Search NASA⌕ Search

SEARCH · Search NASA

Results for “dynamic compression”

Search indexed NASA NTRS and DOE OSTI research on propulsion, heat transfer, battery materials and energy systems. Follow report and document links to the original sources.

Quote a phrase for an exact phrase match. Source license links do not imply unrestricted reuse.

At least 73 records · Page 4
All sourcesNASA NTRSDOE OSTI
DOE OSTI2025

Further Development of the Tamped Richtmyer-Meshkov Instability Method and Application to Molybdenum Dynamic Strength Calibration and Tabulation

The high pressure and high strain rate dynamic strength of Mo is experimentally and computationally investigated in the 3–20 GPa stress, 50–600 C temperature, and 10 5 –10 6 /s strain rate regimes using a modified tamped Richtmyer-Meshkov instability (RMI) method. Modifications to the tamped RMI method include a method to determine loading states during strain, a new strength calibration function based on interface shape, and a robust uncertainty quantification method. These modifications improve fidelity of the tamped RMI method, allowing evaluation of the compensating effects of pressure hardening, strain rate hardening, strain hardening, and thermal softening. The new calibration function based on interface shape is not limited to sinusoidal corrugations and could be applied to additional interface shapes. Plate impact experiments are performed at Argonne National Laboratory’s Advanced Photon Source’s Dynamic Compression Sector operated by Washington State University (DCS), driving a planar shock front through a corrugated Mo-D 2 O or Mo-C 8 F 18 interface, forcing the corrugation to significantly deform. The extent of interfacial deformation, RMI growth, is experimentally observed using X-ray phase contrast imaging at the DCS. RMI jet lengths and jet shapes are extracted from the experimental radiographs, then used to calibrate numerical simulations performed with the Sandia National Laboratories (SNL) hydrocode CTH. Mo yield strength, Y, as a function of shock pressure, P, strain rate, $\dot{\varepsilon }$, accumulated strain, ϵ, relative volumetric compression, RD , and temperature, T , is determined for each impact experiment and presented. The calibrated Mo yield strength values range 1.2–1.8 GPa, with strength generally decreasing as the impact stress increases. This trend is likely caused by thermal softening or strain localization. The tabular yield strength versus loading condition data presented in this paper can be used to fit complex strength models.

Voorhees, T. J. [Sandia National Lab. (SNL-CA), Li↗
DOE OSTI2025

Machine learning for domain transfer between simulated and experimental 2D X-ray diffraction patterns using generative adversarial networks

X-ray diffraction (XRD) is a well-established technique for analyzing materials at an atomic level. Dynamic compression experiments (DCE), in which materials are subject to extreme pressures, can provide fundamental understanding to pressure-induced phase transitions and compression of the crystal lattice. The analysis of XRD patterns from highly compressed samples is non-trivial given the sparsity of data, high experimental costs, and the fact that the data is often marred with X-ray background and other artifacts. While accurate computational frameworks exist, they solve the forward problem—from structures and orientations to XRD patterns. Solving the inverse problem for 2D experimental diffraction patterns is currently a complex manual process of matching and comparing experimentally observed patterns to computationally generated ones. Machine learning is a promising tool for automating the matching process but often requires data-intensive architectures. Here, in this study, we use a CycleGAN to translate the domain of limited experimental data to a domain in which there is readily available simulated data. This domain shift allows data-intensive machine learning models that have only been trained on simulated XRD patterns to be used in the analysis of experiments.

Brozak, Samantha Jean [Sandia National Laboratorie↗
DOE OSTI2024

High strain-rate strength response of single crystal tantalum through in-situ hole closure imaging experiments

The properties of crystalline materials often depend on directionality and operating conditions. Specifically, the strength of materials can depend anisotropically on crystal direction and the loading condition. To probe these effects, a preliminary series of high strain-rate (> 105/s) strength plate-impact hole closure experiments were performed on high purity single crystal Tantalum cubes. The orientation of the single crystals with respect to impact/loading were varied to provide data to inform crystal plasticity modeling efforts. The experiments consist of in-situ high-resolution X-ray radiographic imaging of the hole collapse under dynamic compression conditions to infer the material strength via its resistance to closure at increasing levels of plastic strain. The experiments are compared against hydrocode simulation predictions. Here, a comparison with simple elastic perfectly plastic strength model predictions is presented to elucidate the response of the different crystal orientations at high strain-rate and large plastic strains.

