High Accelerator Research Articles

DyNNamic: Dynamically Reshaping, High Data-Reuse Accelerator for Compact DNNs

Convolutional layers dominate the computation and energy costs of Deep Neural Network (DNN) inference. Recent algorithmic works attempt to reduce these bottlenecks via compact DNN structures and model compression. Likewise, state-of-the-art accelerator designs leverage spatiotemporal characteristics of convolutional layers to reduce data movement overhead and improve throughput. Although both are independently effective at reducing latency and energy costs, combining these approaches does not guarantee cumulative improvements due to inefficient mapping. This inefficiency can be attributed to (1) inflexibility of underlying hardware and (2) inherent reduction of data-reuse opportunities of compact DNN structures. To address these issues, we propose a dynamically reshaping, high data-reuse PE array accelerator, namely <i>DyNNamic</i> . DyNNamic leverages kernel-wise filter decomposition to partition the convolution operation into two compact stages: Shared Kernels Convolution (SKC) and Weighted Accumulation (WA). Because both stages have vastly different dimensions, DyNNamic reshapes its PE array to effectively map the algorithm to the architecture. The architecture then exploits data-reuse opportunities created by the SKC stage, further reducing data movement with negligible overhead. We evaluate our approach on various representative networks and compare against state-of-the-art accelerators. On average, DyNNamic outperforms DianNao by <inline-formula><tex-math notation="LaTeX">$8.4\times$</tex-math></inline-formula> and <inline-formula><tex-math notation="LaTeX">$12.3\times$</tex-math></inline-formula> in terms of inference energy and latency, respectively.

Innovative research and development activities on high energy and high current isochronous proton accelerator

The MW class proton accelerators are expected to play important roles in many fields, attracting institutions to continue researching and tackling key problems. The continuous wave (CW) isochronous accelerator obtains a high-power beam with higher energy efficiency, which is very attractive to many applications. Scholars generally believe that the energy limitation of the isochronous cyclotron is ∼1 GeV. To get higher beam power by the isochronous machine, enhancing the beam focusing become the most important issue.Adjusting the radial gradient of the average magnetic field makes the field distribution match the isochronism. When we adjust the radial gradient of the peak field, the first-order gradient is equivalent to the quadrupole field, the second-order, the hexapole field, and so on. Just like the synchrotron, there are quadrupoles, hexapole magnets, and so on, along the orbits to get higher energy, as all we know.If we adjust the radial gradient for the peak field of an FFA's FDF lattice and cooperate with the angular width (azimuth flutter) and spiral angle (edge focusing) of the traditional cyclotron pole, we can manipulate the working path in the tune diagram very flexibly. During enhancing the axial focusing, both the beam intensity and the energy of the isochronous accelerator are significantly increased. Here a 2 GeV CW FFA with 3 mA of average beam intensity design is presented. It is essentially an isochronous cyclotron although we use 10 FDF lattices. The key difficulty is that the magnetic field and each order of gradient should be accurately adjusted in a large radius range.As a high-power proton accelerator with high energy efficiency, we adopt high-temperature superconducting (HTS) technology for the magnets. 15 RF cavities with a Q value of 90000 provide energy gain per turn of ∼15 MeV to ensure the CW beam intensity reaches 3 mA. A 1:4 scale, 15-ton HTS magnet, and a 1:4 scale, 177 MHz cavity have been completed. The results of such R&D will also be presented in this paper.

A high speed reconfigurable architecture for softmax and GELU in vision transformer

AbstractTransformers have been widely used in various computer vision applications. Compared to traditional convolutional neural networks (CNNs), transformer's inference includes plenty of non‐linear operations, such as softmax and Gaussian error linear units (GELU). As the scale of transformers grows, an efficient hardware implementation of these operations is significant. However, the current works of computer vision neural network accelerators focus on CNN and less attention is paid to transformer. In addition, most current FPGA‐based softmax or GELU accelerators are not designed for vision transformer (ViT). To solve this problem, this work proposes a high speed reconfigurable accelerator. The architecture can support both softmax and GELU functions in ViT by reconfiguring the data path. This architecture on Xilinx XCVU37P is implemented through mathematical transformation and hardware optimization design, and achieve the performance of 102.4 Giga bits per second (Gbps) at 200 MHz. Experimental results show that the architecture achieves a very small accuracy loss in the ViT's inference by using fixed‐point 16‐bit quantization. Compared with existing accelerators, the design has greater throughput and area efficiency.

