Search NASA⌕ Search

SEARCH · Search NASA

Results for “sampling size”

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 91 records · Page 5
All sourcesNASA NTRSDOE OSTI
DOE OSTI2023

Data augmentation and multimodal learning for predicting drug response in patient-derived xenografts from gene expressions and histology images

Patient-derived xenografts (PDXs) are an appealing platform for preclinical drug studies. A primary challenge in modeling drug response prediction (DRP) with PDXs and neural networks (NNs) is the limited number of drug response samples. We investigate multimodal neural network (MM-Net) and data augmentation for DRP in PDXs. The MM-Net learns to predict response using drug descriptors, gene expressions (GE), and histology whole-slide images (WSIs). We explore whether combining WSIs with GE improves predictions as compared with models that use GE alone. We propose two data augmentation methods which allow us training multimodal and unimodal NNs without changing architectures with a single larger dataset: 1) combine single-drug and drug-pair treatments by homogenizing drug representations, and 2) augment drug-pairs which doubles the sample size of all drug-pair samples. Unimodal NNs which use GE are compared to assess the contribution of data augmentation. The NN that uses the original and the augmented drug-pair treatments as well as single-drug treatments outperforms NNs that ignore either the augmented drug-pairs or the single-drug treatments. In assessing the multimodal learning based on the MCC metric, MM-Net outperforms all the baselines. Our results show that data augmentation and integration of histology images with GE can improve prediction performance of drug response in PDXs.

60 APPLIED LIFE SCIENCES↗
DOE OSTI2026

Characterization of Electronic Stress-Induced Changes in Multilayer MoS 2

Transition metal dichalcogenides like molybdenum disulfide (MoS 2 ) are compelling for next-generation electronic devices. In this work, we investigate the impact of electronic stress on MoS 2 to illustrate that observational and phenomenological information on multiple devices can be useful to describe changes in the device, and caution against the rationalization of paltry results as representative or correlative to device behavior. Here, we stress MoS 2 by applying a sustained 20 V DC bias to study the material’s response. Post-stress electronic characterization revealed nonuniform shifts in current–voltage (I–V) behavior alongside microscale changes. Complementary mechanical, spectroscopic, and scanning microwave impedance measurements showed that stress-induced features locally modulate stiffness, surface potential, Raman intensity, and charge carrier density. We correlated I–V behavior with morphological features (wrinkles, tears, folds, height) and device-level geometry (MoS 2 overlap with electrodes, channel area, contact length) on 50 test structures across five chips to move beyond anecdotal conclusions. We found no universal correlations before DC stress. However, device-level geometry was correlated with I–V behavior after DC stress, suggesting that electrode contacts play a more dominant role than morphology in determining performance. Delamination and thinning induced by DC stress led to localized reductions in charge carrier density within the affected regions. Further, delamination and thinning appear to map to I–V device performance in a few samples, but the correlation is lost when a larger sample size is considered. This suggests significant sample-to-sample variability in surface electronic states of the test structures. We also discuss how environmental factors introduced during fabrication may contribute to the observed heterogeneous device response. Progress will require high-resolution, multimodal analysis across many samples constructed under controlled, clean conditions. By building data sets that capture variability, we can better identify the true drivers of performance.

36 MATERIALS SCIENCE↗
DOE OSTI2023

Defect Characterization of Monocrystalline Silicon Solar Cells with Polysilicon Passivated Contact Using Electrically-Detected Magnetic Resonance (EDMR) Spectroscopy

