Online Feature Extraction Algorithms For High Energy Physics Based On Convolutional Neural Networks

The high energy physics(HEP)experiment featuring large research facilities(eg.accelerators)is an important subject in modern physics.It ranges from the theoretical study based on the experimental evidences to the development of detectors and electronics for engineering practice.At present,the HEP experiments are approaching the "energy frontier" and the "precision frontier".With the increase of the event rates and data amount in an event,the researchers have realized that data analysis techniques based on traditional statistics measures are hard to meet the needs in the next surge of detector upgrades,especially when the background events are dominant,the valid data are overwhelming and online triggering and analysis are inevitable.Among the possible aspects,the high dimensionality of data,the pile-up of data/events and the highly structured features are major challenges facing the HEP data analysis.Entering the second decade of the twenty-first century,"deep learning" as a representative of artificial intelligence techniques is developing rapidly and becoming an independent and sophisticated research field It has permeated into the experimental physics and many other areas and drawn great attention as a leading-edge research method.To tackle the high dimensionality,pile-up and structured features,new data analysis tools are needed to utilize the inner features of data so as to reduce the time complexity and improve the precision at the same time.In computer vision,the most successful area of deep learning,new neural network structures are iterating fast,most of which are based on "convolutional neural networks"(CNN).Compared to traditional feed-forward networks,CNNs take advantage of the parameter sharing and the translation invariance,so that they facilitate the training and optimization of the model and improve the performance while keeping the amount of parameters moderate.Results from both theory and practice demonstrate that CNN-based deep learning methods solve the "curse of dimensionality",discriminate multiple instances in the space and make use of the structured features.Inspired by the idea,this thesis proposed several CNN-based network architectures to accomplish application-specific physical tasks to deal with data in different dimensions.The major works and innovations include the following aspects:1.In the neutrinoless double beta decay(0???)experiment with time projection chambers and the Topmetal readout plane,we created a CNN in the three-dimensional(3D)space,which improved the signal/background discrimination and rejected dominant backgrounds.After a concise review of the 0??? experiment principles and relevant equipments,this thesis put forward the residual block in the ResNet to construct the 3D CNN for the first time.To demonstrate the capability of the neural network and its relation to the theoretical upper bound,we devised a toy model of the 0???experiment in time projection chambers,and compared neural networks with bounds from theory.After generating physical simulation data with the 0v?? experiment,this thesis compared the two-dimensional CNN and 3D CNNs with different depths,which confirmed the importance of the 3D structure and the network depth in this problem.When accepting the same ratio of signal events,the background acceptance rate decreased from 11 percent(traditional method)to 4 percent(3D ResNet).Regarding the granularity,diffusion and noise when implementing the detector equipment,thorough simulations showed that the proposed neural network was robust to these variables.2.In the beam measurement system with the gaseous drift chambers and big-array Topmetal detectors,we created an end-to-end network based on segmentation and fitting.This network could acquire the location and orientation of multiple particles without prior knowledge about the number of the particles.The network architecture was comprised of the base network,the binary segmentation part and the semantic segmentation part.The latter two shared the base network and performed initial track determination and precise regression individually.Based on the segmentation parts,this thesis designed a weighted least squares operation implemented inside the deep learning framework to combine the results of both parts.The whole network,including the least squares fitting,could be trained with back-propagation in an end-to-end manner.Besides,this thesis introduced a "center-angle measure" to judge the precision of location and orientation by combining two separate factors.The initial position resolution could achieve 8.8 ?m for the single track and 11.4 ?m for the multiple tracks,and the angle resolution could achieve 0.15° and 0.21°respectively.3.Aimed at the time series of pulses from the PHOS calorimeter for the ALICE experiment,this thesis proposed a denoising autoencoder(DAE)based on one-dimensional CNNs and improved the timing precision more than 20%in the laboratory environment compared to curve fitting.The network contained the DAE part and the regression network part.In the DAE,we used skip connections and channel concatenation to preserve long-range relations.In consideration of three kinds of variations(long-term drift,short-term change and random noise)in HEP experiments,this thesis analyzed the merits and limitations of the traditional curve fitting method,and inferred the advantages of CNNs and deep learning by a comparative study.As regards three kinds of variations,this thesis performed comprehensive simulations to demonstrate the stability of neural networks in non-ideal conditions.Lastly,experiments were conducted with the front-end electronics of the PHOS calorimeter in real-world conditions.Using the specification of 100 ns shaping time,the timing resolution of the proposed neural network could achieve 1.37±0.03 ns.4.On the basis of network architectures and algorithms for time series,we designed the first application specific integrated circuit(ASIC)to accelerate neural network models in HEP experiments.To handle the high throughput in the upgrades of HEP detectors,this thesis proposed to use neural computing ASICs at the front-end electronics to perform specific information-extracting tasks to reduce the amount of transmitted data.The designed ASIC,named PulseDL,targeted the one-dimensional time series with a system structure combining the RISC processor and the customized processing array.In the light of physical tasks,detailed hardware-software co-design was made with an emphasis on the parallelism of convolution and effects of fixed-point quantization.With the GSMCR013 digital process,logic synthesis and physical layout were done,and the final area of the chip was 24 mm2.The simulation and validation results showed that the peak power efficiency could achieve 12.351 GOPS/W.At last,the chip was tested under the hardware condition in the laboratory,and the ability to support multiple channels was demonstrated.