SEARCH · Search NASA

Results for “Observational Networks”

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 199 records · Page 11

All sources NASA NTRS DOE OSTI

DOE OSTI2025

Time-Varying Output Delay Compensation-A Model-Free Approach and its Application on Cooperative On-Ramp Merging

This paper presents a model-free approach to compensate for time-varying output delay in networked control systems. The proposed architecture combines a model-free observer and the Smith predictor. The model-free observer estimates the current state while handling modeling errors and uncertainties of the system. The Smith predictor moves the effect of time delay outside the control closed-loop using the estimated delayed output and the actual output of the plant. The proposed method is applied to a cooperative on-ramp merging problem. First, an ultra-local model predictive control is implemented to provide a computationally efficient online speed planner agnostic to the vehicle dynamics. After that, a model-free observer is designed to estimate the current state. Finally, the proposed architecture is tested against a time-varying output delay with an upper bound of 200 milliseconds. The results demonstrate the effectiveness of the proposed method with improved tracking of intervehicle distance.

Waleed khan, Muhammad [The University of Texas at ↗

DOE OSTI2025

Experimental Characterization of Hydrogen Diffusion in Shale Rocks for Geologic Storage Applications

As global energy systems undergo a transition to cleaner alternatives, geologic hydrogen storage has emerged as a promising solution for large-scale energy storage. A critical factor in determining the feasibility of this approach is the effectiveness of caprock formations, such as shale, in preventing hydrogen migration. This study investigates the diffusion behavior of hydrogen through shale to assess its suitability as a caprock for geologic hydrogen storage. Using a novel double-seal core holder design and a through-diffusion apparatus, hydrogen diffusion was measured through shale rock from the Eagle Ford and Wolfcamp Formations under dry conditions. These measurements were complemented by microstructural and mineralogical analyses using low-pressure nitrogen adsorption and X-ray diffraction. The effective diffusion coefficient of hydrogen in these shale caprocks ranged from 2.51 × 10 –8 to 9.85 × 10 –8 m 2 /s. Notably, we observed that the diffusion behavior was more related to the pore network structure and could not be attributed to differences in the total pore volume between shale types alone. Here, to further understand the role of pore network complexity, a fractal pore model was developed to correlate tortuosity with the fractal dimension of the pore structure (a measure of pore network complexity). The proposed model closely matched tortuosity values obtained from diffusion experiments, outperforming existing theoretical tortuosity–porosity correlations. These findings provide key quantitative parameters needed to assess the feasibility of geologic hydrogen storage as well as insights that can be applied to hydrogen storage in a range of geologic formations.

08 HYDROGEN↗

DOE OSTI2024

Point spread function deconvolution using a convolutional autoencoder

A major issue in optical astronomical image analysis is the combined effect of the instrument’s point spread function (PSF) and the atmospheric seeing that blurs images and changes their shape in a way that is band and time-of-observation dependent. In this work we present a very simple neural network based approach to nonblind image deconvolution that relies on feeding a convolutional autoencoder (CAE) input images that have been preprocessed by convolution with the corresponding PSF and its regularized inverse, a method which is both conceptually simple and computationally less intensive. We also present here, a new approach for dealing with limited input dynamic range of neural networks compared to the dynamic range present in astronomical images.

79 ASTRONOMY AND ASTROPHYSICS↗

DOE OSTI2025

Precise Linker Length and Dynamic Bond Exchange Control Penetrant Diffusion in Dense Vitrimers

Polymer networks with dynamic covalent bonds have been investigated for their self-healing ability, recyclability, and potential as more sustainable materials. Recent results have indicated that in some cases, bond exchange can enhance the transport of penetrants in dense networks, pointing to their potential for separations of membranes. Here, imine dynamic bonds in ethylene oxide (EO) networks with precise linker lengths were synthesized to investigate the transport of N,N′-bis(2,5-di-tert-butylphenyl)-3,4,9,10-perylenedicarboximide (BTBP), a large, anisotropic dye molecule. Networks with mesh sizes smaller than, comparable to, and greater than the size of the penetrant axes were investigated to probe the effects of bond exchange and network confinement on transport. Mesh sizes, which ranged from 0.5 to 1.62 nm, were determined from shear rheology, glass transitions by calorimetry, and probe diffusion coefficients by fluorescence recovery after photobleaching. Permanent networks with identical EO chain lengths were prepared as control samples, and up to a 3 orders of magnitude increase in diffusion coefficient is observed in the dynamic systems for short linkers containing 13 backbone atoms. The longest linkers with 71 backbone atoms show no difference between the permanent and dynamic networks. Linkers shorter than 11 backbone atoms, corresponding to a mesh size smaller than the penetrant small axis, diffusion is no longer observable on the experimental time scale, indicating a sharp cutoff attributed to the precise linkers and narrow mesh size distribution. The dynamic imine exchange time scales were compared to the diffusive hopping times of penetrants and indicate that exchange can occur during a diffusive displacement. Furthermore, these findings provide insights into the factors affecting penetrant transport in dense polymers and inspire the development of next-generation selective polymer membranes.

