SEARCH · Engineering Papers

Results for “Training Analysis”

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.

SourceResearch connectionFrom yearThrough year

Quote a phrase for an exact phrase match. Source license links do not imply unrestricted reuse.

At least 1,567 records · Page 87

All sources NASA NTRS DOE OSTI

DOE OSTI2024

Physics-inspired spatiotemporal-graph AI ensemble for the detection of higher order wave mode signals of spinning binary black hole mergers

We present a new class of AI models for the detection of quasi-circular, spinning, non-precessing binary black hole mergers whose waveforms include the higher order gravitational wave modes ($\ell$, |m|) = {(2,2), (2,1), (3,3), (3,2), (4,4)}, and mode mixing effects in the $\ell$ = 3, |m| = 2 harmonics. These AI models combine hybrid dilated convolution neural networks to accurately model both short- and long-range temporal sequential information of gravitational waves; and graph neural networks to capture spatial correlations among gravitational wave observatories to consistently describe and identify the presence of a signal in a three detector network encompassing the Advanced LIGO and Virgo detectors. We first trained these spatiotemporal-graph AI models using synthetic noise, using 1.2 million modeled waveforms to densely sample this signal manifold, within 1.7 h using 256 NVIDIA A100 GPUs in the Polaris supercomputer at the Argonne Leadership Computing Facility. This distributed training approach exhibited optimal classification performance, and strong scaling up to 512 NVIDIA A100 GPUs. With these AI ensembles we processed data from a three detector network, and found that an ensemble of 4 AI models achieves state-of-the-art performance for signal detection, and reports two misclassifications for every decade of searched data. We distributed AI inference over 128 GPUs in the Polaris supercomputer and 128 nodes in the Theta supercomputer, and completed the processing of a decade of gravitational wave data from a three detector network within 3.5 h. Finally, we fine-tuned these AI ensembles to process the entire month of February 2020, which is part of the O3b LIGO/Virgo observation run, and found 6 gravitational waves, concurrently identified in Advanced LIGO and Advanced Virgo data, and zero false positives. This analysis was completed in one hour using one NVIDIA A100 GPU.

79 ASTRONOMY AND ASTROPHYSICS↗

DOE OSTI2021

TX$^2$: Transformer eXplainability and eXploration

The Transformer eXplainability and eXploration (Martindale & Stewart, 2021), or TX 2 software package, is a library designed for artificial intelligence researchers to better understand the performance of transformer models (Vaswani et al., 2017) used for sequence classification. The tool is capable of integrating with a trained transformer model and a dataset split into training and testing populations to produce an ipywidget (Project Jupyter Contributors, 2021) dashboard with a number of visualizations to understand model performance with an emphasis on explainability and interpretability. The TX 2 package is primarily intended to integrate into a workflow centered around Jupyter Notebooks (Kluyver et al., 2016), and currently assumes the use of PyTorch (Paszke et al., 2019) and Hugging Face transformers library (Wolf et al., 2020). The dashboard includes visualization and data exploration features to aid researchers, including an interactive UMAP embedding graph (McInnes et al., 2018) to understand classification clusters, a word salience map that can be updated as researchers alter textual entries in near real time, a set of tools to understand word frequency and importance based on the clusters in the UMAP embedding graph, and a set of traditional confusion matrix analysis tools.

97 MATHEMATICS AND COMPUTING↗

DOE OSTI2024

CMOS-Based Single-Cycle in-Memory XOR/XNOR

Big data applications are on the rise, and so is the number of data centers. The ever-increasing massive data pool needs to be periodically backed up in a secure environment. Moreover, a massive amount of securely backed-up data is required for training binary convolutional neural networks for image classification. XOR and XNOR operations are essential for large-scale data copy verification, encryption, and classification algorithms. The disproportionate speed of existing compute and memory units makes the von Neumann architecture inefficient to perform these Boolean operations. Compute-in-memory (CiM) has proved to be an optimum approach for such bulk computations. The existing CiM-based XOR/XNOR techniques either require multiple cycles for computing or add to the complexity of the fabrication process. Here, we propose a CMOS-based hardware topology for single-cycle in-memory XOR/XNOR operations. Our design provides at least 2× improvement in the latency compared with other existing CMOS-compatible solutions. We verify the proposed system through circuit/system-level simulations and evaluate its robustness using a 5000-point Monte Carlo variation analysis. This all-CMOS design paves the way for practical implementation of CiM XOR/XNOR at scaled technology nodes.