As the c-Si based solar cell efficiencies are approaching over 26%, it is becoming critical to characterize the low concentrations of the defects – as low as 10^10-10^11 cm-3 (for e.g., iron contamination in high-lifetime Ga-doped wafers3 and n-type wafers), and further reduce them. Also, atomistic level understanding of the mechanisms of the low concentration process-induced-defects and reliability limiting defects (such as light and elevated temperature induced degradation, surface passivation degradation) is needed to design the mitigation strategies. The conventional characterization techniques are limited due to their detection limitations. Some of the techniques based on lifetime spectroscopies can still be used for low concentration characterization however, they are based on estimations and theoretical models and hence, indirect and cannot fully reveal information about the microscopic mechanism of the defects. Thus, we present the application of an ultrasensitive magnetic resonance-based technique for the direct spectroscopic detection of the defects in Si PV - electrically detected magnetic resonance (EDMR). In this work, we aim to focus on establishing a process flow for fabrication of minicells with (miniature replica of the larger-area cells) and setting up the routine for EDMR measurements on them with the EDMR instrumentation capability at NREL. For the EDMR measurements, sample size is limited by the dimensions of sample holder tube (width less than 3.2 mm, active area - 20 mm). Thus, we have designed c-Si based minicells with polysilicon (poly-Si) passivated contacts same as the larger-area cells that we fabricate in our group at NREL. We also modified our minicell process flow for fabricating the textured minicells for preserving the texture during processing and taking care of the laser-ablation edge damage which can significantly affect the performance of such small devices. We have achieved comparable performance on these newly fabricated minicells as that of our 4 cm2 devices with same structure (comparable VOC, JSC, FF). We also conducted EDMR measurements on the minicells and observed a distinct EDMR signal at g-value ~2.005 at temperatures 30K and above, as shown in Fig. 2. We associate this signal to the presence of silicon dangling bonds based on the g-value. We also observed an EDMR signal at g-value ~1.998 at temperature ~5K. The origin of this signal is still being investigated. Thus, we show the proof of concept of minicells and EDMR measurements with which we now aim to study some of the unknown defects in silicon solar cell devices.

EDMR↗
DOE OSTI2023

Heat capacity of microgram oxide samples by fast scanning calorimetry

Quantitative scanning calorimetry on microgram-sized samples opens a broad, new range of opportunities for studying the thermodynamic properties of quantity-limited materials, including those produced under extreme conditions or found as rare accessory minerals in nature. We calibrated the Mettler Toledo Flash DSC 2+ calorimeter to obtain quantitative heat capacities in the range 200–350 °C, using samples weighing between 2 and 11.5 μg. Our technique is applied to a new set of oxide materials to which it has never been used before, without the need for melting, glass transitions, or phase transformations. Heat capacity data were obtained for silica in the high pressure stishovite (rutile) structure, dense post-stishovite glass, standard fused quartz, and for TiO2 rutile. These heat capacities agree within 5%–15% with the literature values reported for rutile, stishovite, and fused SiO2 glass. The heat capacity of post-stishovite glass, made by heating stishovite to 1000 °C, is a newly reported value. After accurate calibrations, measured heat capacities were then used to calculate masses for samples in the microgram range, a substantial improvement over measurement in conventional microbalances, which have uncertainties approaching 50%–100% for such small samples. Since the typical uncertainty of heat capacities measured on 10–100 mg samples in conventional differential scanning calorimetry is typically 7% (1%–5% with careful work), flash differential scanning calorimetry, using samples a factor of 1000 smaller, increases the uncertainty of heat capacity measurements by a factor of <3, opening the door for meaningful measurements on ultra-small, high-pressure samples and other quantity-limited materials.

Instruments & Instrumentation↗
DOE OSTI2022

X-ray fluorescence microscopy methods for biological tissues

Abstract Synchrotron-based X-ray fluorescence microscopy is a flexible tool for identifying the distribution of trace elements in biological specimens across a broad range of sample sizes. The technique is not particularly limited by sample type and can be performed on ancient fossils, fixed or fresh tissue specimens, and in some cases even live tissue and live cells can be studied. The technique can also be expanded to provide chemical specificity to elemental maps, either at individual points of interest in a map or across a large field of view. While virtually any sample type can be characterized with X-ray fluorescence microscopy, common biological sample preparation methods (often borrowed from other fields, such as histology) can lead to unforeseen pitfalls, resulting in altered element distributions and concentrations. A general overview of sample preparation and data-acquisition methods for X-ray fluorescence microscopy is presented, along with outlining the general approach for applying this technique to a new field of investigation for prospective new users. Considerations for improving data acquisition and quality are reviewed as well as the effects of sample preparation, with a particular focus on soft tissues. The effects of common sample pretreatment steps as well as the underlying factors that govern which, and to what extent, specific elements are likely to be altered are reviewed along with common artifacts observed in X-ray fluorescence microscopy data.

Pushie, M. Jake (ORCID:0000000174945427)↗
DOE OSTI2020

Reconciling Chord Length Distributions and Area Distributions for Fields of Fractal Cumulus Clouds

While the total cover of broken cloud fields can in principle be obtained from one-dimensional measurements, the cloud size distribution normally differs between two-dimensional (area) and one-dimensional retrieval (chord length) methods. In this study, we use output from high-resolution Large Eddy Simulations to generate a transfer function between the two. We retrieve chord lengths and areas for many clouds, and plot the one as a function of the other, and vice versa. We find that the cloud area distribution conditional on the chord length behaves like a gamma distribution with well-behaved parameters, with a mean μ=1.1L and a shape parameter β=L−0.645. Using this information, we are able to generate a transfer function that can adjust the chord length distribution so that it comes much closer to the cloud area distribution. Our transfer function improves the error in predicting the mean cloud size, and is performs without strong biases for smaller sample sizes. However, we find that the method is still has difficulties in accurately predicting the frequency of occurrence of the largest cloud sizes.