36 MATERIALS SCIENCE↗
DOE OSTI2025

A surprising proliferation of detwinning in β -tin at extreme loading rates

Integrating data from dynamic compression experiments of condensed matter across three national laboratories has led to insight and quantitative calibration of materials strength over decades of loading rate. For many materials, a single strength model (such as PTW) is sufficient to capture the flow-stress strain rate relationship which is monotonic. Here, we show here that β -tin, a tetragonal metal, exhibits dramatic deviations from this behavior. Naive fitting to a single PTW model is insufficient to capture the behavior; indeed, such resulting inferred flow stress versus strain exhibits a non-monotonic behavior. We suggest a resolution to this by proposing that in β -tin there are important Bauschinger effects arising from favorable conditions for twinning and detwinning. A simple yield surface model when paired with PTW hardening captures the experimental data.

36 MATERIALS SCIENCE↗
DOE OSTI2024

Static and Dynamic Thermomechanical Properties of Phase-Separated Epoxy Networks with Tuned Microstructures

Here, polymerization-induced phase separation is a useful method for the construction of heterogeneous epoxy networks with properties exceeding their homogeneous counterparts. In this work, we examine the static and dynamic thermomechanical properties of phase-separated epoxy networks salient to their application as encapsulants. Three heterogeneous epoxy-amine networks with nano-, meso-, and macro-phase-separated morphologies comprised of hard and soft domains are compared to a rigid, unstructured network. The glass transition profiles of the heterogeneous networks are complex, spanning many decades in the frequency domain. The nanophase-separated morphology leads to higher coefficient of thermal expansion, yet surprisingly is characterized by reduced residual stress. Under both quasi-static and dynamic compression (strain rates of order 10 –3 and 10 3 s –1 , respectively), the nanophase-separated network also exhibits higher modulus and strength. In split-Hopkinson bar experiments, the energy dissipation characteristics of the epoxy networks were nearly identical. Curiously, however, the Hugoniot response of the macro-phase-separated network determined by ballistic shockwave analysis indicates a remarkable ability of this material to mitigate shockwave propagation in comparison to many homogeneous and heterogeneous polymer materials. Collectively, this work reveals several previously unreported phenomena with respect to structure–property relationships in phase-separated epoxy networks, illustrating the potential value of systematically tuned microstructures for optimization of application-specific physical properties.

36 MATERIALS SCIENCE↗
DOE OSTI2024

Quantifying motion blur by imaging shock front propagation with broadband and narrowband X-ray sources

Time-integrated radiography using MeV Bremsstrahlung X-ray sources is the norm for imaging during system-level testing of components and structures under dynamic condition. One source of error in the analysis of the time-integrated radiography data sets stems from motion blur which smears out sharp interfaces to a greater degree with longer exposure times, which become necessary to provide sufficient signal-to-noise with low X-ray penetration of objects of interest. To quantify motion blur, a 1D shock wave through PMMA was investigated experimentally at The Dynamic Compression Sector at The Advanced Photon Source (DCS@APS) with tapered broadband and 25.46 ± 1.06 keV narrowband X-rays. Four cameras with different exposure times were used for each experiment to compare the effect that exposure time has on motion blur. In addition, our methodology to accurately simulate motion blur in terms of transmission and shape is presented and compared to our experimental results and quantified. There is a high level of agreement between the experimental and simulation results across the range of data sets investigated in this study with a percent difference range of 0.29–1.31% for the four shots. The methodology of this work serves as a steppingstone towards a physically validated model that could be used in conjunction with experimental results to deconvolve physical parameters, densities, and interfaces of interest in a way that would not be possible with experimental results alone.

46 INSTRUMENTATION RELATED TO NUCLEAR SCIENCE AND ↗
DOE OSTI2024

Developing time-resolved x-ray diffraction diagnostics at the National Ignition Facility (invited)

