A Machine Learning-Driven Solution for Denoising Inertial Confinement Fusion Images
📝 Abstract
Neutron imaging is essential for diagnosing and optimizing inertial confinement fusion implosions at the National Ignition Facility. Due to the required 10-micrometer resolution, however, neutron image require image reconstruction using iterative algorithms. For low-yield sources, the images may be degraded by various types of noise. Gaussian and Poisson noise often coexist within one image, obscuring fine details and blurring the edges where the source information is encoded. Traditional denoising techniques, such as filtering and thresholding, can inadvertently alter critical features or reshape the noise statistics, potentially impacting the ultimate fidelity of the iterative image reconstruction pipeline. However, recent advances in synthetic data production and machine learning have opened new opportunities to address these challenges. In this study, we present an unsupervised autoencoder with a Cohen-Daubechies- Feauveau (CDF 97) wavelet transform in the latent space, designed to suppress for mixed Gaussian-Poisson noise while preserving essential image features. The network successfully denoises neutron imaging data. Benchmarking against both simulated and experimental NIF datasets demonstrates that our approach achieves lower reconstruction error and superior edge preservation compared to conventional filtering methods such as Block-matching and 3D filtering (BM3D). By validating the effectiveness of unsupervised learning for denoising neutron images, this study establishes a critical first step towards fully AI-driven, end-to-end reconstruction frameworks for ICF diagnostics.
💡 Analysis
Neutron imaging is essential for diagnosing and optimizing inertial confinement fusion implosions at the National Ignition Facility. Due to the required 10-micrometer resolution, however, neutron image require image reconstruction using iterative algorithms. For low-yield sources, the images may be degraded by various types of noise. Gaussian and Poisson noise often coexist within one image, obscuring fine details and blurring the edges where the source information is encoded. Traditional denoising techniques, such as filtering and thresholding, can inadvertently alter critical features or reshape the noise statistics, potentially impacting the ultimate fidelity of the iterative image reconstruction pipeline. However, recent advances in synthetic data production and machine learning have opened new opportunities to address these challenges. In this study, we present an unsupervised autoencoder with a Cohen-Daubechies- Feauveau (CDF 97) wavelet transform in the latent space, designed to suppress for mixed Gaussian-Poisson noise while preserving essential image features. The network successfully denoises neutron imaging data. Benchmarking against both simulated and experimental NIF datasets demonstrates that our approach achieves lower reconstruction error and superior edge preservation compared to conventional filtering methods such as Block-matching and 3D filtering (BM3D). By validating the effectiveness of unsupervised learning for denoising neutron images, this study establishes a critical first step towards fully AI-driven, end-to-end reconstruction frameworks for ICF diagnostics.
📄 Content
NERTIAL confinement fusion (ICF) is the process of compressing fusion fuel at high temperature and pressure and allowing inertia to confine the fuel for long enough to induce fusion of atomic nuclei [1,2]. Most commonly, ICF uses deuterium-deuterium (DD) or deuterium-tritium (DT) fusion. It can be designed with direct or indirect drive [3,4]. Direct drive ICF relies on multiple lasers directly heating, ablating and compressing a small capsule containing fusion fuel. In indirectdrive ICF, lasers heat a hohlraum around the capsule, and radiation from the hohlraum then heats, ablates and compresses the capsule [1,5]. Laser-driven indirect-drive ICF was the method used to achieve scientific breakeven at the Lawrence Livermore National Laboratory’s National Ignition Facility (NIF) [1] A. Neutron Aperture Imaging
Neutron pinhole and penumbral aperture imaging are employed at NIF for measuring the shape of the burning fuel in ICF events [5]. The pinhole apertures are smaller in size than the source while penumbral apertures are larger than the source. To create a larger effective field of view, each NIF neutron imaging system uses an array of apertures that is a mix of pinhole and penumbral apertures.
Due to neutrons’ ability to penetrate dense materials, neutron imaging at NIF requires thick apertures. Thus, the images captured are integrations over the source and the translationvariant point-spread function (PSF) of the source. For penumbral apertures, the source information is encoded within the penumbral shadow of the aperture, and image processing and reconstruction techniques are required to determine the source [5]. In contrast, pinhole apertures provide a clearer, more direct image of the source, but for small sources, the images may still require image reconstruction to remove the aperture PSF and make a better estimate of the source.
The image reconstructions at NIF currently use iterative generalized expectation maximization algorithms, which require significant time and frequent human interventions that will be incompatible with any future higher repetition rate ICF facility. An alternative approach for the reconstructions would be to use a fully artificial-intelligence-driven framework.
In this work, we take a first step towards such a framework by investigating using an autoencoder for denoising neutron images from NIF. Such methods would also be directly applicable to denoising many x-ray or gamma-ray images.
The NIF images used for this denoising analysis study were captured from image plates that collected neutrons passing through thick pinhole and penumbral apertures, which produces challenges to visual data quality [6]. The low solid angle of pinhole apertures limits their ability to image certain smaller ICF cases appropriately due to decreased signal strength and neutron concentration per pinhole making them more I > REPLACE THIS LINE WITH YOUR MANUSCRIPT ID NUMBER (DOUBLE-CLICK HERE TO EDIT) < susceptible to noise than penumbral or annular apertures [5]. Even on high-yield neutron shots, there are numerous sources of noise [4].
Poisson shot noise is produced due to discrete photons or particles colliding with the detector and is most significant in low-yield conditions. Numerous effects can also produce Gaussian noise in the data. Uniform noise, much like Poisson noise, is seen in situations where signals are poor. Salt-andpepper noise arises during transmission errors or because of sensor defects in camera-based systems, which are also used at NIF. Finally, defects or scratches present on the plate or detector can directly affect the image, causing visible artifacts in those areas.
While there are many noise combinations possible, for the lower-yield shots used in this study, image analysis led to the conclusion that Poisson noise, Gaussian noise, and plate defects are the most common sources of noise to be addressed. After different particle conversions, gain stages, and background subtractions, the noise is closer to Gaussian noise than Poisson or combined Gaussian-Poisson noise in certain NIF images [5]. However, for low-yield shots, Poisson shot noise often combines with Gaussian and is the predominant source of noise in some images [5]. This necessitates extensive work in denoising pinhole images while still retaining the basic properties of the original image, with a specific focus on tackling edge preservation and source reconstruction for NIF data outputs.
Currently, ICF images are usually denoised via filtering methods such as Gaussian filtering, Wiener filtering, and BM3D [7,8]. In recent years, there has been an uptick in the usage of Convolutional Neural Networks (CNNs) to denoise real and simulated neutron pinhole images [9]. While unsupervised methods have been explored more frequently in the recent past due to their small-training-set robustness, they have not been popular for denoising ICF images due to the complex noise environment [10].
The current lack of ground t
This content is AI-processed based on ArXiv data.