54 ENVIRONMENTAL SCIENCES↗
DOE OSTI2021

A review of imputation strategies for isobaric labeling-based shotgun proteomics

The throughput efficiency and increased depth of coverage provided by isobaric-labeled proteomics measurements have led to increased usage of these techniques. However, the structure of missing data is uniquely different than unlabeled studies. In this review, we compare the efficacy of nine imputation methods on a CPTAC proteomics iTRAQ dataset. Imputation methods were evaluated with regard to accuracy, variability, statistical hypothesis test inference and run time over datasets consisting of varying number of iTRAQ plexes and percentages of missing data. In general, expectation maximization and random forest imputation methods yielded the best performances, and constant-based methods performed poorly consistently across all dataset sizes and percentages of missing values. For datasets with small sample sizes and higher percentages of missing data, results indicate that statistical inference with no imputation may be preferable. Based on the findings in this review, there are core imputation methods that perform higher for isobaric-labeled proteomics data, but great care and consideration as to whether imputation should be used should be given for datasets comprised of a small number of samples, as well as to factors such as computational time and reproducibility of imputation values.

Bramer, Lisa M.↗
DOE OSTI2023

Efficient Bayesian inference with latent Hamiltonian neural networks in No-U-Turn Sampling

When sampling for Bayesian inference, one popular approach in the computational field is to use Hamiltonian Monte Carlo (HMC) and specifically the No-U-Turn Sampler (NUTS), which automatically decides the end time of the Hamiltonian trajectory. However, HMC and NUTS can require numerous numerical gradients of the target density and can prove slow in practice when relying on computationally expensive forward models. We propose Latent Hamiltonian neural networks (L-HNNs) with HMC and NUTS for solving Bayesian inference problems. Once trained, L-HNNs do not require numerical gradients of the target density during sampling, and hence numerous evaluations of the forward computational model. Moreover, L-HNNs satisfy important properties such as perfect time reversibility and Hamiltonian conservation, making them well-suited for use within HMC and NUTS because stationarity can be shown. We also propose the integration of L-HNNs in an online error monitoring scheme, in which numerical gradients of the target density are used for a few samples whenever the L-HNNs prediction errors are large. This online error monitor scheme prevents sample degeneracy in regions of low probability density and ensures robust uncertainty quantification. We demonstrate L-HNNs in NUTS with online error monitoring on several analytical examples involving complex, heavy-tailed, and high-local-curvature probability densities. We then demonstrate the applicability of L-HNNs in NUTS to two computational case studies, namely the Allen-Cahn stochastic partial differential equation and an elliptic partial differential equation with 25 and 50 inference parameters, respectively. Overall, the L-HNNs in NUTS with online error monitoring satisfactorily inferred these probability densities. In conclusion, compared to traditional NUTS, L-HNNs in NUTS with online error monitoring required 1–2 orders of magnitude fewer numerical gradients of the target density and improved the effective sample size (ESS) per gradient (which is a measure of both the sampling quality and the computational expense) by an order of magnitude.

97 MATHEMATICS AND COMPUTING↗
DOE OSTI2020

Finite-size effects in lead scandium tantalate relaxor thin films

Large electromechanical effects in relaxor ferroelectrics are generally attributed to the collective response of an ensemble of correlated, nanometer-sized polar structures induced by chemical and charge disorder. Here, we study finite-size effects on such polar order (i.e., how it evolves when sample dimensions approach the polarization correlation length) in 7–70-nm-thick films of the relaxor ferroelectric PbSc 0.5 Ta 0.5 O 3 . Temperature-dependent polarization studies reveal a linear suppression of the polarization and nonlinearity associated with relaxor order as the film thickness decreases to ≈ 30 nm. Below this thickness, however, the suppression rapidly accelerates, and polarization is completely absent by film thicknesses of ≈ 7 nm, despite the continued observation of a broad peak in dielectric permittivity and frequency dispersion. Diffuse-scattering measurements reveal the diffuse-scattering symmetry, and analysis suggests the films have a polarization correlation length of ≈ 23 nm. Finally, taken together, it is apparent that reduction of sample size and the resulting distribution of polar structures drive suppression and eventual quenching of the electrical response of relaxors, which may be attributed to increasing dipole-dipole and dipole-interface interactions.

