Machine Learning for Charged Particle Imaging Applications

自1987年首次演示以来，用于分子动力学研究的带电粒子成像已经显著地改变了气相光化学和分子散射实验。自该技术出现以来，已经对其进行了许多改进，其中最受欢迎的是速度图成像（VMI）。这些实验使用激光脉冲探测化学过程（即光解离）并电离光碎片。这创造了一个后坐力分布，然后通过使用静电透镜组件投射到位置敏感探测器上。生成的图像随后被相机捕获，提供了一个非常有用的多路能量和角度分辨数据来源，可以对气相分子的光化学动力学产生很大的了解。大量的后处理技术可以在这些图像上实现许多应用，并且许多这些转换（例如去除噪声或超分辨率成像）可以使用机器学习技术显着增强：在传统计算中，输入数据通过预先确定的算法进行转换以产生特定的输出，与之相反，机器学习通过使用许多对已知的输入/输出组合来切换这一过程，而不是开发实际的转换算法。一旦这个“训练”阶段完成，就可以根据需要处理额外的输入图像。这使我们能够解决原本非常具有挑战性的问题-特别是在分析数学解决方案不可能的情况下。先前利用机器学习的VMI应用包括从实验数据中去除噪声[chemphychem， 22, 76,(2021)]，以及从单个2D投影中“重新膨胀”离子/电子原始3D球形分布的能力[Rev. Sci]。Instrum。[j].农业工程学报，2016,33(2):481 - 481。拟议的博士项目将建立在这项工作的基础上，并开发更多的神经网络来解决与带电粒子图像处理相关的更广泛的问题。这包括，例如，下一代成像探测器的图像上采样，真空紫外光谱区偏振诊断的新策略，以及时域数据稀疏图像的去噪。这将涉及广泛的计算/数值建模工作，以准确模拟相关VMI训练数据，以及实现各种神经网络架构。此外，该项目将包括一些实验性VMI测试数据的获取，以帮助验证正在开发的网络。

Charged particle imaging for molecular dynamics studies has significantly revolutionised experiments in gas-phase photochemistry and molecular scattering since its initial demonstration in 1987. Many improvements to this technique have been developed since its emergence, the most popular of which is known as velocity map imaging (VMI). These experiments use laser pulses to probe chemical processes (i.e. photodissociation) and ionise the photofragments. This creates a recoiling distribution which is then projected onto a position sensitive detector by use of an electrostatic lens assembly. The resulting image is then captured by a camera, providing an incredibly useful source of multiplexed energy- and angle-resolved data that can yield a great deal of insight into the photochemical dynamics of gas-phase molecules. A plethora of post-processing techniques can be implemented on these images for numerous applications, and many of these transformations (such as the removal of noise, or super-resolution imaging) can be significantly enhanced using machine learning techniques: In contrast to traditional computing, where input data is transformed via a predetermined algorithm to produce a specific output, machine learning switches up this process by using many pairs of known input/output combinations to instead develop the actual transformation algorithm. Once this "training" phase is complete, additional input images may be processed as required. This allows us to tackle problems that would otherwise be very challenging - particularly where analytical mathematical solutions are not possible. Previous VMI applications exploiting machine learning include the removal of noise from experimental data [ChemPhysChem, 22, 76, (2021)] and the ability to "reinflate" the original 3D spherical distribution of ions/electrons from just a single 2D projection [Rev. Sci. Instrum., 93, 023303, (2022)]. The proposed PhD project will build upon this work and develop additional neural networks to tackle a much wider range of problems relevant to charged particle image processing. This includes, for example, image up-sampling for next generation imaging detectors, novel strategies for polarization diagnostics in the vacuum ultraviolet spectral region, and denoising data sparse images in the temporal domain. This will involve extensive computational/numerical modelling work for the accurate simulation of relevant VMI training data, as well as implementing a variety of neural network architectures. Additionally, the project will include some acquisition of experimental VMI test data to help validate the networks under development.