Filtern
Dokumenttyp
- Vortrag (5)
- Zeitschriftenartikel (3)
- Beitrag zu einem Tagungsband (1)
- Sonstiges (1)
Sprache
- Englisch (10)
Schlagworte
- Neural network (10) (entfernen)
Organisationseinheit der BAM
Eingeladener Vortrag
- nein (5)
Getting more efficient – The use of Bayesian optimization and Gaussian processes at the BAMline
(2022)
For more than 20 years, BAM is operating the BAMline at the synchrotron BESSY II in Berlin Adlershof. During this time, the complexity of the setup and the amount of data generated have multiplied. To increase the effectiveness and in preparation for BESSY III, algorithms from the field of machine learning are increasingly used.
In this paper, several examples in the areas of beamline alignment and measurement time optimization based on Bayesian optimization (BO) with Gaussian processes (GP) are presented. BO is a method for finding the global optimum of a function using a probabilistic model represented by a GP. The advantage of this method is that it can handle high-dimensional problems, does not depend on the initial estimate, and also provides uncertainty estimates.
After a short introduction to BO and GP, the first example is the automatic alignment of our double multilayer monochromator (DMM). To achieve optimal performance, up to three linear and two angular motor positions have to be optimized. To achieve this with a grid scan, at least 100^5 measurement points would be required. Assuming that all positions can be aligned independently, 100*5 points are still necessary. We show that with BO and GP less than 100 points are sufficient to achieve equal or better results.
The second example is the optimization of measurement time in XRF scanning. Here we will show the advantage of the BO GP approach over point-by-point scanning. As can be seen in Fig. 1, the number of points required and thus the measurement time can be reduced by a factor of 50, while the loss in image quality is acceptable. The advantages and limitations of this approach will be discussed.
In this talk an overview about artificial intelligence/machine learning applications @BAMline is given. In the first part, the use of neural networks for the quantification of XRF measurements and the decoding of coded-aperture measurements are shown. Then it is shown how Gaussian processes and Bayesian statistics can be used to achieve an optimal alignment of the set-up and in general for optimization of measurements.
Learned block iterative shrinkage thresholding algorithm for photothermal super resolution imaging
(2022)
Block-sparse regularization is already well known in active thermal imaging and is used for multiple-measurement-based inverse problems. The main bottleneck of this method is the choice of regularization parameters which differs for each experiment. We show the benefits of using a learned block iterative shrinkage thresholding algorithm (LBISTA) that is able to learn the choice of regularization parameters, without the need to manually select them. In addition, LBISTA enables the determination of a suitable weight matrix to solve the underlying inverse problem. Therefore, in this paper we present LBISTA and compare it with state-of-the-art block iterative shrinkage thresholding using synthetically generated and experimental test data from active thermography for defect reconstruction. Our results show that the use of the learned block-sparse optimization approach provides smaller normalized mean square errors for a small fixed number of iterations. Thus, this allows us to improve the convergence speed and only needs a few iterations to generate accurate defect reconstruction in photothermal super-resolution imaging.
This article presents deep unfolding neural networks to handle inverse problems in photothermal radiometry enabling super-resolution (SR) imaging. The photothermal SR approach is a well-known technique to overcome the spatial resolution limitation in photothermal imaging by extracting high-frequency spatial components based on the deconvolution with the thermal point spread function (PSF). However, stable deconvolution can only be achieved by using the sparse structure of defect patterns, which often requires tedious, handcrafted tuning of hyperparameters and results in computationally intensive algorithms. On this account, this article proposes Photothermal-SR-Net, which performs deconvolution by deep unfolding considering the underlying physics. Since defects appear sparsely in materials, our approach includes trained block-sparsity thresholding in each convolutional layer. This enables to super-resolve 2-D thermal images for nondestructive testing (NDT) with a substantially improved convergence rate compared to classic approaches. The performance of the proposed approach is evaluated on various deep unfolding and thresholding approaches. Furthermore, we explored how to increase the reconstruction quality and the computational performance. Thereby, it was found that the computing time for creating high-resolution images could be significantly reduced without decreasing the reconstruction quality by using pixel binning as a preprocessing step.
Learned block iterative shrinkage thresholding algorithm for photothermal super resolution imaging
(2020)
Block-sparse regularization is already well-known in active thermal imaging and is used for multiple measurement based inverse problems. The main bottleneck of this method is the choice of regularization parameters which differs for each experiment. To avoid time-consuming manually selected regularization parameters, we propose a learned block-sparse optimization approach using an iterative algorithm unfolded into a deep neural network. More precisely, we show the benefits of using a learned block iterative shrinkage thresholding algorithm that is able to learn the choice of regularization parameters. In addition, this algorithm enables the determination of a suitable weight matrix to solve the underlying inverse problem. Therefore, in this paper we present the algorithm and compare it with state of the art block iterative shrinkage thresholding using synthetically generated test data and experimental test data from active thermography for defect reconstruction. Our results show that the use of the learned block-sparse optimization approach provides smaller normalized mean square errors for a small fixed number of iterations than without learning. Thus, this new approach allows to improve the convergence speed and only needs a few iterations to generate accurate defect reconstruction in photothermal super resolution imaging.