97 MATHEMATICS AND COMPUTING↗

DOE OSTI2026

Optimal Client Sampling in Federated Learning with Client-level Heterogeneous Differential Privacy

Federated Learning with client-level differential privacy (DP) provides a promising framework for collaboratively training models while rigorously protecting clients’ privacy. However, classic approaches like DP-FedAvg struggle when clients have heterogeneous privacy requirements, as they must uniformly enforce the strictest privacy level across all clients, leading to excessive DP noise and significant degradation in model utility. Existing methods to improve the model utility in such heterogeneous privacy settings often assume a trusted server and are largely heuristic, resulting in suboptimal performance and lacking strong theoretical foundations. Here, in this work, we address these challenges under a practical attack model where both clients and the server are honest-but-curious. We propose GDPFed, which partitions clients into groups based on their privacy budgets and achieves client-level DP within each group to reduce the privacy budget waste and hence improve the model utility. Based on the privacy and convergence analysis of GDPFed, we find that the magnitude of DP noise depends on both model dimensionality and the per-group client sampling ratios. To further improve the performance of GDPFed, we introduce GDPFed+, which integrates model sparsification to eliminate unnecessary noise and optimizes per-group client sampling ratios to minimize convergence error. Extensive empirical evaluations on multiple benchmark datasets demonstrate the effectiveness of GDPFed+, showing substantial performance gains compared with state-of-the-art methods.

Xu, Jiahao [Univ. of Nevada, Reno, NV (United Stat↗

DOE OSTI2023

FY23 Progress Report: SRNL Analysis of ICCWR LCM and WAMS data for Corrosion and Cracking

12 MANAGEMENT OF RADIOACTIVE AND NON-RADIOACTIVE W↗

DOE OSTI2020

Learning Spectral Templates for Photometric Redshift Estimation from Broadband Photometry

Estimating redshifts from broadband photometry is often limited by how accurately we can map the colors of galaxies to an underlying spectral template. Current techniques utilize spectrophotometric samples of galaxies or spectra derived from spectral synthesis models. Both of these approaches have their limitations: either the sample sizes are small and often not representative of the diversity of galaxy colors, or the model colors can be biased (often as a function of wavelength), which introduces systematics in the derived redshifts. In this paper, we learn the underlying spectral energy distributions from an ensemble of ∼100 K galaxies with measured redshifts and colors. We show that we are able to reconstruct emission and absorption lines at a significantly higher resolution than the broadband filters used to measure the photometry for a sample of 20 spectral templates. We find that our training algorithm reduces the fraction of outliers in the derived photometric redshifts by up to 28%, bias up to 91%, and scatter up to 25%, when compared to estimates using a standard set of spectral templates. We discuss the current limitations of this approach and its applicability for recovering the underlying properties of galaxies. Our derived templates and the code used to produce these results are publicly available in a dedicated Github repository: https://github.com/dirac-institute/photoz-template-learning.

79 ASTRONOMY AND ASTROPHYSICS↗

DOE OSTI2024

Improved Subseasonal Forecasting of Extreme Polar Vortices Using Machine Learning

Our research was focused on forecasting the position and shape of the winter stratospheric polar vortex at a subseasonal timescale of 15 days in advance. To achieve this, we employed both statistical and neural network machine learning techniques. The analysis was performed on 42 winter seasons of reanalysis data provided by NASA giving us a total of 6,342 days of data. The state of the polar vortex for determined by using geometric moments to calculate the centroid latitude and the aspect ratio of an ellipse fit onto the vortex. Timeseries for thirty additional precursors were calculated to help improve the predictive capabilities of the algorithm. Feature importance of these precursors was performed using random forest to measure the predictive importance and the ideal number of precursors. Then, using the precursors identified as important, various statistical methods were tested for predictive accuracy with random forest and nearest neighbor performing the best. An echo state network, a type of recurrent neural network that features sparsely connected hidden layer and a reduced number of trainable parameters that allows for rapid training and testing, was also implemented for the forecasting problem. Hyperparameter tuning was performed for each methods using a subset of the training data. The algorithms were trained and tuned on the first 41 years of data, then tested for accuracy on the final year. In general, the centroid latitude of the polar vortex proved easier to predict than the aspect ratio across all algorithms. Random forest outperformed other statistical forecasting algorithms overall but struggled to predict extreme values. Forecasting from echo state network suggested a strong predictive capability past 15 days, but further work is required to fully realize the potential of recurrent neural network approaches.