Evaluation and modelling of the lithium fluoride based thermoluminescent detector response at the CERN-EU high-energy reference field (CERF)

Lithium fluoride (LiF) thermoluminescent detectors (TLDs) are commonly used for measurements in mixed high energy radiation fields such as space, hadron radiotherapy and other high energy accelerator facilities. However, the results of these measurements are often challenging to interpret since the radiation field compositions are not known a priori. The goal of this work was to model and understand the LiF TLD response in the CERN-EU high-energy Reference Field (CERF).Six different LiF TLD types, namely MTS (LiF:Mg,Ti) and MCP (LiF:Mg,Cu,P) TLDs with different Li abundances, were irradiated at CERF. For the modelling of their response, CERF reference fluence energy spectra from FLUKA simulations and TLD relative luminescence efficiencies according to the Microdosimetric d(z) Model were collected from previous works and supplemented with new relative luminescence efficiency calculations for neutrons with energies above 0.01 MeV. Dedicated MCNP6.2 simulations were performed to obtain the required fluence to absorbed dose conversion factors.Results demonstrated that the response of LiF detectors can be well predicted even for complicated, mixed radiation fields. The modelling allowed to assess the dose contributions from different radiation types in mixed fields. Based on the results of this work, a methodology was proposed to perform LiF TLD measurements in well-characterized mixed fields in terms of absorbed dose in water or tissue. In unknown mixed radiation fields MTS-7 TLDs are recommended to assess the non-neutron component of the absorbed dose in water or tissue. Assessing the dose from neutrons requires dedicated neutron detectors.

Cherenkov Radiation in Optical Fibres as a Versatile Machine Protection System in Particle Accelerators.

Machine protection systems in high power particle accelerators are crucial. They can detect, prevent, and respond to events which would otherwise cause damage and significant downtime to accelerator infrastructure. Current systems are often resource heavy and operationally expensive, reacting after an event has begun to cause damage; this leads to facilities only covering certain operational modes and setting lower limits on machine performance. Presented here is a new type of machine protection system based upon optical fibres, which would be complementary to existing systems, elevating existing performance. These fibres are laid along an accelerator beam line in lengths of ∼100 m, providing continuous coverage over this distance. When relativistic particles pass through these fibres, they generate Cherenkov radiation in the optical spectrum. This radiation propagates in both directions along the fibre and can be detected at both ends. A calibration based technique allows the location of the Cherenkov radiation source to be pinpointed to within 0.5 m with a resolution of 1 m. This measurement mechanism, from a single device, has multiple applications within an accelerator facility. These include beam loss location monitoring, RF breakdown prediction, and quench prevention. Detailed here are the application processes and results from measurements, which provide proof of concept for this device for both beam loss monitoring and RF breakdown detection. Furthermore, highlighted are the current challenges for future innovation.

Differentiable self-consistent space-charge simulation for accelerator design

The nonlinear space-charge effects in a high intensity or high brightness accelerator can have a significant impact on the beam properties through the accelerator. These effects are included in the accelerator design via self-consistent multi-particle tracking simulations. In order to study the sensitivity of the final beam's properties with respect to the accelerator design parameters, one has to carry out the time-consuming space-charge simulation multiple times. In this paper, we propose a differentiable self-consistent space-charge simulation model that enables the study of the beam sensitivity through only one simulation. Such a model can also be used with gradient-based numerical optimizers for accelerator design optimizations including the self-consistent space-charge effects.