Diffusion↗

DOE OSTI2025

Climate warming enhances biodiversity and stability of grassland soil phosphorus-cycling microbial communities

Abstract Climate warming poses significant challenges to global phosphorus sustainability, an essential component of Earth biogeochemistry cycling and water-food-energy nexus. Despite the crucial role of polyphosphate-accumulating organism as key functional microbial agents in phosphorus cycling, the impacts of global climate warming on polyphosphate accumulating organism communities remain largely enigmatic. This study investigates the effects of climate warming on the taxonomic, network, and functional profiles of soil bacterial polyphosphate-accumulating organisms, leveraging fluorescence-activated cell sorting and single-cell Raman spectroscopy. Climate warming enhances both taxonomic and functional biodiversity of polyphosphate-accumulating organisms via biotic interactions and environmental filtering, with observed functionality-biodiversity relationships supporting the functional redundancy theory. Furthermore, polyphosphate-accumulating organism network complexity and stability rise under warming with strengthened positive relationships, supporting stress gradient hypothesis and the belief that complexity begets stability. Finally, polyphosphate-accumulating organisms are significantly correlated to key ecosystem functioning in carbon and phosphorus cycling under warming. Our study suggests that preserving polyphosphate-accumulating organism communities is crucial for maintaining soil ecosystem functioning and sustainable phosphorus management in a warming world and opens avenues for predicting the responses of other functional microbial groups to climate change, beneficially or maliciously.

Environmental Sciences & Ecology↗

DOE OSTI2024

Errant Beam Prognostics with Machine Leaning at SNS Accelerator

Particle Accelerators are complex machine with many pieces of equipment running in synchronization to deliver required beam. However, faults in particle accelerators reduce the availability of the beam for experiments affecting the overall science output. To avoid these faults, we apply anomaly detection techniques to predict any unusual behavior and perform preemptive actions to improve the total availability. Many researchers have adopted semi-supervised Machine Learning (ML) methods such as auto-encoders and variational auto-encoders for such tasks. However, supervised ML techniques designed for similarity learning such as Siamese Neural Network (SNN) can outperform semi-supervised or unsupervised methods for anomaly prediction. One of the challenges associated with application of ML models to particle accelerators is the variability in observed data over time due to system configuration changes. We employ conditional models such as Conditional Siamese Neural Networks (CSNN), and Conditional-VAE (CVAE) to learn the variability in the data by using beam configuration parameters as conditional input. We apply these models for errant beam prediction at Spallation Neutron Source accelerator under different system configurations and compare their performance. We demonstrate that CSNN outperforms CVAE in our application. This talk will present the data source, collection, analysis, data-preparation, model development, hyper-parameter studies and the results.

Rajput, Kishansingh↗

DOE OSTI2025

Mitigating imaging systematics for DESI 2024 emission Line Galaxies and beyond

Emission Line Galaxies (ELGs) are one of the main tracers that the Dark Energy Spectroscopic Instrument (DESI) uses to probe the universe. However, they are afflicted by strong spurious correlations between target density and observing conditions known as imaging systematics. In this paper, we present the imaging systematics mitigation applied to the DESI Data Release 1 (DR1) large-scale structure catalogs used in the DESI 2024 cosmological analyses. We also explore extensions of the fiducial treatment. This includes a combined approach, through forward image simulations (Obiwan) in conjunction with neural network-based regression, to obtain an angular selection function that mitigates the imaging systematics observed in the DESI DR1 ELGs target density. We further derive a line of sight selection function from the forward model that removes the strong redshift dependence between imaging systematics and low redshift ELGs. Combining both angular and redshift-dependent systematics, we construct a three-dimensional selection function and assess the impact of all selection functions on clustering statistics. We quantify differences between these extended treatments and the fiducial treatment in terms of the measured 2-point statistics. We find that the results are generally consistent with the fiducial treatment and conclude that the differences are far less than the imaging systematics uncertainty included in DESI 2024 full-shape measurements. We extend our investigation to the ELGs at 0.6 < z < 0.8, i.e., beyond the redshift range (0.8 < z < 1.6) adopted for the DESI clustering catalog, and demonstrate that determining the full three-dimensional selection function is necessary in this redshift range. Our tests showed that all changes are consistent with statistical noise for BAO analyses indicating they are robust to even severe imaging systematics. Specific tests for the full-shape analysis will be presented in a companion paper.

