Particle reconstruction of volumetric particle image velocimetry with the strategy of machine learning

Three-dimensional particle reconstruction with limited two-dimensional projections is an under-determined inverse problem that the exact solution is often difficult to be obtained. In general, approximate solutions can be obtained by iterative optimization methods. In the current work, a practical particle reconstruction method based on a convolutional neural network (CNN) with geometry-informed features is proposed. The proposed technique can refine the particle reconstruction from a very coarse initial guess of particle distribution that is generated by any traditional algebraic reconstruction technique (ART) based methods. Compared with available ART-based algorithms, the novel technique makes significant improvements in terms of reconstruction quality, robustness to noise, and at least an order of magnitude faster in the offline stage.

Advanced iterative particle reconstruction for Lagrangian particle tracking

The method of iterative particle reconstruction (IPR), introduced by Wieneke (Meas Sci Technol 24:024008, 2013), constitutes a major step toward Lagrangian particle tracking in densely seeded flows (Schanz et al. in Exp Fluids 57:1–27, 2016). Here we present novel approaches in several key aspects of the algorithm, which, in combination, triple the working range of IPR in terms of particle image densities. The updated method is proven to be fast, accurate and robust against image noise and other imaging artifacts. Most of the proposed changes to the original processing are easy to implement and come at low computational cost. Furthermore, a bundle adjustment scheme that simultaneously updates the 3D locations of all particles and the camera calibrations is introduced. While the particle position optimization proved to be more effective using localized ‘shake’ schemes, this so-called global shake scheme constitutes an effective measure to correct for decalibrations and vibrations, acting as an in-situ single-image volume-self-calibration. Further optimization strategies using such approaches are conceivable. Graphic abstract

Challenges in solving structures from radiation-damaged tomograms of protein nanocrystals assessed by simulation

Structure-determination methods are needed to resolve the atomic details that underlie protein function. X-ray crystallography has provided most of our knowledge of protein structure, but is constrained by the need for large, well ordered crystals and the loss of phase information. The rapidly developing methods of serial femtosecond crystallography, micro-electron diffraction and single-particle reconstruction circumvent the first of these limitations by enabling data collection from nanocrystals or purified proteins. However, the first two methods also suffer from the phase problem, while many proteins fall below the molecular-weight threshold required for single-particle reconstruction. Cryo-electron tomography of protein nanocrystals has the potential to overcome these obstacles of mainstream structure-determination methods. Here, a data-processing scheme is presented that combines routines from X-ray crystallography and new algorithms that have been developed to solve structures from tomograms of nanocrystals. This pipeline handles image-processing challenges specific to tomographic sampling of periodic specimens and is validated using simulated crystals. The tolerance of this workflow to the effects of radiation damage is also assessed. The simulations indicate a trade-off between a wider tilt range to facilitate merging data from multiple tomograms and a smaller tilt increment to improve phase accuracy. Since phase errors, but not merging errors, can be overcome with additional data sets, these results recommend distributing the dose over a wide angular range rather than using a finer sampling interval to solve the protein structure.

Distance-Weighted Graph Neural Networks on FPGAs for Real-Time Particle Reconstruction in High Energy Physics

Graph neural networks have been shown to achieve excellent performance for several crucial tasks in particle physics, such as charged particle tracking, jet tagging, and clustering. An important domain for the application of these networks is the FGPA-based first layer of real-time data filtering at the CERN Large Hadron Collider, which has strict latency and resource constraints. We discuss how to design distance-weighted graph networks that can be executed with a latency of less than one μs on an FPGA. To do so, we consider a representative task associated to particle reconstruction and identification in a next-generation calorimeter operating at a particle collider. We use a graph network architecture developed for such purposes, and apply additional simplifications to match the computing constraints of Level-1 trigger systems, including weight quantization. Using the hls4ml library, we convert the compressed models into firmware to be implemented on an FPGA. Performance of the synthesized models is presented both in terms of inference accuracy and resource usage.

Sparseness and Smoothness Regularized Imaging for improving the resolution of Cryo-EM single-particle reconstruction

In this paper, we present a refinement method for cryo-electron microscopy (cryo-EM) single-particle reconstruction, termed as OPUS-SSRI (Sparseness and Smoothness Regularized Imaging). In OPUS-SSRI, spatially varying sparseness and smoothness priors are incorporated to improve the regularity of electron density map, and a type of real space penalty function is designed. Moreover, we define the back-projection step as a local kernel regression and propose a first-order method to solve the resulting optimization problem. On the seven cryo-EM datasets that we tested, the average improvement in resolution by OPUS-SSRI over that from RELION 3.0, the commonly used image-processing software for single-particle cryo-EM, was 0.64 Å, with the largest improvement being 1.25 Å. We expect OPUS-SSRI to be an invaluable tool to the broad field of cryo-EM single-particle analysis. The implementation of OPUS-SSRI can be found at https://github.com/alncat/cryoem.

Multi-particle reconstruction in the High Granularity Calorimeter using object condensation and graph neural networks

The high-luminosity upgrade of the LHC will come with unprecedented physics and computing challenges. One of these challenges is the accurate reconstruction of particles in events with up to 200 simultaneous protonproton interactions. The planned CMS High Granularity Calorimeter offers fine spatial resolution for this purpose, with more than 6 million channels, but also poses unique challenges to reconstruction algorithms aiming to reconstruct individual particle showers. In this contribution, we propose an end-to-end machine-learning method that performs clustering, classification, and energy and position regression in one step while staying within memory and computational constraints. We employ GravNet, a graph neural network, and an object condensation loss function to achieve this task. Additionally, we propose a method to relate truth showers to reconstructed showers by maximising the energy weighted intersection over union using maximal weight matching. Our results show the efficiency of our method and highlight a promising research direction to be investigated further.