Increased storage stability of marble adhesive based on unsaturated polyester resin

In this paper, different combinations of accelerators and inhibitors were used to cure marble adhesive based on an unsaturated polyester resin. N, N- dimethyl aniline (DMA, NL-63), N, N-dimethylparatoluidine (DEP, NL-65), and ethoxylates of p-toluidine (Pergaquick A150 p.m.) were used as amine accelerators. Benzoyl peroxide (BPO) was used as a medium temperature decomposition initiator. Hydroquinone and 1,4- naphthoquinone were used as polymerization inhibitors. To study the curing behavior, gel time, pseudo-adiabatic exothermic, and storage stability measurements have been used. The low-temperature initiator decomposition rate has also influenced the exothermic peak and cure rate at various accelerators. It is shown in this study that for N, N- dimethyl aniline (DMA, NL-63), N, N-dimethylparatoluidine (DEP, NL-65) and Ethoxylates of p-toluidine (Pergaquick A-150) promoters systems, with the increasing of pergaquick A-150 concentration, namely high reactive accelerator, causes the exothermic parameters sharply change. It might be due to the more reactive nature of pergaquick A-150 in comparison with other accelerators reactivity. Due to this study, using a dual system of inhibitor, the storage stability of marble adhesive was increased. Now we can say, a judicious choice of a dual inhibitor and a dual accelerator can prevent short time exothermic reactions with high storage stability, so a dual system of the inhibitor can be much more effective than a single one.

Tutorial: Modeling of the extraction and acceleration of negative ions from plasma sources using particle-based methods

In this Tutorial, we consider plasma sources with applications to fusion devices and high energy accelerators. These ion sources typically produce negative ions from hydrogen-isotope gases, which are extracted through one or multiple apertures and accelerated to high kinetic energies. Next, they are either double stripped of two electrons to form positive ions used as precursors in accelerator devices or neutralized to produce a neutral beam injected in tokamak reactors. Contrary to the working conditions of most ion sources where volume production prevails, the mechanism of negative ion production by dissociative electron attachment on vibrationally excited molecules inside the plasma volume of fusion-type hydrogen-fueled high power discharges is mostly balanced by their destruction by detachment before being extracted rendering this means of producing negative ions rather inefficient. Surface production through the transfer of electrons from low work function metallic materials to the impacting atoms is the alternative solution to fulfill the requirements for the applications concerned. Negative ions are produced close to the aperture from which they are extracted. As a result, the analysis and understanding of the extraction mechanisms through experimental diagnostics is rather difficult due to the lack of accessibility and can only give a partial view. In addition, most of the experimental work is focused on the validation of requirements for the applications and not to the investigation of the fundamental processes that take place inside these types of sources. This Tutorial is focused on the description and understanding of the physical mechanisms behind the extraction and acceleration of negative ions from hydrogen plasma sources through modeling methods. We describe the numerical techniques of particle-based methods with a specific emphasis on particle-in-cell Monte Carlo collision algorithms. An analysis of the physical processes involved in driving the negative ions from the plasma source, across the apertures and inside the accelerator as reported in the literature, is presented in detail. This Tutorial concludes with additional and future works to be addressed in the coming years.

Study on baking deformation of the thin-walled arc chamber

A large scientific facility, High Intensity heavy-ion Accelerator Facility (HIAF), was being built to study basic and interdisciplinary sciences. To meet the requirements of magnetic rigidity for high-energy heavy ions, pulsed dipole magnets were used in Booster Ring (BRing). Therefore, the ceramic-lined thin-walled arc chambers with 0.3 mm shell were developed to reduce the eddy current resulting from the rapidly changing magnetic field. Furthermore, to satisfy the requirements of the beam lifetime, the pressure of BRing of the HIAF must be lower than 1 × 10−9 Pa. In-situ baking methods were used to obtain ultra-high vacuum. As the stiffness of the thin-walled arc chamber was poorer for evacuation and in-situ baking as a conventional way, it had to be controlled in the deformation and deflection angle. Thus, a deformation control method using springs was adopted. The results show that the longitudinal deformation of both different thin-walled chambers was controlled to within 0.3 mm, the radial deformation was controlled to within 0.8 mm and the variation of deflection angle was within 0.2°.