79 ASTRONOMY AND ASTROPHYSICS↗

DOE OSTI2025

Neuronal Plasma Membranes as Supramolecular Assemblies for Biological Memory

Biological memory is the ability to develop, retain, and retrieve information over time. Currently, it is widely accepted that memories are stored in synapses (i.e., connections between brain cells throughout the brain) through a process known as synaptic plasticity, which leads to either long-term potentiation (LTP) or long-term depression (LTD). However, the strengthening (LTP) and weakening (LTD) of synapses involve post-translational modifications to neural networks requiring de novo gene expression, a lengthy and energetically expensive process. Recently, we observed that lipid bilayers in the absence of peptides/proteins are capable of LTP, not unlike what has been observed in mammals and birds. As such, this finding has prompted us to postulate that the lipid bilayer provides a good model for understanding the molecular basis of biological memory. Here, in this article, we discuss the status, challenges, and opportunities of neuronal plasma membranes as structures for biological memory and learning, therapeutic targets for various brain disorders, and platforms for neural network developments.

59 BASIC BIOLOGICAL SCIENCES↗

DOE OSTI2025

Isomerization kinetics of azobenzene crosslinked to a liquid crystal polymer network

Azobenzene is an important chromophore having numerous applications that are a direct implication of its ability to change shape upon irradiation with a suitable wavelength. Here, in this study, photoisomerization kinetics of a diacrylate azobenzene molecule (A3MA) cross-linked to a liquid crystal polymer network (LCN) is investigated. Isomerization experiments were conducted on planar LCN thin films with varying cross-linking densities for a fixed azobenzene concentration. The experiments were conducted at approximately 26 °C, which is significantly below each film’s glass transition temperature. The forward trans-cis and reverse cis-trans isomerization slows down in crosslinked azobenzene liquid crystal polymer networks (ALCNs) as compared to a solution and an uncrosslinked polymer mixture. The observed slowdown is due to the reduction in the free volume and increased steric hindrances with an increase in crosslinking. A single-term first-order kinetics rate equation is known to describe the isomerization kinetics in solutions or uncrosslinked mixtures while the same form fails to represent the reaction kinetics in moderate to highly cross-linked ALCNs. Our results for the first time indicate that a two-term first-order reaction rate is an accurate way to describe the forward trans-cis and reverse thermal cis-trans isomerization in crosslinked azobenzene liquid crystal polymer networks. In addition, the effect of film thickness and incident intensities on the reaction kinetics is investigated. Finally, the two-term isomerization kinetics is incorporated into a finite element photo-chemo-mechanical model to reveal the consequent slowdown in the actuation response of ALCN films. The current findings will be of particular interest to theoreticians and experimentalists alike, for it will help predict and understand accurately the responses of azobenzene-based coatings, and actuators and may have potential applications in underwater soft robotics.