As part of a program to measure phase transition timescales in materials under dynamic compression, we have designed new x-ray imaging diagnostics to record multiple x-ray diffraction measurements during a single laser-driven experiment. Our design places several ns-gated hybrid CMOS (hCMOS) sensors within a few cm of a laser-driven target. The sensors must be protected from an extremely harsh environment, including debris, electromagnetic pulses, and unconverted laser light. Another key challenge is reducing the x-ray background relative to the faint diffraction signal. Building on the success of our predecessor (Target Diffraction In Situ), we implemented a staged approach to platform development. First, we built a demonstration diagnostic (Gated Diffraction Development Diagnostic) with two hCMOS sensors to confirm we could adequately protect them from the harsh environment and also acquire acceptable diffraction data. This allowed the team to quickly assess the risks and address the most significant challenges. Here, we also collected scientifically useful data during development. Leveraging what we learned, we recently developed a much more ambitious instrument (Flexible Imaging Diffraction Diagnostic for Laser Experiments) that can field up to eight hCMOS sensors in a flexible geometry and participate in back-to-back shots at the National Ignition Facility (NIF). The design also allows for future iterations, such as faster hCMOS sensors and an embedded x-ray streak camera. The enhanced capabilities of the new instrument required a much more complex design, and the unexpected issues encountered on the first few shots at NIF remind us that complexity has consequences. Our progress in addressing these challenges is described herein, as is our current focus on improving data quality by reducing x-ray background and quantifying the uncertainties of our diffraction measurements.

36 MATERIALS SCIENCE↗
DOE OSTI2025

Phase-based velocity extraction method for photonic Doppler velocimetry with potential higher time resolution

We present an extension of the [Takeda et al., J. Opt. Soc. Am. 72, 156 (1982)] phase extraction method to heterodyne photonic Doppler velocimetry applications. The method yields results equivalent to those obtained by the short-time Fourier transform (STFT), while offering potential improvements in time resolution. Unlike STFT, which relies on window functions, such as the Hamming window, that emphasize central data points and diminish the influence of edges, the extended Takeda method utilizes all data uniformly. This uniform treatment allows for the derivation of empirical equations that directly relate velocity error to the actual time resolution rather than to the local analysis duration. The established equation provides a useful metric for both optimizing hardware configuration and guiding data analysis. Simulation and experimental results confirm that, for a given dataset, specifying a target time resolution yields consistent velocity errors for both methods. These findings underscore the Takeda method’s advantages, particularly its potential higher time resolution and reduced computational burden, making it a valuable tool for high-throughput applications such as laser dynamic compression experiments.

Computer simulation↗
DOE OSTI2025

Implementation of compound refractive lenses for large field-of-view x-ray phase-contrast imaging during hypervelocity impact experiments

Synchrotron x-ray phase-contrast imaging (XPCI) offers time-resolved visualization of dynamic compression phenomena, but its intrinsically small field-of-view (FOV) limits the time that key features remain in frame. A novel approach to enlarge the FOV is achieved by positioning a two-dimensional parabolic compound refractive lens (CRL) upstream of the sample to deliberately defocus the white beam. Ray-tracing simulations and XPCI measurements show that this CRL configuration can expand the beam by ∼50% vertically and ∼15% horizontally based on the full width at half-maximum of the beam. Implementing the CRL, however, attenuates the photon flux and lowers signal-to-noise ratio (SNR). Task-based analysis using a calibration grid (30 μm dots) showed that both setups fail to consistently meet the Rose criterion (SNR ≥ 5) for features of this size in single-bunch imaging. Extrapolating the measured SNR Rose values suggests that the minimum consistently detectable feature lies closer to 30–40 μm for the standard XPCI setup and above 40 μm for CRL-XPCI. Despite this limitation, the CRL configuration nearly doubles the illuminated area, enabling simultaneous tracking of front and rear observations of boron carbide targets subjected to rod and sphere impacts at 1.0–2.6 km/s. Image tracking algorithms and photonic Doppler velocimetry were used to measure penetration and rear-surface velocity histories. Together, these measurements capture crack fronts, penetration, and material breakout, offering new benchmark data for validating high-strain-rate constitutive models of ceramic materials.

Ceramic materials↗
DOE OSTI2025

X-ray diffraction strains in laser-ablated aluminum, nickel, sodium, and invar: Pressures to 475 GPa