Proposal of a VHEE Linac for FLASH radiotherapy

Translation of electron FLASH radiotherapy in clinical practice requires the use of high energy accelerators to treat deep tumours and Very High Electron Energy (VHEE) could represent a valid technique to achieve this goal. In this scenario, a VHEE FLASH linac is under study at the University La Sapienza of Rome (Italy) in collaboration with the Italian Institute for Nuclear Research (INFN) and the Institut Curie (France). Here we present the preliminary results of a compact C-band system aiming to reach an high accelerating gradient and an high pulse current necessary to deliver high dose per pulse and ultra-high dose rate required for FLASH effect. We propose a system composed of a low energy high current injector linac followed by a modular section of high accelerating gradient structures. CST code is used to define the required LINAC’s RF parameters and beam dynamics simulations are performed using TSTEP and ASTRA.

Time-Series Deep Learning Anomaly Detection for Particle Accelerators

High energy particle accelerators rely on superconducting radio frequency cavities to transfer energy and accelerate the beam. Such particle accelerators are complex and expensive systems prone to failures which lead to downtime of the whole experimental facility: it is thus of primary importance to anticipate and prevent these faults to improve the uptime and cost-effectiveness of particle accelerators. Data-driven methods are especially fit for this task as they can leverage all the data recorded and archived by a typical control system. Previous works used classical machine learning (ML) models for anomaly detection to detect early signs of an upcoming fault. We propose here a different approach based on deep learning (DL) models, exploiting the temporal correlation of the raw data. Three different models are tested on data from the ALPI (Linear Accelerator for Ions) linear accelerator in INFN (National Institute for Nuclear Physics) Legnaro National Laboratories in Italy and they are compared with the classical ML approach.

Thermal moderator-reflector design of the 24 Hz target station for the High Brilliance Neutron Source

The High Brilliance Neutron Source (HBS) project develops a high current accelerator driven neutron source (HiCANS), aiming to substitute existing neutron research reactors, that reach the end of life. This study focuses on the thermal neutron moderator design for a low frequency, long pulse target station of HBS. We investigate the neutronic characteristics of the D2O+H2O mixture moderator material using the Monte Carlo particle transport code PHITS. Our findings show that, compared to D2O with lower neutron absorption, confinement of H2O is the dominant factor in achieving a high neutron flux within the moderator and thus a high neutron brightness. We find that surrounding the central thermal water moderator with a D2O+H2O mixture can achieve a better balance between confinement and absorption. However, because of multiple angles of reflecting neutron, this design will also increase the divergence of the neutron beam.

Towards a direct measurement of the 17O(p, γ)18F 65 keV resonance strength at LUNA

The 17O(p, γ)18F reaction plays a crucial role in the hydrogen burning phases of different stellar scenarios. At temperature of interest for AGB nucleosynthesis (20 MK &lt; T &lt; 80 MK) the main contribution to the astrophysical reaction rate comes from the poorly constrained 65 keV resonance. The strength of this resonance is presently determined only through indirect measurements, with a reported value of ωγ = (1.6 ± 0.3) 10−11 eV. With typical experimental quantities for beam current, isotopic enrichment and detection efficiency, this strength yields to an expected count rate of less than one count per Coulomb, making the direct measurement of this resonance extremely challenging. A new high sensitivity setup has been installed at LUNA (Laboratory for Underground Nuclear Astrophysics) of Laboratori Nazionali del Gran Sasso. The high performance LUNA 400kV accelerator underground location guarantees, indeed, a reduction of cosmic ray background by several orders of magnitude. The residual background was further reduced by a devoted shielding of lead and borated (5%) polyethylene. On the other hand, the 4π BGO detector efficiency was optimized installing aluminum target chamber and holder. With about 400 C accumulated on Ta2O5 targets, with nominal 17O enrichment of 90%, the LUNA collaboration has performed the first direct measurement of the 65 keV resonance strength.

