregarding the large-scale anomalies that persist at similar scales in WMAP and Planck, observations of the cosmic microwave background are largely consistent with the standard cosmological model within 2σ.

79 ASTRONOMY AND ASTROPHYSICS↗