75 CONDENSED MATTER PHYSICS, SUPERCONDUCTIVITY AND↗
DOE OSTI2021

Learning curves for drug response prediction in cancer cell lines

Motivated by the size and availability of cell line drug sensitivity data, researchers have been developing machine learning (ML) models for predicting drug response to advance cancer treatment. As drug sensitivity studies continue generating drug response data, a common question is whether the generalization performance of existing prediction models can be further improved with more training data. We utilize empirical learning curves for evaluating and comparing the data scaling properties of two neural networks (NNs) and two gradient boosting decision tree (GBDT) models trained on four cell line drug screening datasets. The learning curves are accurately fitted to a power law model, providing a framework for assessing the data scaling behavior of these models. The curves demonstrate that no single model dominates in terms of prediction performance across all datasets and training sizes, thus suggesting that the actual shape of these curves depends on the unique pair of an ML model and a dataset. The multi-input NN (mNN), in which gene expressions of cancer cells and molecular drug descriptors are input into separate subnetworks, outperforms a single-input NN (sNN), where the cell and drug features are concatenated for the input layer. In contrast, a GBDT with hyperparameter tuning exhibits superior performance as compared with both NNs at the lower range of training set sizes for two of the tested datasets, whereas the mNN consistently performs better at the higher range of training sizes. Moreover, the trajectory of the curves suggests that increasing the sample size is expected to further improve prediction scores of both NNs. These observations demonstrate the benefit of using learning curves to evaluate prediction models, providing a broader perspective on the overall data scaling characteristics. A fitted power law learning curve provides a forward-looking metric for analyzing prediction performance and can serve as a co-design tool to guide experimental biologists and computational scientists in the design of future experiments in prospective research studies.

60 APPLIED LIFE SCIENCES↗
DOE OSTI2024

Validation of the Scientific Program for the Dark Energy Spectroscopic Instrument

The Dark Energy Spectroscopic Instrument (DESI) was designed to conduct a survey covering 14,000 deg$^{2}$ over 5 yr to constrain the cosmic expansion history through precise measurements of baryon acoustic oscillations (BAO). The scientific program for DESI was evaluated during a 5 month survey validation (SV) campaign before beginning full operations. This program produced deep spectra of tens of thousands of objects from each of the stellar Milky Way Survey (MWS), Bright Galaxy Survey (BGS), luminous red galaxy (LRG), emission line galaxy (ELG), and quasar target classes. These SV spectra were used to optimize redshift distributions, characterize exposure times, determine calibration procedures, and assess observational overheads for the 5 yr program. In this paper, we present the final target selection algorithms, redshift distributions, and projected cosmology constraints resulting from those studies. We also present a One-Percent Survey conducted at the conclusion of SV covering 140 deg$^{2}$ using the final target selection algorithms with exposures of a depth typical of the main survey. The SV indicates that DESI will be able to complete the full 14,000 deg$^{2}$ program with spectroscopically confirmed targets from the MWS, BGS, LRG, ELG, and quasar programs with total sample sizes of 7.2, 13.8, 7.46, 15.7, and 2.87 million, respectively. These samples will allow exploration of the Milky Way halo, clustering on all scales, and BAO measurements with a statistical precision of 0.28% over the redshift interval z < 1.1, 0.39% over the redshift interval 1.1 < z < 1.9, and 0.46% over the redshift interval 1.9 < z < 3.5.

79 ASTRONOMY AND ASTROPHYSICS↗
DOE OSTI2022

An Easy‐to‐Use Custom‐Built Cell for Neutron Powder Diffraction Studies of Rechargeable Batteries

Abstract In operando powder diffraction remains one of the most powerful tools for non‐destructive investigation of battery electrode materials. While in operando X‐ray, especially synchrotron radiation, powder diffraction is by now a routine experimental technique, in operando neutron powder diffraction is still less established. We present a new electrochemical cell for in operando neutron powder diffraction, which is, first and foremost, easy to use, but can also cycle electrode materials under electrochemical conditions close to those achieved using standard laboratory cells. The cell has been designed in multiple sizes, and high‐quality electrochemical and neutron powder diffraction data is presented for sample sizes as low as 48 mg total active material. The cell handles lithium‐ion and sodium‐ion materials equally well, with no difference in how the cell is prepared and assembled. The cell is intended to be used as sample environment at powder diffractometers at the neutron facilities MLZ, ORNL and ACNS.