The PSI meson target facility and its upgrade IMPACT-HIMB

The high intensity proton accelerator complex (HIPA) at the Paul Scherrer Institute (PSI) delivers a 590 MeV CW proton beam with currents up to 2.4 mA (1.4 MW). Besides two spallation targets for thermal/cold neutrons (SINQ) and for ultracold neutrons (UCN), the beam feeds two meson production targets Target M and Target E. The targets consist of graphite wheels of effective thickness 5 mm (M) and 40/60 mm (E). The target stations M and E are of quite different design; however, both of them rotate at 1 Hz to dissipate the heat (20 kW/mA for the 40 mm target E) efficiently. Recent progress was made by a new type of bearings and a new target geometry able to increase the rate of surface muons by up to 50 %. This is also foreseen for the upgrade of the target station M within the High Intensity Muon Beam (HIMB) initiative aiming to increase the surface muons available for experiment by two orders of magnitude. HIMB is part of IMPACT (Isotope and Muon Production with Advanced Cyclotron and Target Technology), an application for the Swiss Roadmap of Research Infrastructure.

The High Brilliance Neutron Source Target Stations

High Current Accelerator driven Neutron Sources (HiCANS) are a cost-efficient alternative for future large scale neutron facilities. They excel in transforming neutrons released by the primary nuclear reaction into a spectral range usable for applications. In particular, the cost of a target station represents only a minor fraction of the overall construction cost due to the lower energy of the primary neutrons, which requires less shielding. They can be designed to provide optimized pulse and spectral properties for applications in neutron scattering, analytics and imaging experiments. For the High Brilliance neutron Source (HBS) project at Forschungszentrum Jülich, we have developed a modular design that meets the radiation protection requirements while providing sufficient space in its core to adapt the target-moderator-reflector assembly to the different applications. In the following, we present the basic target station design which will be used at HBS for three different realizations with their own instrument suite. All relevant components have been designed, built and tested at the JULIC Neutron Platform which has produced neutrons since December 12th 2022. The simulated performance of a target station shows that the brightnesses of the moderators are in the range of modern research reactors and sub-MW power spallation sources.

BOHA: A High Performance VSLAM Backend Optimization Hardware Accelerator using Recursive Fine-Grain H-Matrix Decomposition and Early-Computing with Approximate Linear Solver

Visual simultaneous localization and mapping (VSLAM) technology has been used in many intelligent applications such autonomous car, unmanned aerial vehicle (UAV) and augmented reality (AR) for localization and trajectory generation. The VSLAM usually contains the frontend computation and the backend optimization. The backend optimization is an essential part of VSLAM which helps improve the precision and dominates the computation time ( 80 VSLAM. However, it includes large inner data dependency which is difficult to be accelerated using GPU. To address this issue, in this work, we propose a high performance VSLAM backend optimization hardware accelerator (named BOHA). Diverse design techniques have been proposed to reduce computation latency by avoiding the data redundancy in the backend optimization, including 1) a recursive fine-grain H-matrix decomposition computing architecture to improve the hardware utilization and reduce the data movement. 2) an early-computing technique with approximate linear solver to reduce the computation latency while maintaining high precision. 3) a twostage deep Schur elimination technique to reduce the computation latency. The proposed design has been implemented and evaluated on a Xilinx FPGA ZCU104 and achieves low computation latency (5.1 ms) with low trajectory error (0.27 state-of-the-art designs.

Design of a cylindrical corrugated waveguide for a collinear wakefield accelerator

We present the design of a cylindrical corrugated waveguide for use in the A-STAR accelerator under development at Argonne National Laboratory. A-STAR is a high gradient, high bunch repetition rate collinear wakefield accelerator that uses a 1-mm inner radius corrugated waveguide to produce a $90\text{ }\text{ }\mathrm{MV}\text{ }{\mathrm{m}}^{\ensuremath{-}1}$, 180-GHz accelerating field when driven by a 10-nC drive bunch. To select a corrugation geometry for A-STAR, we analyze three types of corrugation profiles in the overmoded regime with $a/\ensuremath{\lambda}$ ranging from 0.53 to 0.67, where $a$ is the minor radius of the corrugated waveguide and $\ensuremath{\lambda}$ is the free-space wavelength. We find that the corrugation geometry that optimizes the accelerator performance is a rounded profile with vertical sidewalls and a corrugation period $p\ensuremath{\ll}a$. Trade-offs between the peak surface fields and thermal loading are presented along with calculations of pulse heating and steady-state power dissipation. In addition to the ${\mathrm{TM}}_{01}$ accelerating mode, properties of the ${\mathrm{HEM}}_{11}$ mode and contributions from higher order modes are discussed.