37 INORGANIC, ORGANIC, PHYSICAL, AND ANALYTICAL CH↗

DOE OSTI2025

Evolved Gas Analysis–Mass Spectrometry Exposes Polymer Network Structures

Polymer network structures in epoxy thermosets play an important role in the final thermoset material properties. However, analytical characterization of these network structures is difficult due to their amorphous nature. In this work, the application of evolved gas analysis–mass spectrometry (EGA-MS) to characterize the polymer network structures of bisphenol A (BPA)-based thermosets is demonstrated. Analytical characterization of the polymer network structures is accomplished by monitoring the Product-Specific Kinetics (PSK) of BPA monomer formation during thermal degradation investigations. We relate observed differences in the activation energy (E a ) of BPA monomer formation to the local packing environment around the BPA monomer units within the polymer network. Variations in the local environment related to the polymer networks manifest qualitatively as broadening in the thermal profile of the BPA monomer evolution and quantitatively as changes in the activation energy (E a ). Three BPA thermoset formulations were investigated; two amine-cured thermoset with 4,4′-diaminodiphenylmethane (DDM) or poly(propylene glycol) bis(2-amino-propyl ether) (PPG400) and a homopolymerized thermoset via curing with Epikure 3253 catalyst (3253). Results revealed that the 3253 thermoset contained two distinct packing densities in the polymer network, while DDM and PPG400 thermosets had uniform distributions of packing densities. Results from the DDM thermoset revealed a gradually decreasing E a , while the apparent E a of PPG400 was consistent over the entire degradation. Furthermore, these differences in E a were concluded to stem from the flexibility of the corresponding polymer networks and the ability of the network components to rearrange and occupy formed voids. Due to the minimal sample required for analysis (100–200 μg), this EGA-MS technique has great potential for postproduction evaluation of composite parts to identify changes in the polymer networks from use and aging, which could signal compromised performance.

Degradation↗

DOE OSTI2024

Elucidating the role of Cr migration in Ni-Cr exposed to molten FLiNaK via multiscale characterization

Structural materials used in nuclear reactor environments are subjected to coupled extremes such as high temperature, irradiation, and corrosion which act in concert to degrade their functional performance. Connecting alloy microstructure such as grain boundaries, and accumulating point defects with corrosion attack and pore morphology is critical to understanding underlying mechanisms. We uncover the compositional variations and morphology at multiple length scales in corrosion-damaged Ni-Cr alloy after exposure to oxidants in molten fluoride salts. A complex network of dense corrosion pores is detected by surface-level SEM observations. The corrosion pores take on a 1-dimensional morphology and are enriched with Ni and depleted of Cr 1–5 µm from the pore surface. STEM-EDS and 4D-STEM strain maps acquired simultaneously highlight the local variations in composition and structure of a ≤ 200 nm Cr-rich layer identified from a cross-section taken at the bottom of an isolated corrosion pore between the Ni-Cr alloy matrix and the embedded salt. However, the absence of an observed interface between the Ni-Cr alloy matrix and the FCC-structured Cr-rich layer suggests that Cr plating from the salt did not transpire. These findings support a proposed Cr lattice diffusion mechanism rather than Cr-precipitation from the salt to accommodate temperature transient conditions during sample cooling. Through the development of a 1D phase field model, these results are rationalized by formation energies for the Ni- and Cr-oxidation into the molten salt. This study reveals the locally altered microstructure caused by high temperature corrosion in non-steady-state molten salt nuclear reactor environments.

36 MATERIALS SCIENCE↗

DOE OSTI2025

On the Source of Conductivity in Alkaline Zn Anodes: Zn Percolation and ZnO Bridging

Alkaline Zn anodes are fundamental to commercial Zn-MnO 2 batteries as well as emerging rechargeable designs. In these electrodes, Zn particles are both the active material and source of electronic conduction. However, there are known cases in which electronic connectivity between Zn particles or the current collecting pin is lost even though the battery continues to function. Here we use X-ray computed tomography (CT) of AA batteries to demonstrate several examples of Zn particle-to-particle connectivity breakdown, which is observed even in cases at relatively high discharge rate. This indicates maintenance of the electronic network through the less-conductive ZnO discharge product. We introduce a new equation for electronic conductivity maintained through bridges of ZnO formed between percolating Zn particles. This necessitates discarding the Bruggeman correlation and instead redefining effective electronic conductivity using percolation theory. We demonstrate that such a model for conduction enables prediction of an inverted reaction zone, which is an experimentally observed case in which significant Zn dissolution and ZnO formation occurs heavily near the current collecting pin. Current computational Zn-MnO 2 models never predict an inverted reaction zone, and thus the updated conductivity enables models to accurately explain a wider range of experimental conditions.