Dynamically compressed materials in longitudinal waves are described by two physical models: hydrostatic pressure, with equal, normal, principal stresses or material uniaxially strained in the wave propagation direction. These models are disparate, so experimental comparisons and evaluations are important. Polycrystalline material in a state of hydrostatic pressure will have no eccentricity of x-ray-diffracted Debye–Scherrer rings. A general three-dimensional solution of Bragg diffracted x rays based on principal crystallographic strains in the compression wave was found. The distortion of x-ray diffraction beams has been used for strain measurements; the analysis developed incorporates a strained reciprocal lattice and the incident x-ray beam. Strain-distorted Polanyi surfaces form an annulus of compression with an ellipsoid of revolution in reciprocal space, which is intersected by Ewald's sphere for Bragg diffraction. The in situ measurements for strain describe nanosecond diffraction evaluated using two planes, (hkl) and (h′k′l′), both in the same crystallographic phase. Diffraction from Al, Ni, Na, and Invar quantifies the compression axial strains in these materials: The compression axial ratios are 0.65, 1.05, 0.88, and 1.58 at pressures of 291, 402, 409, and 367 GPa for the respective materials. Crystal structure transformations with homogeneous pressurized stresses, mandating equal normal strains, should not be anticipated to agree with heterogeneous, uniaxially strained, and sheared crystalline phases. Measurements support in-plane strains increasing with pressure, p , in fcc and hcp aluminum as ε 22 = −0.21 p /TPa.

Alloys↗
DOE OSTI2026

Enabling accurate chemical modeling of shocked energetic materials using a machine learning interatomic potential

Understanding the complex chemistry of organic materials under dynamic compression is important for many applications, but it is challenging due to the large number of reactions occurring at various time scales. Here, in this study, we develop a machine learning potential based on Chebyshev polynomials to study the insensitive energetic material 1,3,5-triamino-2,4,6-trinitrobenzene (TATB) under detonation. We discuss a strategy for constructing diverse training data needed to capture the complex chemistry of TATB. Our potential demonstrates strong transferability across a wide range of thermodynamic conditions and other explosives, enabling accurate and reliable chemical modeling of organic materials under extreme conditions. The efficiency of our approach allows for simulations over several nanoseconds and for large system sizes, providing detailed insights into the chemistry of shocked TATB. The model accurately reproduces experimental Hugoniot equation of state data, and our simulations reveal the rapid formation of nitrogen-rich carbon clusters following shock. The methods and datasets developed here offer a robust framework for accurate chemical modeling of other shocked organic energetic materials.

Chemistry↗
DOE OSTI2026

Experimental shock Hugoniot and melt boundary temperatures of aluminum

Aluminum is a ubiquitous component in dynamic compression, pulsed power, and other high energy density physics studies. Its high-pressure behavior and phase diagram are extensively studied standards in shockwave physics. While theoretical calculations and multiphase equations of state have been benchmarked to velocity measurements of loading and unloading waves, pressure and density under shock, and other mechanical data, experimental temperature data under these conditions have not been reported. We conducted a series of experiments shocking and releasing aluminum 6061 and 1100 samples into lithium fluoride windows. We measured temperature at the sample–window interface under steady compression and subsequent isentropic release. These results allow us to constrain the temperature of the solid Hugoniot and the boundary between the liquid and face-centered cubic solid phases.

Hartsfield, Thomas Murray [Sandia National Laborat↗
DOE OSTI2024

DENNIS: a design and analysis tool for dynamic material x-ray diffraction experiments

We present DENNIS (Diffraction Experiment desigN and aNalysiS): a graphical software tool useful for the design and analysis of dynamic x-ray diffraction experiments, such as those performed on the Z Pulsed Power Facility, Thor Pulsed Power Generator, and Dynamic Compression Sector (DCS) of the Advanced Photon Source. DENNIS provides rapid powder and single-crystal diffraction pattern predictions and powder diffraction pattern image integration in three-dimensional geometries. Additional features include crystallographic information file reading, image processing, and synthetic diffraction pattern image generation. We overview the software's capabilities, detail the prediction and integration methodologies, and provide example implementations on Z and DCS experiments.