In this work, we describe the use of artificial neural networks for the quantification of X-ray fluorescence measurements. The training data were generated using Monte Carlo simulation, which avoided the use of adapted reference materials. The extension of the available data set by means of an ANN to generate additional data was demonstrated. Particular emphasis was put on the comparability of simulated and experimental data and how the influence of deviations can be reduced. The search for the optimal hyperparameter, manual and automatic, is also described. For the presented case, we were able to train a network with a mean absolute error of 0.1 weight percent for the synthetic data and 0.7 weight percent for a set of experimental data obtained with certified reference materials.
In X-ray fluorescence (XRF), a sample is excited with X-rays, and the resulting characteristic radiation is detected to detect elements quantitatively and qualitatively. Quantification is traditionally done in several steps:
1. Normalization of the data
2. Determination of the existing elements
3. Fit of the measured spectrum
4. Calculation of concentrations with fundamental parameters / MC simulations / standard based
The problem with standard based procedures is the availability of corresponding standards. The problem with the calculations is that the measured intensities for XRF measurements are matrix-dependent. Calculations must, therefore, be performed iteratively (= time consuming) in order to determine the chemical composition.
First experiments with gold samples have shown the feasibility of machine learning based quantification in principle. A large number of compositions were simulated (> 10000) and analyzed with a deep learning network. For first experiments, an ANN (Artificial Neural Network) with 3 hidden layers and 33x33x33 neurons was used. This network learned the mapping of spectra to concentrations using supervised learning by multidimensional regression. The input layer was formed by the normalized spectrum, and the output layer directly yielded the searched values. The applicability for real samples was shown by measurements on certified reference materials.
Artificial intelligence (AI), machine learning, and neural networks have revolutionized fields such as self-driving cars or machine translation. Indeed, AI has progressed so far that scientists such as Stephen Hawking now fear that it may destroy humankind altogether. So, let’s get started and use AI in surface science.
Here we train neural networks to use raw measurement data as input and immediately return the desired output parameters. We discuss the example of X-ray reflectivity measurements of ultrathin films, which contain reciprocal space information and must traditionally be fitted with dynamic scattering theory. Instead, we train the neural network with simulated X-ray data of multilayer structures and then apply it to measurement data. The neural network yields high accuracy results, is robust against noise, and performs significantly faster than fitting algorithms in real-time experiments. The presented neural network data analysis is becoming increasingly attractive, because free software has become very accessible and specialized computer chips (NPU) are currently being rolled out.
An approach to develop an arc sensor for gap-width estimation during automated NG-GMAW with a weaving electrode motion is introduced by combining arc sensor readings with optical measurement of the groove shape to allow precise analyses of the process. The two test specimen welded for this study were designed to feature a variable groove geometry in order to maximize efficiency of the conducted experimental efforts, resulting in 1696 individual weaving cycle records with associated arc sensor measurements, process parameters and groove shape information. Gap width was varied from 18 to 25 mm and wire feed rates in the range of 9 to 13 m/min were used in the course of this study. Artificial neural networks were used as a modelling tool to derive an arc sensor for estimation of gap width suitable for online process control that can adapt to changes in process parameters as well as changes in the weaving motion of the electrode. Wire feed rate, weaving current, sidewall dwell currents and angles were used as inputs to calculate the gap width. Evaluation the proposed arc sensor model show very good estimation capabilities for parameters sufficiently covered during experiments.
An approach to develop an arc sensor for gap-width estimation during automated NG-GMAW with a weaving electrode motion is introduced by combining arc sensor readings with optical measurement of the groove shape to allow precise analyses of the process. The two test specimen welded for this study were designed to feature a variable groove geometry in order to maximize efficiency of the conducted experimental efforts, resulting in 1696 individual weaving cycle records with associated arc sensor measurements, process parameters and groove shape information. Gap width was varied from 18 to 25 mm and wire feed rates in the range of 9 to 13 m/min were used in the course of this study. Artificial neural networks were used as a modelling tool to derive an arc sensor for estimation of gap width suitable for online process control that can adapt to changes in process parameters as well as changes in the weaving motion of the electrode. Wire feed rate, weaving current, sidewall dwell currents and angles were used as inputs to calculate the gap width. Evaluation the proposed arc sensor model show very good estimation capabilities for parameters sufficiently covered during experiments.