54 ENVIRONMENTAL SCIENCES↗

DOE OSTI2024

FY24 Progress Report: SRNL Analysis of ICCWR LCM and WAMS data for Corrosion and Cracking

Algorithms for Machine Learning (ML) and data analysis for the 3013 Surveillance Program have been developed in an ongoing collaborative effort by the Savannah River National Laboratory (SRNL) and the University of South Carolina (USC). The objective of the algorithms is to automate the identification of corrosion and crack formation in the Inner Container Closure Weld Region (ICCWR) of the canister system used to store Pu-bearing material. Data for corrosion and cracking is collected from large binary files generated by a Laser Confocal Microscope (LCM), the Wide Area 3D Measurement System (WAMS), or,in a recent proposal, by a Scanning Electron Microscope (SEM). The ML software uses the physical attributes in the data files (e.g., one or more of: height, color, and 16-bit grayscale values as functions of position in a plane projection) to detect signs of surface corrosion and cracking after being trained on similar data, with the features to be detected. Although the initial scope included screening for broader indicators of corrosion, e.g., pitting, identification of potential cracks was prioritized for the past several years at the request of program leadership. Labeled training data is essential to developing the ML algorithm, and enhancements to data labeling capability have been developed to address this essential precursor to application of ML routines. Efficient labeling is particularly important in view of the large volume of data required to train ML algorithms and the relative rarity of cracks in the ICCWR data set. The updated program will read binary data from either LCM, WAMS or SEM files, interrogate data attributes, facilitate user labeling of data for training ML algorithms, execute ML algorithms, output parameters from trained ML algorithms, report ML model accuracy with respect to labeled data, and generate graphical representations for various analyses. In FY24, hourglass neural networks (HNNs) that were initiated in FY22 were further developed and tested using available LCM data, and their performance was tested against that of the alternative U-Net Neural Network algorithm structure. HNNs along with previously developed Convolutional Neural Networks (CNNs) and Deep Neural Networks (DNNs) comprise a suite of ML tools for identification of cracks in the ICCWR

12 MANAGEMENT OF RADIOACTIVE AND NON-RADIOACTIVE W↗

DOE OSTI2024

Deep-Learning-Based Segmentation of Keyhole in In-Situ X-ray Imaging of Laser Powder Bed Fusion

In laser powder bed fusion processes, keyholes are the gaseous cavities formed where laser interacts with metal, and their morphologies play an important role in defect formation and the final product quality. The in-situ X-ray imaging technique can monitor the keyhole dynamics from the side and capture keyhole shapes in the X-ray image stream. Keyhole shapes in X-ray images are then often labeled by humans for analysis, which increasingly involves attempting to correlate keyhole shapes with defects using machine learning. However, such labeling is tedious, time-consuming, error-prone, and cannot be scaled to large data sets. To use keyhole shapes more readily as the input to machine learning methods, an automatic tool to identify keyhole regions is desirable. In this paper, a deep-learning-based computer vision tool that can automatically segment keyhole shapes out of X-ray images is presented. The pipeline contains a filtering method and an implementation of the BASNet deep learning model to semantically segment the keyhole morphologies out of X-ray images. The presented tool shows promising average accuracy of 91.24% for keyhole area, and 92.81% for boundary shape, for a range of test dataset conditions in Al6061 (and one AliSi10Mg) alloys, with 300 training images/labels and 100 testing images for each trial. Prospective users may apply the presently trained tool or a retrained version following the approach used here to automatically label keyhole shapes in large image sets.