47 OTHER INSTRUMENTATION↗
DOE OSTI2026

Jupiter Laser Facility Annual Report, FY 2025

Dear JLF community, I cannot believe I am now entering my third year as JLF director — time definitely flies when you are having fun! FY25 was another pivotal year for the Jupiter Laser Facility, marked by both scientific achievement and growing visibility for our community. Building on the successful reopening and refurbishment of the facility, we continued to demonstrate how JLF drives innovation in high energy density and fusion energy science, laser technology, and workforce development. Across Janus, Titan, and COMET, users executed a diverse portfolio of experiments, from dynamic compression and opacity measurements to laser plasma interactions, laboratory astrophysics, and advanced diagnostics. These efforts are highlighted in this report, including the development of new probes that capture the time evolution of plasmas on a single shot, and diagnostics and platforms that are already impacting experiments at NIF and other large facilities. JLF continues to serve as both a testbed for new ideas and a bridge to larger scale campaigns. FY25 also showcased the broader role of JLF within the Laboratory and the national HED science ecosystem. The NIF JLF User Groups Meeting in February brought nearly 180 participants to Livermore and highlighted the scientific progress made during JLF’s first full year of renewed operations. JLF research and users were recognized with Director’s Institutional Awards and Early and Mid Career awards, underscoring the quality and impact of the work performed here. Our team also contributed prominently to national conversations about laser safety, plasma physics, and inertial fusion energy through invited talks, conferences, and professional society leadership. JLF’s integration with LaserNetUS deepened this year as well. We launched a new technical exchange program across LaserNetUS facilities and kicked it off with a JLF team visit to the BELLA Center at Lawrence Berkeley National Laboratory. These exchanges are strengthening operations, sharing best practices, and improving the user experience across the network. Filming for the LaserNetUS “Behind the Scenes” series and participation in the annual LaserNetUS meeting further increased the visibility of our facility and our users. At the same time, JLF continues to play a central role in ambitious new programs, such as the Big Aperture Thulium laser effort funded through one of the DOE Office of Science Microelectronics Science Research Centers, which will use JLF infrastructure to explore next generation high rep rate lasers for EUV and x-ray source development. A core part of our mission remains training the next generation of scientists. In FY25, we welcomed another cohort of summer students, who joined experimental teams on Titan and presented their research at LLNL’s student poster symposium and national inertial fusion energy meetings. JLF users and early career scientists showcased their work at conferences across the country, highlighting experiments performed at the facility. These hands on experiences, and the mentoring provided by our staff and user teams, are central to JLF’s identity as a true user facility. Finally, FY25 reinforced JLF’s role as a focal point for partnerships and outreach. We hosted visits from international collaborators, science leaders, and we shared the story of the facility through venues such as the Big Ideas Lab podcast. These interactions help connect our work to a broader scientific and policy audience and open new pathways for collaboration. As we look ahead, the combination of refurbished hardware, new capabilities like STILETTO and enhanced short pulse performance on Titan, strong partnerships across LLNL and LaserNetUS, and a growing user community positions JLF for an even more ambitious program in the coming years. I am deeply grateful to our technical and operations staff for their dedication, to our LLNL partners for their continued support, and to our users for bringing bold, creative ideas to the facility. I look forward to more experiments, capabilities, partnerships, and groundbreaking science in the years to come! With brightest regards, Félicie Albert, JLF Director.

70 PLASMA PHYSICS AND FUSION TECHNOLOGY↗
DOE OSTI2025

Adaptive variational quantum dynamics simulations with compressed circuits and fewer measurements

The adaptive variational quantum dynamics simulation (AVQDS) method performs real-time evolution of quantum states using automatically generated parametrized quantum circuits that often contain substantially fewer gates than Trotter circuits. Here we report an improved version of the method, which we call AVQDS(T), by porting the tiling efficient trial circuits with rotations implemented simultaneously technique. The algorithm adaptively adds layers of disjoint unitary gates to the ansatz circuit so as to keep the McLachlan distance, a measure of the accuracy of the variational dynamics, below a fixed threshold. Here we perform benchmark noiseless AVQDS(T) simulations of quench dynamics in local spin models and compare with an alternative adaptive variational approach on quantum resource requirement. Quantum dynamical simulations implementing realistic noise channels are also reported. Finally, we propose a way to substantially alleviate the measurement overhead of AVQDS(T) while maintaining high accuracy by synergistically integrating quantum circuit calculations on quantum processing units with classical calculations using, e.g., tensor networks to evaluate the quantum geometric tensor. We showcase that this approach enables AVQDS(T) to deliver more accurate results than simulations using a fixed ansatz of comparable final depth for a significant time duration with fewer quantum resources.

75 CONDENSED MATTER PHYSICS, SUPERCONDUCTIVITY AND↗
DOE OSTI2025

Numerical simulation of compressible fluid-dynamics in the chamber of inertial fusion energy systems