25 ENERGY STORAGE↗
DOE OSTI2020

Additional Analysis of the 2H-Evaporator Wall Scale Sample HTF-17-57

Savannah River National Laboratory previously analyzed scale samples from both the wall and cone sections of the 242-16H Evaporator prior to chemical cleaning. The samples were analyzed for uranium and plutonium isotopes required for a Nuclear Criticality Safety Assessment of the scale removal process. The analysis of the scale samples found the material to contain crystalline nitrated cancrinite and clarkeite and significant amounts of mercury. Savannah River Remediation subsequently requested additional analyses on the samples. Inspection of the sample bottles revealed that no sample remained of the cone scale (HTF-17-56) while approximately 4 g of the wall scale sample (HTF-17-57) remained. The wall sample was subsequently ground with a mortar and pestle, at which time it was discovered that the sample contained a significant amount of liquid, metallic mercury. The portion of the sample that was not mercury was approximately 3 g. This portion of the sample was analyzed to determine the concentrations of selected radionuclides. The following radionuclide concentrations were measured: <9.6E+02 dpm/g H-3, 2.9E+02 dpm/g C-14, 5.6E+03 dpm/g Tc-99, and 1.6E+02 dpm/g I-129. Blank samples were also analyzed to evaluate sample contamination from the Shielded Cells environment. All blank sample results were below detectable limits. Due to the limited sample size and the anticipated low levels of the target radionuclides, no replicate sample analysis was conducted.

12 MANAGEMENT OF RADIOACTIVE AND NON-RADIOACTIVE W↗
DOE OSTI2024

Utilization of Synthetic Near-Infrared Spectra via Generative Adversarial Network to Improve Wood Stiffness Prediction

Near-infrared (NIR) spectroscopy is widely used as a nondestructive evaluation (NDE) tool for predicting wood properties. When deploying NIR models, one faces challenges in ensuring representative training data, which large datasets can mitigate but often at a significant cost. Machine learning and deep learning NIR models are at an even greater disadvantage because they typically require higher sample sizes for training. In this study, NIR spectra were collected to predict the modulus of elasticity (MOE) of southern pine lumber (training set = 573 samples, testing set = 145 samples). To account for the limited size of the training data, this study employed a generative adversarial network (GAN) to generate synthetic NIR spectra. The training dataset was fed into a GAN to generate 313, 573, and 1000 synthetic spectra. The original and enhanced datasets were used to train artificial neural networks (ANNs), convolutional neural networks (CNNs), and light gradient boosting machines (LGBMs) for MOE prediction. Overall, results showed that data augmentation using GAN improved the coefficient of determination (R 2 ) by up to 7.02% and reduced the error of predictions by up to 4.29%. ANNs and CNNs benefited more from synthetic spectra than LGBMs, which only yielded slight improvement. All models showed optimal performance when 313 synthetic spectra were added to the original training data; further additions did not improve model performance because the quality of the datapoints generated by GAN beyond a certain threshold is poor, and one of the main reasons for this can be the size of the initial training data fed into the GAN. LGBMs showed superior performances than ANNs and CNNs on both the original and enhanced training datasets, which highlights the significance of selecting an appropriate machine learning or deep learning model for NIR spectral-data analysis. The results highlighted the positive impact of GAN on the predictive performance of models utilizing NIR spectroscopy as an NDE technique and monitoring tool for wood mechanical-property evaluation. Further studies should investigate the impact of the initial size of training data, the optimal number of generated synthetic spectra, and machine learning or deep learning models that could benefit more from data augmentation using GANs.

59 BASIC BIOLOGICAL SCIENCES↗
DOE OSTI2021

CIS Project 22359, Final Technical Report. Discretized Posterior Approximation in High Dimensions

Our primary aim in this work is to understand how to efficiently obtain reliable uncertainty quantification in automatic learning algorithms with limited training datasets. Standard approaches rely on cross-validation to tune hyper parameters. Unfortunately, when our datasets are too small, holdout datasets become unreliable—albeit unbiased—measures of prediction quality due to the lack of adequate sample size. We should not place confidence in holdout estimators under conditions wherein the sample variance is both large and unknown. More poigniantly, our training experiments on limited data (Duersch and Catanach, 2021) show that even if we could improve estimator quality under these conditions, the typical training trajectory may never even encounter generalizable models.