36 MATERIALS SCIENCE↗

DOE OSTI2020

Thermal Model Heat Rate Predictions of the AGR-5/6/7 Experiment

Several fuel and material irradiation experiments have been planned for the U.S. Department of Energy Advanced Gas Reactor Fuel Development and Qualification Program, which supports the development and qualification of tristructural isotropic (TRISO) coated particle fuel for use in high-temperature gas-cooled reactors. The goals of these experiments are to provide irradiation performance data to support fuel process development, qualify fuel for normal operating conditions, support development of fuel performance models and codes, and provide irradiated fuel and materials for post-irradiation examination and safety testing. Originally planned and named as separate fuel experiments, but subsequently combined into a single test train. AGR-5/6/7 will test low-enriched uranium oxycarbide (UCO – a heterogeneous mixture of uranium oxide and uranium carbide) TRISO fuel. The AGR-5/6 portion of the experiment will provide data to support qualification of the selected reference fuel design, while the AGR-7 portion will serve as a margin test, irradiating the fuel beyond normal operating conditions. During irradiation, the temperature in each capsule is controlled by varying the helium / neon gas mixture flowing through the capsule until the desired thermocouple temperature is reached. Further adjustments can be made by changing the lobe power and the neutron filter. Three neutron filters are available to use as the experiment burns up the fissionable fuel. A thermal finite element model has been created for the five capsules comprising the AGR-5/6/7 experiment. Heat rates from a physics analysis are imported into the model along with fast neutron fluence. Graphite shrinkage due to the fast neutron fluence is incorporated into the model. Gas gaps change as a function of fast neutron fluence. Radiation heat transfer is a major contributor in this model. This is a large model with more than 1 million finite element brick elements. More than 150 different parts are modeled in the finite element model and communicate with each other from a heat transfer sense. More than 50 thermocouples are used in the experiment and are compared to actual measurements. The experiment is composed of five separate stainless steel capsules all welded together. This paper presents a method used to predict the gas mixture in the event that all of the thermocouples fail in a particular capsule. Curve fitting of heat rates from previous cycles and projecting them into the future is the basis of the method.

36 MATERIALS SCIENCE↗

DOE OSTI2022

Machine learning approaches to streamline and enhance the analysis of multiscale imaging data for bioaerosol and soil particles

Bioaerosol and soil particles are ubiquitous in the environment. They are multicomponent and complex in nature displaying mixed inorganic and organic components. The way components are mixed in a bioaerosol sample is referred to as its mixing state. Soil particles are also a mixture of inorganic (mineral) and organic (soil organic matter) components. Bioaerosol particles contribute to a major fraction of coarse mode atmospheric particles, especially in the tropical areas, contributing up to 80 % of the particle mass concentration. The mixing state of particles is crucial to evaluate because it impacts several important environmental processes such as warm and cold cloud formation and radiation budget. Mixing states in aerosols are accompanied by chemical reactions across solid-liquid-gas interfaces. In this study, we utilized elemental compositions and microcopy images of thousands of atmospheric particles acquired by computer-controlled scanning electron microscope equipped with an energy-dispersive x-ray spectrometer to compute the mixing state of atmospheric particles. A 2D convolutional neural network (CNN), also known as convnet, was used to model the relationship between low resolution imaging data and higher resolution spectroscopy data, with the former as training input and the latter as target output. Two types of CNNs were implemented and tested; a basic CNN and an Inception-v3 network. For binary classification, the basic CNN achieved an accuracy of 84.29 % across all atom types, and the Inception-v3-like network achieved an accuracy 85.51 %. This study demonstrates the applicability of deep learning to handle large amounts of imaging/chemical spectroscopy data efficiently and evaluate particle mixing state from a range of environmental samples.

54 ENVIRONMENTAL SCIENCES↗

DOE OSTI2025

E-scooter safety: How attitudinal factors influence risky behavior among shared e-scooter riders

In recent years, e-scooter usage for short-distance trips has grown rapidly. This surge in e-scooter use, combined with the high exposure of e-scooter riders to accident risk, has sparked concerns regarding e-scooter safety. Despite some studies focusing on e-scooter safety, little is known about how attitudinal factors lead e-scooter riders to engage in risky riding behaviors. In this paper, we developed a survey-based empirical model to identify the attitudinal factors influencing engagement in risky behaviors among e-scooter users. We used survey data collected from 420 shared e-scooter users in Chicago in 2022. The survey showed that 47.7% of respondents had experienced at least one collision or fall-off while riding e-scooters. We employed the Partial Least Squares Structural Equation Model (PLS-SEM) to examine the relationships between latent attitudinal factors and risky behavior engagement. Moreover, we conducted Permutation Multi-group Analysis (PMGA) to assess the moderating effect of socio-demographic factors within the estimated model. The findings suggest that riders’ unsafe riding attitude and riding confidence are the most influential factors shaping their risky behavior engagement. In addition, accident experience, infrastructure suitability, perceived enjoyment, traffic risk perception, and operational risk perception are among the other significant predictors. Among socio-demographic factors, gender, age, education, and car use frequency significantly influence riders’ engagement in risky behaviors. The results highlight the importance of infrastructure suitability and accident experience in analyzing e-scooter users’ riding behavior. The developed model advances our understanding of factors contributing to e-scooter riders’ risky behavior engagement. The findings offer valuable insights for policymakers and e-scooter vendors aiming to mitigate e-scooter users’ accident risk. Specifically, we recommend three safety countermeasures: (1) safety training programs to encourage a safer attitude, (2) practice-based initiatives to enhance riding confidence, and (3) infrastructure improvements, especially the expansion of bike lanes.

