Search

We demonstrate live-updating ptychographic reconstruction with ePIE, an iterative ptychography method, during ongoing data acquisition. The reconstruction starts with a small subset of the total data, and as the acquisition proceeds the data used for reconstruction is extended. This creates a live-updating view of object and illumination that allows monitoring the ongoing experiment and adjusting parameters with quick turn-around. This is particularly advantageous for long-running acquisitions. We show that such a gradual reconstruction yields interpretable results already with a small subset of the data. We show simulated live processing with various scan patterns, parallelized reconstruction, and real-world live processing at the hard X-ray ptychographic nanoanalytical microscope PtyNAMi at the PETRA III beamline.

Phase retrieval is an ill-posed inverse problem with several applications in the fields of medical imaging and materials science. Conventional phase retrieval algorithms either simplify the problem by assuming certain object properties and optical propagation regimes or tuning a large number of free parameters. While the latter most often leads to good solutions for a wider application range, it is still a time-consuming process, even for experienced users. One way to circumvent this is by introducing a self-optimizing machine learning- based algorithm. Basing this on invertible networks such as normalising flows ensures good inversion, effi- cient sampling, and fast probability density estimation for large images and generally, complex-valued dis- tributions. Here, complex wavefield datasets are trained and tested on a normalising flows-based machine learning model for phase retrieval called conditional Wavelet Flow (cWF) and benchmarked against other conventional algorithms and baseline models. The cWF algorithm adds a conditioning network on top of the Wavelet Flow algorithm that is able to model the conditional data distribution of high resolution images of up to 1024 x 1024 pixels, which was not possible in other flow-based models. Additionally, cWF takes advantage of the parallelized training of different image resolutions, allowing for more efficient and fast training of large datasets. The trained algorithm is then applied to X-ray holography data wherein fast and high-quality image reconstruction is made possible.

Understanding laser-solid interactions is important for the development of laser-driven particle and photon sources, e.g., tumor therapy, astrophysics, and fusion. Currently, these interactions can only be modeled by simulations that need to be verified experimentally. Consequently, pump-probe experiments were conducted to examine the laser-plasma interaction that occurs when a high intensity laser hits a solid target. Since we aim for a femtosecond temporal and nanometer spatial resolution at European XFEL, we employ Small-Angle X-ray Scattering (SAXS) and Phase Contrast Imaging (PCI) that can each be approximated by an analytical propagator. In our reconstruction of the target, we employ a gradient descent algorithm that iteratively minimizes the error between experimental and synthetic patterns propagated from proposed target structures. By implementing the propagator in PyTorch we leverage the automatic differentiation capabilities, as well as the speed-up by computing the process on a GPU. We perform a scan of different initial parameters to find the global minimum, which is accelerated by batching multiple parallel reconstructions.

In the exploration of universe and matter, dealing with inverse problems is often a central challenge. In many experimental investigations, which are carried out in particular at large-scale research facilities such as FRM II, DESY or European XFEL, the essential phase information in the experimental data is lost due to the measurement principle (phase problem). Therefore, methods based on direct inversion are not applicable, so that the solution of the underlying non-convex optimization problem is usually very time-consuming and expensive to implement. Here we present our project “Versatile Inverse Problem fRamework” (VIPR), recently funded by the Federal Ministry of Education and Research (Grant 05D23CJ1). The stated goal of our project is to develop a flexible software framework for data-driven solutions to inverse problems. First we plan to focus on using invertible neural networks. Other architectures can be also considered at later stages of the project. The main application areas envisioned include grazing incidence small- and wide-angle scattering with both neutrons and x-rays, neutron/x-ray reflectivity, and ptychography. Development will also take into account requirements from spectroscopy and particle physics.

Grazing-Incidence Small-Angle X-ray Scattering (GISAXS) is a modern imaging technique used in material research to study nanoscale materials. Reconstruction of the parameters of an imaged object imposes an ill-posed inverse problem that is further complicated when only an in-plane GISAXS signal is available. Traditionally used inference algorithms such as Approximate Bayesian Computation (ABC) rely on computationally expensive scattering simulation software, rendering analysis highly time-consuming. We propose a simulation-based framework that combines variational auto-encoders and normalizing flows to estimate the posterior distribution of object parameters given its GISAXS data. We apply the inference pipeline to experimental data and demonstrate that our method reduces the inference cost by orders of magnitude while producing consistent results with ABC.