Here, this paper aims to establish new and innovative modeling capabilities for analyzing chambers in Inertial Fusion Energy (IFE) systems. IFE is emerging as a promising method to achieve fusion power production, but several challenges must be overcome to develop an IFE pilot plant or deploy commercial IFE systems. These challenges are both theoretical and technical, encompassing a deeper understanding of the underlying physical phenomena and the development of new technologies and materials. One of the needs is to develop mathematical models to describe IFE systems and numerical tools to simulate them. This paper contributes to this endeavor by presenting a new OpenFOAM solver for IFE systems, focusing on gas dynamics in their chambers. The analysis and development of chamber designs will play a significant role in the transition from single-shot experiments to high-repetition rates, as there is a need to protect the chamber walls from the intense radiation fields produced by fusion reactions. A promising design option, normally referred to as thick wall chamber design, consists in using lithium or molten salt jet arrays within the chamber. A critical phenomenon is the venting of high-pressure gases from the center to the external part of the chamber, passing through the blanket array. This process involves the propagation and attenuation of strong pressure waves, requiring suitable modeling approaches for compressible fluid-dynamics. The solver proposed in this work implements a multi-material hydrodynamics model tailored to accurately describe the non-linear propagation of pressure waves while avoiding numerical oscillation issues typical of high-velocity compressible simulation. This solver is verified against numerical test cases, validated against experimental data, and applied to the analysis of the High-Yield Lithium-Injection Fusion-Energy (HYLIFE-I) concept. The relevance of this paper is threefold. Firstly, it contributes to developing and testing modeling approaches for compressible fluid-dynamics phenomena, with specific focus on the new and unexplored topic of IFE thick-liquid-wall blanket modeling. Secondly, it marks one of the first applications of the OpenFOAM library in the research field of IFE systems. Finally, the investigated problem is of practical interest for IFE developers, as it provides useful indications about relevant phenomena in pressure wave propagation in the chamber of these systems.

70 PLASMA PHYSICS AND FUSION TECHNOLOGY↗
DOE OSTI2024

A Method for Dynamic Kolsky Bar Compression at High Temperatures: Application to Ti-6Al-4V

Abstract An experimental apparatus for measuring the dynamic behavior of materials subjected to strain rates on the order of 10 $$^3$$ 3 s $$^{-1}$$ - 1 and temperatures up to 800°C with a unique triple actuation system is developed in this work. This system is based on the traditional Kolsky (or split-Hopkinson pressure) bar design, with the addition of an external furnace used to heat the specimen to the desired temperature. A synchronized triple pneumatic actuation system is used to control the motion and timing of the the sample, incident, and transmitted bars. The cold contact time (CCT), or the time during which the heated sample is in contact with the room temperature bars before compression, is measured experimentally and carefully controlled to minimize the development of a temperature gradient across the sample and avoid heating of the bars. Experiments are performed in conjunction with ultra high speed imaging and 2D digital image correlation (DIC), as well as high speed thermal imaging. To verify the viability of the proposed system, experiments were conducted on Ti-6Al-4V (wt.%) at temperatures from 25°C up to and 800°C at an average strain rate of approximately 1200 s $$^{-1}$$ - 1 .

36 MATERIALS SCIENCE↗
DOE OSTI2025

SymbolNet: neural symbolic regression with adaptive dynamic pruning for compression

Abstract Compact symbolic expressions have been shown to be more efficient than neural network (NN) models in terms of resource consumption and inference speed when implemented on custom hardware such as field-programmable gate arrays (FPGAs), while maintaining comparable accuracy (Tsoi et al 2024 EPJ Web Conf. 295 09036). These capabilities are highly valuable in environments with stringent computational resource constraints, such as high-energy physics experiments at the CERN Large Hadron Collider. However, finding compact expressions for high-dimensional datasets remains challenging due to the inherent limitations of genetic programming (GP), the search algorithm of most symbolic regression (SR) methods. Contrary to GP, the NN approach to SR offers scalability to high-dimensional inputs and leverages gradient methods for faster equation searching. Common ways of constraining expression complexity often involve multistage pruning with fine-tuning, which can result in significant performance loss. In this work, we propose S y m b o l N e t , a NN approach to SR specifically designed as a model compression technique, aimed at enabling low-latency inference for high-dimensional inputs on custom hardware such as FPGAs. This framework allows dynamic pruning of model weights, input features, and mathematical operators in a single training process, where both training loss and expression complexity are optimized simultaneously. We introduce a sparsity regularization term for each pruning type, which can adaptively adjust its strength, leading to convergence at a target sparsity ratio. Unlike most existing SR methods that struggle with datasets containing more than O ( 10 ) inputs, we demonstrate the effectiveness of our model on the LHC jet tagging task (16 inputs), MNIST (784 inputs), and SVHN (3072 inputs).

Tsoi, Ho Fung (ORCID:0000000225502184)↗