E-scooter↗

DOE OSTI2019

Phase Segmentation in Atom-Probe Tomography Using Deep Learning-Based Edge Detection

Abstract Atom-probe tomography (APT) facilitates nano- and atomic-scale characterization and analysis of microstructural features. Specifically, APT is well suited to study the interfacial properties of granular or heterophase systems. Traditionally, the identification of the interface between, for precipitate and matrix phases, in APT data has been obtained either by extracting iso-concentration surfaces based on a user-supplied concentration value or by manually perturbing the concentration value until the iso-concentration surface qualitatively matches the interface. These approaches are subjective, not scalable, and may lead to inconsistencies due to local composition inhomogeneities. We introduce a digital image segmentation approach based on deep neural networks that transfer learned knowledge from natural images to automatically segment the data obtained from APT into different phases. This approach not only provides an efficient way to segment the data and extract interfacial properties but does so without the need for expensive interface labeling for training the segmentation model. We consider here a system with a precipitate phase in a matrix and with three different interface modalities—layered, isolated, and interconnected—that are obtained for different relative geometries of the precipitate phase. We demonstrate the accuracy of our segmentation approach through qualitative visualization of the interfaces, as well as through quantitative comparisons with proximity histograms obtained by using more traditional approaches.

47 OTHER INSTRUMENTATION↗

DOE OSTI2025

Machine learning-powered data cleaning for LEGEND: a semi-supervised approach using affinity propagation and support vector machines

Neutrinoless double-beta decay ($0\nu\beta\beta$) is a rare nuclear process that, if observed, will provide insight into the nature of neutrinos and help explain the matter-antimatter asymmetry in the Universe. The large enriched germanium experiment for neutrinoless double-beta decay (LEGEND) will operate in two phases to search for $0\nu\beta\beta$. The first (second) stage will employ 200 (1000) kg of High-Purity Germanium (HPGe) enriched in 76 Ge to achieve a half-life sensitivity of 10 27 (10 28 ) years. In this study, we present a semi-supervised data-driven approach to remove non-physical events captured by HPGe detectors powered by a novel artificial intelligence model. We utilize affinity propagation to cluster waveform signals based on their shape and a support vector machine to classify them into different categories. We train, optimize, and test our model on data taken from a natural abundance HPGe detector installed in the Full Chain Test experimental stand at the University of North Carolina at Chapel Hill. We demonstrate that our model yields a maximum sacrifice of physics events of $0.024 ^{+0.004}_{-0.003} \%$ after data cleaning. Our model is being used to accelerate data cleaning development for LEGEND-200 and will serve to improve data cleaning procedures for LEGEND-1000.

artificial intelligence↗

DOE OSTI2025

Neural Network Analysis of Nuclear Magnetic Resonance and Infrared Spectra

Nuclear magnetic resonance (NMR) spectroscopy and infrared (IR) spectroscopy are powerful chemical characterization techniques with broad general usage. However, the manual evaluation of the resulting spectra is time-consuming and requires significant expertise, preventing insights from being used in real-time applications. With recent advances in computation and artificial intelligence (AI), new tools are available for automating spectral interpretation. In this work, machine learning (ML) algorithms using 1-dimensional convolutional neural networks (CNNs) were applied to identify common functional groups from spectral information. Raw spectra were collected virtually from the Human Metabolome Database (HMDB) and National Institute of Standards and Technology (NIST) Chemistry WebBook and processed into a suitable standard. Algorithm design was tailored to best fit the nature of the problem, with built-in flexibility to accommodate relevant parameters beyond the raw spectral input, specifically solvent identity and magnetic frequency for NMR. The predictive capability of the algorithm in identifying functional groups is displayed in several examples. This methodology has been compiled into a code repository and could easily be modified to adapt alternative data sources, including other spectrum types. To mitigate overfitting, a common problem in mathematical modeling where overfamiliarity with training data produces trends that are not representative of the general data, a novel metric was developed, referred to as Accufit. Accufit includes a parameter that penalizes substantial differences in the training accuracy and the accuracy of an independent validation set. Examples are presented showing the effectiveness of Accufit in maintaining the model’s predictive capability while controlling the overfitting when used as a custom metric for hyperparameter tuning.