Model of a 5-MA Linear-Transformer-Driver Accelerator: Transmission-Line-Circuit Method and Three-Dimensional Field-Circuit Coupling Method

A 5-MA-LTD-based, high current accelerator is modeled using transmission line code. And the machine performances are presented for driving short-circuit loads, <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$Z$ </tex-math></inline-formula> -pinch loads, and bremsstrahlung electron beam diodes. With simulation analysis, it is expected that the short-circuit load current can reach 4.7 MA in 100 ns with ±80 kV LTD charge voltage, and the peak of <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$Z$ </tex-math></inline-formula> -pinch load current when the machine is conducted with a 0.45 mg wire-array can be high as 4.2 MA, the peak kinetic energy is 49 kJ, and the energy coupling efficiency is estimated as 38%. The physical model of the reflex triode load which is typical electron beam diode is presented in detail. By incorporating the physics-based reflex triode model into a 3-D transient electromagnetic model, the field-circuit coupling simulation can be performed and the results are compared with the circuit model. The results from these two methods are in good agreement. The peak load current when the machine is configured with a <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$\sim 0.15~\Omega $ </tex-math></inline-formula> reflex triode load is ~2.8 MA. The energy coupling efficiency to the reflex triode load is 35.8%.

Suppression of eddy-current effects in beam injection using a pulsed sextupole magnet with a new ceramic chamber

At the Photon Factory storage ring (PF ring) of the High Energy Accelerator Research Organization (KEK), a pulsed sextupole magnet (PSM) has been used for top-up injection for the first time. Following its successful operation at the KEK, this new injection scheme was expanded to one with a nonlinear kicker in BESSY II and was then applied at UVSOR and Aichi SR in Japan. Also, the PSM system in the PF ring was improved by introducing a rectangular shape for the magnet's bore to strengthen the magnetic field. However, this upgraded PSM was not implemented for user operation because it induced a horizontal oscillation of the stored beam with a maximum amplitude of 5.0 mm. This oscillation was assumed to be caused by eddy currents originating from the uniform inner coating of the ceramic chamber. We, therefore, developed a modified ceramic chamber with a new patterned coating designed to suppress the eddy current--induced magnetic field. The pattern shape was optimized to suppress this magnetic field and, simultaneously, to reduce the resistive wall impedance. A fine-line pattern-coating process, originally developed for a ceramic chamber with an integrated pulsed magnet, was modified for application with the magnet's chamber of the sextupole. The maximum horizontal oscillation amplitude was suppressed by a factor of 16 by using the new PSM chamber without any issues related to beam instability. Details of the historical background of the upgrading of the PSM system, an investigation of the cause of the abnormal oscillation problem, the development of the new process for coating the ceramic chamber, and the results of the beam-response measurements for the new ceramic chamber installed in the PF ring are reported in this article.

Classifying exploratory experimentation – three case studies of exploratory experimentation at the LHC

Along three measurements at the Large Hadron Collider (LHC), a high energy particle accelerator, we analyze procedures and consequences of exploratory experimentation (EE). While all of these measurements fulfill the requirements of EE: probing new parameter spaces, being void of a target theory and applying a broad range of experimental methods, we identify epistemic differences and suggest a classification of EE. We distinguish classes of EE according to their respective goals: the exploration where an established global theory cannot provide the details of a local phenomenon, exploration of an astonishing discovery and exploration to find a new entity. We find that these classes also differ with respect to the existence of an identifiable target and their impact on the background theory. The characteristics distinguish EE from other kinds of experimentation, even though these different kinds have not yet been systematically studied. The formal rigor and precision of LHC physics facilitates to analyze concept formation in its early state. In particular we emphasize the importance for nil–results for conceptualization and argue that conceptualization can also be achieved from nil–results only.