Guida, Dominick P. [Northeastern Univ., Boston, MA↗

DOE OSTI2025

Analysis and optimization of seismic monitoring networks with Bayesian optimal experimental design

SUMMARY Monitoring networks increasingly aim to assimilate data from a large number of diverse sensors covering many sensing modalities. Bayesian optimal experimental design (OED) seeks to identify data, sensor configurations or experiments which can optimally reduce uncertainty and hence increase the performance of a monitoring network. Information theory guides OED by formulating the choice of experiment or sensor placement as an optimization problem that maximizes the expected information gain (EIG) about quantities of interest given prior knowledge and models of expected observation data. Therefore, within the context of seismo-acoustic monitoring, we can use Bayesian OED to configure sensor networks by choosing sensor locations, types and fidelity in order to improve our ability to identify and locate seismic sources. In this work, we develop the framework necessary to use Bayesian OED to optimize a sensor network’s ability to locate seismic events from arrival time data of detected seismic phases at the regional-scale. This framework requires five elements: (i) A likelihood function that describes the distribution of detection and traveltime data from the sensor network, (ii) A prior distribution that describes a priori belief about seismic events, (iii) A Bayesian solver that uses a prior and likelihood to identify the posterior distribution of seismic events given the data, (iv) An algorithm to compute EIG about seismic events over a data set of hypothetical prior events, (v) An optimizer that finds a sensor network which maximizes EIG. Once we have developed this framework, we explore many relevant questions to monitoring such as: how to trade off sensor fidelity and earth model uncertainty; how sensor types, number and locations influence uncertainty; and how prior models and constraints influence sensor placement.

58 GEOSCIENCES↗

DOE OSTI2025

Introduction: Neuromorphic Materials

The explosive growth in data collection and the need to process it efficiently, as well as the desire to automate increasingly complex tasks in transportation, medical care, manufacturing, security and many other fields have motivated a growing interest in neuromorphic computing. Unlike the binary, transistorbased ON/OFF logic gates and separate logic and memory functionalities employed in digital computing, neuromorphic computing is inspired by animal brains that use interconnected synapses and neurons to perform processing, storage and transmission of information at the same location, while only consuming ~20 W or less of power. Motivated by the brain’s efficiency, adaptability, self-learning and resiliency qualities, neuromorphic computing can be broadly defined as an approach to processing and storing information using hardware and algorithms inspired by models of biological neural systems. Present research in neuromorphic computing encompasses approaches that vary significantly in their degree of neuro-inspiration, from systems that only incorporate features such as asynchronous, event-driven operation or use crossbar arrays of non-volatile memory (NVM) elements to accelerate deep neural networks (DNNs), to designs that embrace the extreme parallelism, sparsity, reconfigurability, adaptability, complexity and stochasticity observed in nervous systems. The term ‘neuromorphic’ computing is often credited to Carver Mead, who in the 1980s investigated Si-based analog electronics to replicate functions of the animal retina. Earlier important advances in this field include the work of Frank Rosenblatt, who proposed the concept of the perceptron, Bernard Widrow, who used this concept to build one of the first analog neural networks, the Adaline and many other researchers (see ref. 6 for an historical perspective on neuromorphic computing). With the recent increase in the use of artificial intelligence and large language models, and rising concerns over the associated energy costs, interest in neuromorphic hardware has expanded rapidly. According to some estimates, driven largely by the drastic growth in the training use of artificial intelligence (AI) models using the current computing architectures, the energy cost of computing is projected to reach the energy supply worldwide by 2045. Furthermore, while this is not a realistic outcome, it means that, if more efficient computing technologies are not developed -- soon -- the world will soon become one where demand for energy and market constraints limit the continued increase of societal access to AI and cloud services from data centers. Data centers used for training and use of these models consume hundreds of terawatt hours of electricity, already past 4% of the US electricity demand.

Circuits↗

DOE OSTI2024

Kilonova light-curve interpolation with neural networks

Kilonovae are the electromagnetic transients created by the radioactive decay of freshly synthesized elements in the environment surrounding a neutron star merger. To study the fundamental physics in these complex environments, kilonova modeling requires, in part, the use of radiative transfer simulations. The microphysics involved in these simulations results in high computational cost, prompting the use of emulators for parameter inference applications. Utilizing a training set of 22 248 high-fidelity simulations (composed of 412 unique ejecta parameter combinations evaluated at 54 viewing angles), we use a neural network to efficiently train on existing radiative transfer simulations and predict light curves for new parameters in a fast and computationally efficient manner. Our neural network can generate millions of new light curves in under a minute. We discuss our emulator's degree of off-sample reliability and parameter inference of the AT2017gfo observational data. Finally, we discuss tension introduced by multiband inference in the parameter inference results, particularly with regard to the neural network's recovery of viewing angle. Published by the American Physical Society 2024