97 MATHEMATICS AND COMPUTING↗
DOE OSTI2025

Efficient Monte Carlo event generation for neutrino-nucleus exclusive cross sections

Modern neutrino-nucleus cross section computations need to incorporate sophisticated nuclear models to achieve greater predictive precision. However, the computational complexity of these advanced models often limits their practicality for experimental analyses. To address this challenge, we introduce a new Monte Carlo method utilizing normalizing flows to generate surrogate cross sections that closely approximate those of the original model while significantly reducing computational overhead. As a case study, we built a Monte Carlo event generator for the neutrino-nucleus cross section model developed by the Ghent group. This model employs a Hartree-Fock procedure to establish a quantum mechanical framework in which both the bound and scattering nucleon states are solutions to the mean-field nuclear potential. The surrogate cross sections generated by our method demonstrate excellent accuracy with a relative effective sample size of more than 98.4%, providing a computationally efficient alternative to traditional Monte Carlo sampling methods for differential cross sections.

72 PHYSICS OF ELEMENTARY PARTICLES AND FIELDS↗
DOE OSTI2025

Optical 14 C Tracing for Biological and Pharmaceutical Applications Using Two-Color Cavity Ringdown Spectroscopy

Laser-based 14 C quantitation has been proposed as a more affordable, higher-throughput, table-top alternative to accelerator mass spectrometry (AMS). Here, we demonstrate the feasibility of a mid-IR 14 C detector based on two-color cavity ringdown spectroscopy (2C-CRDS) for low-level 14 C isotope tracing in biological studies. The 2C-CRDS technique quantifies the sample 14 C content by measuring the 14 CO 2 absorption signals from the combusted samples with mid-IR lasers. With 2C-CRDS, we previously demonstrated the most sensitive and accurate optical measurements of 14 CO 2 . The current detection sensitivity and quantitation accuracy of the instrument, at a few parts per quadrillion (where a quadrillion = 10 15 ) 14 C/C mole fraction, is competitive against AMS. Here, by applying the 2C-CRDS 14 C sensor to two applications relevant to 14 C-labeled biochemical analysis and pharmaceutical studies, we demonstrate sub-fCi level (where 1 fCi = 10 –15 Ci) quantitation of sample 14 C activity, with a minimum sample-size requirement of 3 mg of carbon. The current measurement throughput, ~25 min/sample, is largely limited by the sampling efficiency of the online combustion and CO 2 processing interface to the 2C-CRDS instrument. The possibility of a significantly improved measurement throughput of a few minutes per sample is suggested by the results of a flow-through 14 CO 2 sampling scheme. In conclusion, this improved measurement efficiency, combined with the relatively low cost and compact size of a 2C-CRDS sensor, could potentially revolutionize high-sensitivity 14 C tracing in biological, pharmaceutical, and clinical studies.

60 APPLIED LIFE SCIENCES↗
DOE OSTI2021

Deep Learning Estimation of Daily Ground–Level NO 2 Concentrations from Remote Sensing Data

The limited number of nitrogen dioxide (NO 2 ) surface measurements calls for the development of highly accurate approaches to estimating surface NO 2 concentrations. In this study, we leverage a new satellite instrument, the TROPOspheric Monitoring Instrument (TROPOMI), along with other predictor variables, to estimate daily surface NO 2 concentrations over Texas in 2019. We use the deep convolutional neural network (Deep-CNN), an advanced deep learning algorithm, to obtain estimates and achieve a correlation coefficient (R) of 0.91, an index of agreement (IOA) of 0.95, and a mean absolute bias (MAB) of 1.75 ppb in surface NO 2 estimation. Additionally, we leverage a novel approach, SHapley Additive exPlanations (SHAP), to describe how Deep-CNN understands each predictor variable. The SHAP results show that the Deep-CNN model has an advanced understanding of the dataset, revealing that TROPOMI closely captures levels of NO 2 . In addition, we show the superiority of our Deep-CNN model at estimating surface NO 2 over other well-known machine learning and regression models in the field, including the support vector machines (SVM), random forest (RF), and multiple linear regression (MLR). Although SVM and RF show strong capabilities at estimating surface NO 2 concentrations, their accuracy is inferior to that of the Deep-CNN model, ranking second and third in model accuracy in this study. The MLR, however, shows a poor ability at NO 2 estimation and ranks last among all models. Furthermore, testing the impact of sample size on model performance, we also show that, compared to other models, Deep-CNN needs more samples to trigger its strength at surface NO 2 estimation.

54 ENVIRONMENTAL SCIENCES↗