Sturgill, James↗

DOE OSTI2023

Machine Learning for Daily Forecasts of Arctic Sea Ice Motion: An Attribution Assessment of Model Predictive Skill

Physics-based simulations of Arctic sea ice are highly complex, involving transport between different phases, length scales, and time scales. Resultantly, numerical simulations of sea ice dynamics have a high computational cost and model uncertainty. We employ data-driven machine learning (ML) to make predictions of sea ice motion. The ML models are built to predict present-day sea ice velocity given present-day wind velocity and previous-day sea ice concentration and velocity. Models are trained using reanalysis winds and satellite-derived sea ice properties. We compare the predictions of three different models: persistence (PS), linear regression (LR), and a convolutional neural network (CNN). We quantify the spatiotemporal variability of the correlation between observations and the statistical model predictions. Additionally, we analyze model performance in comparison to variability in properties related to ice motion (wind velocity, ice velocity, ice concentration, distance from coast, bathymetric depth) to understand the processes related to decreases in model performance. Results indicate that a CNN makes skillful predictions of daily sea ice velocity with a correlation up to 0.81 between predicted and observed sea ice velocity, while the LR and PS implementations exhibit correlations of 0.78 and 0.69, respectively. The correlation varies spatially and seasonally: lower values occur in shallow coastal regions and during times of minimum sea ice extent. LR parameter analysis indicates that wind velocity plays the largest role in predicting sea ice velocity on 1-day time scales, particularly in the central Arctic. Regions where wind velocity has the largest LR parameter are regions where the CNN has higher predictive skill than the LR.

54 ENVIRONMENTAL SCIENCES↗

DOE OSTI2022

A Machine Learning Framework to Predict Images of Edge-on Protoplanetary Disks

The physical structure and properties of protoplanetary disks are typically derived from spatially resolved disk images. Edge-on disks in particular provide an important view point on the vertical structure and degree of settling of disks. Such analyses rely on radiative transfer (RT) calculations that are generally computationally intensive due to the high optical depth of disks. Here we present a machine learning framework that has the potential to dramatically speed up the forward modeling process by approximating the results of RT calculations. This framework, trained on an initial set of RT calculations, utilizes an autoencoder neural network to enable the generation of synthetic scattered light images of edge-on disks directly from a set of physical parameters. We demonstrate that this framework generates synthetic images 2–3 orders of magnitude faster than using RT calculations. These machine learning-generated images appear to approximate the RT images well, in particular preserving their size and shape. We also find a strong correlation between the latent space representations of the generated disk images and several of their associated physical parameters. Finally, we discuss potential changes to the framework, such as methods to further improve the image quality, extending the framework to multiple wavelengths, and inverting the process to infer physical parameters from observed images. Overall, these new tools have the potential to enable a more efficient and uniform analysis of edge-on disk properties and the initial conditions of planet formation.

79 ASTRONOMY AND ASTROPHYSICS↗

DOE OSTI2020

The PAU Survey: Photometric redshifts using transfer learning from simulations

In this paper, we introduce the DEEPZ deep learning photometric redshift (photo-z) code. As a test case, we apply the code to the PAU survey (PAUS) data in the COSMOS field. DEEPZ reduces the σ68 scatter statistic by 50 percent at iAB = 22.5 compared to existing algorithms. This improvement is achieved through various methods, including transfer learning from simulations where the training set consists of simulations as well as observations, which reduces the need for training data. The redshift probability distribution is estimated with a mixture density network (MDN), which produces accurate redshift distributions. Our code includes an autoencoder to reduce noise and extract features from the galaxy SEDs. It also benefits from combining multiple networks, which lowers the photo-z scatter by 10 percent. Furthermore, training with randomly constructed coadded fluxes adds information about individual exposures, reducing the impact of photometric outliers. In addition to opening up the route for higher redshift precision with narrow bands, these machine learning techniques can also be valuable for broad-band surveys.

79 ASTRONOMY AND ASTROPHYSICS↗