79 ASTRONOMY AND ASTROPHYSICS↗

DOE OSTI2024

The influence of burn severity on dissolved organic carbon concentrations across a stream network differs based on seasonal wetness conditions

Large, high-severity wildfires in many regions across the globe have increased concerns about their impacts on carbon cycling in watersheds. Altered sources of carbon and changes in catchment hydrology after wildfire can lead to shifts in dissolved organic carbon (DOC) concentrations in streams, which can have negative impacts on aquatic ecosystem health and downstream drinking-water treatment. Despite its importance, post-fire DOC responses remain relatively unconstrained in the literature, and we lack critical knowledge of how burn severity, landscape elements, and climate interact to affect DOC concentrations. To improve our understanding of the impact of burn severity on DOC concentrations, we measured DOC at 129 sites across a stream network extending upstream, within, and downstream of a large, high-severity wildfire in Oregon, USA. We collected samples across the study sub-basin during four distinct seasonal wetness conditions. We used our high-spatial-resolution data to develop spatial stream network (SSN) models to predict DOC across the stream network and to improve our understanding of the controls on DOC concentrations. Spatially, we found no obvious wildfire signal – instead, we observed a pattern of increasing DOC concentrations from the high-elevation headwaters to the sub-basin outlet, while the mainstem maintained consistently low DOC concentrations. This suggests that effects from large wildfires may be “averaged” out at higher stream orders and larger spatial scales. When we grouped DOC concentrations by burn severity group, we observed a significant decrease in the variability of DOC concentrations in the moderate and high burn severity sub-catchments. However, our SSN models were able to predict decreases in DOC concentrations with increases in burn severity across the stream network. Decreases in DOC concentrations were also highly variable across seasonal wetness conditions, with the greatest (–1.40 to –1.64 mg L –1 ) decrease occurring in the high-severity group during the wetting season. Additionally, our models indicated that in all seasons, baseflow index was more influential in predicting DOC concentrations than burn severity was, indicating that groundwater discharge can obscure the impacts of wildfire in a stream network. Overall, our results suggested that landscape characteristics can regulate the DOC response to wildfire. Moreover, our results also indicated that the seasonal timing of sampling can influence the observed response of DOC concentrations to wildfire.

54 ENVIRONMENTAL SCIENCES↗

DOE OSTI2024

HPC Network Simulation Tuning via Automatic Extraction of Hardware Parameters

Popular HPC network interconnection simulators such as SST/macro provide a variety of configurable parameters to explore the design space of hardware components such as network interface cards (NIC), switches, and links among them. While such knobs provide flexibility to explore design trade-offs for novel hardware, manually configuring simulations for matching configurations of the existing hardware to focus on topology exploration can be cumbersome and error-prone, leading to widely inaccurate simulations. This challenge is compounded when specifications of various (proprietary) technologies are not readily available or intentionally omitted. In this work, we propose a framework to autotune the multiple network models’ simulation configurations within SST/macro using Tree-structured Parzen Estimator-based Bayesian optimization to observe the effect on simulation accuracy across different message regimes. These regimes consist of small to large message sizes and latency to bandwidth-bound messages. We provide a detailed analysis of the simulation error for four representative HPC systems. Our Bayesian optimization based autotuning framework for network models achieves a maximum of 5x improvement in accuracy over best-effort manual configurations based on available hardware specifications.

Simulation, autotuning↗

DOE OSTI2024

Using the Metropolis algorithm to explore the loss surface of a recurrent neural network

In the limit of small trial moves the Metropolis Monte Carlo algorithm is equivalent to gradient descent on the energy function in the presence of Gaussian white noise. This observation was originally used to demonstrate a correspondence between Metropolis Monte Carlo moves of model molecules and overdamped Langevin dynamics, but it also applies in the context of training a neural network: making small random changes to the weights of a neural network, accepted with the Metropolis probability, with the loss function playing the role of energy, has the same effect as training by explicit gradient descent in the presence of Gaussian white noise. We explore this correspondence in the context of a simple recurrent neural network. We also explore regimes in which this correspondence breaks down, where the gradient of the loss function becomes very large or small. In these regimes the Metropolis algorithm can still effect training, and so can be used as a probe of the loss function of a neural network in regimes in which gradient descent struggles. We also show that training can be accelerated by making purposely-designed Monte Carlo trial moves of neural-network weights.

Casert, Corneel↗