In our HELIPORT workshop, we provided insights into our project and share our results. In addition, we would like to provide a platform for the presentation of similar projects, as well as extensions or integrations from the surrounding research areas. The overall goal of the workshop is bringing together different institutions with similar challenges and establishing a community around our HELIPORT project. We therefore encouraged our community to submit an abstract for a talk or poster. The submissions are dedicated to four thematic points: • HELIPORTuse-cases, • Scientificprojectandmetadatamanagement, • Experiment-specificandoverallmetadataand • Scientificworkflows. HELIPORT (Helmholtz ScIentific Project WORkflow PlaTform) itself is a project funded by the Helmholtz Metadata Collaboration, and runs from July 2021 until June 2023. HELIPORT aims to make the entire life cycle of a scientific project findable, accessible, interoperable and reusable according to the FAIR principles. In particular, our data management solution deals with the areas from the generation of the data to the publication of primary research data, the workflows carried out and the actual research results. For this purpose, a concept was developed which shows the various essential components and their connections.

Steerable convolutional neural networks (CNNs) provide a general framework for building neural networks equivariant to translations and transformations of an origin-preserving group G, such as reflections and rotations. They rely on standard convolutions with G-steerable kernels obtained by analytically solving the group-specific equivariance constraint imposed onto the kernel space. As the solution is tailored to a particular group G, implementing a kernel basis does not generalize to other symmetry transformations, complicating the development of general group equivariant models. We propose using implicit neural representation via multi-layer perceptrons (MLPs) to parameterize G-steerable kernels. The resulting framework offers a simple and flexible way to implement Steerable CNNs and generalizes to any group G for which a G-equivariant MLP can be built. We prove the effectiveness of our method on multiple tasks, including N-body simulations, point cloud classification and molecular property prediction.

Recently, free electron lasing at UV wavelength has been demonstrated by deploying the COXINEL beamline driven by HZDR plasma accelerator in a seeded configuration[1]. Further control and optimization of such an FEL radiation require full knowledge of strongly-coupled multivariate parameters involved in laser plasma acceleration, electron beam transport and radiation generation. For this purpose, one has to solve an inverse problem, i.e. finding matching parameters of the simulation to reproduce the experiment. Such inverse problems are ill-posed and cannot be easily resolved due to high computational complexity. Here, machine learning-based methods have a high potential to accelerate theoretical comprehension of the system, novel means for design space exploration and promise reliable in-situ analysis of experimental diagnostics and parameters. We apply simulation-based inference technique for this purpose. This method is a combination of deep learning and statistical approaches to resolve an inverse problem up to a posterior distribution of the simulation parameters given an experimental sample. In addition, we have developed machine learning-based surrogate models that can significantly accelerate forward computations for even faster results of the inverse solver. [1] M. Labat, et al. "Seeded free-electron laser in driven by a compact laser plasma accelerator", Nat. Photonics, 17, 150(2023)

The output of simulations can be extremely challenging to work with. For example, in large-scale particle-in-cell simulations in plasma physics, trillions of particles are simulated over millions of time steps, creating Petabytes of data. In this project, we develop methods to compress particle data by training conditional invertible neural networks (cINNs) on the particle data. The particles then can be reconstructed by running the trained model in generative mode. This allows us to reach up to millionfold compression, and a controlled loss of a accuracy. The models can be conditioned not only on the temporal axis but also on other types of simulation outputs of smaller data volume, leading potentially to even higher compression factors. In order to enable the neural network model to represent simulation data over long time spans, we apply methods from Continual Learning, where each new learning task is represented by a dataset produced by a simulation time step. This approach enables us to efficiently solve the inverse problem of particle data reconstruction from radiation data in a time-resolved manner by side-stepping demanding simulations.

Scanning coherent X-ray microscopy (ptychography) has gained considerable interest during the last decade since the performance of this indirect imaging technique does not necessarily rely on the quality of the X-ray optics and, in principle, can achieve highest spatial resolution in X-ray imaging. The method can be easily extended to 3D by adding standard tomographic reconstruction schemes. However, the tomographic reconstruction is often applied in a subsequent step using a sequence of aligned ptychographic 2D projections. In this contribution, we outline current developments of a GPU-accelerated framework for direct 3D ptychography, coupling 2D ptychography and tomography. The program utilizes a custom GPU-accelerated framework for ptychography that offers three distinct ptychographic reconstruction algorithms. The tomographic reconstruction runs simultaneously and uses numerical routines of the ASTRA Toolbox. This parallel-computing approach results in a high performance increase considerably reducing the reconstruction time of 3D ptychographic datasets.

While the interaction of ultra-intense ultra-short laser pulseswith near- and overcritical plasmas cannot be directly observed, experimentally accessible quantities (observables) often only indirectly giveinformation about the underlying plasma dynamics. Furthermore, theinformation provided by observables is incomplete, making the inverseproblem highly ambiguous. Therefore, in order to infer plasma dynamicsas well as experimental parameter, the full distribution over parameters given an observation needs to considered, requiring that models areflexible and account for the information lost in the forward process. Invertible Neural Networks (INNs) have been designed to efficiently modelboth the forward and inverse process, providing the full conditional posterior given a specific measurement. In this work, we benchmark INNsand standard statistical methods on synthetic electron spectra. First, weprovide experimental results with respect to the acceptance rate, whereour results show increases in acceptance rates up to a factor of 10. Additionally, we show that this increased acceptance rate also results in anincreased speed-up for INNs to the same extent. Lastly, we propose acomposite algorithm that utilizes INNs and promises low runtimes whilepreserving high accuracy.

Modelling the temperature of Electric Vehicle (EV) batteries is a fundamental task of EV manufacturing. Extreme temperatures in the battery packs can affect their longevity and power output. Although theoretical models exist for describing heat transfer in battery packs, they are computationally expensive to simulate. Furthermore, it is difficult to acquire data measurements from within the battery cell. In this work, we propose a data-driven surrogate model (LiFe-net) that uses readily accessible driving diagnostics for battery temperature estimation to overcome these limitations. This model incorporates Neural Operators with a traditional numerical integration scheme to estimate the temperature evolution. Moreover, we propose two further variations of the baseline model: LiFe-net trained with a regulariser and LiFe-net trained with time stability loss. We compared these models in terms of generalization error on test data. The results showed that LiFe-net trained with time stability loss outperforms the other two models and can estimate the temperature evolution on unseen data with a relative error of 2.77 % on average.

High-fidelity (computer-) experiments are very expensive to perform, hence extracting as much information as possible from the collected data is vital. We present methods to estimate experiment outcomes based on ingested past data, to steer further sample acquisition both by suggesting paths for maximized uncertainty reduction and highlighting sensitivities on inputs, with use-cases from laser-light propagation and laser-plasma interactions to electron-bunch kinetics.

Innovation in mass spectrometry (MS) and the rapidly increasing throughput and sensitivity of MS instrumentation require adaptations and innovations in data processing tools. Here, we introduce MZmine 3, a scalable MS data analysis platform that supports hybrid datasets from various instrumental setups, including liquid and gas chromatography (LC and GC)–MS, ion mobility spectrometry (IMS)–MS and MS imaging. In particular, the integration of IMS–MS imaging and LC–IMS–MS datasets provides opportunities for spatial metabolomics analyses with increased annotation confidence.

Understanding laser-solid interactions is important for the development of laser-driven particle and photon sources, e.g., tumor therapy, astrophysics, and fusion. Currently, these interactions can only be modeled by simulations that need to be verified experimentally. Consequently, pump-probe experiments were conducted to examine the laser-plasma interaction that occurs when a high intensity laser hits a solid target. Since we aim for a femtosecond temporal and nanometer spatial resolution at European XFEL, we employ Small-Angle X-ray Scattering (SAXS) and Phase Contrast Imaging (PCI) that can each be approximated by an analytical propagator. In our reconstruction of the target, we employ a gradient descent algorithm that iteratively minimizes the error between experimental and synthetic patterns propagated from proposed target structures. By implementing the propagator in PyTorch we leverage the automatic differentiation capabilities, as well as the speed-up by computing the process on a GPU. We perform a scan of different initial parameters to find the global minimum, which is accelerated by batching multiple parallel reconstructions. The fully differentiable implementation of a forward function may serve as a physically-constraining loss, enabling training with unpaired data or unsupervised training of neural networks to predict the initial parameters for the gradient descent fit.

Building high-fidelity digitial twins through start-to-end models to better understand and control advanced laser-plasma accelerators, as well as compact free-electron laser beamlines, requires direct comparison to experimental data. We highlight recent results in start-to-end simulations and developments with a focus on their connection to experiment, such as by synthetic diagnostics and experimental data reconstruction analyses. Live-visualization for experimental data aquisition Providing data integration for laser-plasma accelerator experiments requires a flexible infrastructure that is easily extendable and can be accessed straight-forward. This not only eases the everyday operations of experimentalists, but also enables fast development and deployment of machine learning methods. Therefore, we propose a novel combination of several existing components such as mongoDB, Grafana, and Kafka, to approach this task. The data, which is ingested live during the experiments, is visualized using interactive plots in an easy-to-access Grafana dashboard. Surrogate model to identify electron injection The plasma dynamics in laser-plasma accelerators are in most cases inaccessible to direct measurements. The emitted radiation from the electrons inside the plasma, however, contains information about the electron motion and thus about the relevant plasma dynamics but is hard to interpret and thus does not allow for direct interpretation. We propose an invertible-neural-network (INN) based method, radINN, that is capable of translating the emitted radiation spectra to the amount of injected charge over time. The model was trained using synthetic radiation spectra from large-scale particle-in-cell simulations. Its capabilities in identifying the injection process using "wave-breaking radiation" already exceed the capability of human domain experts. Towards digital twins of LPA driven FELs Understanding and control of Laser-driven Free Electron Lasers remain to be difficult prob