Engineering Papers⌕ Search

SEARCH · Engineering Papers

Results for “weight loss”

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 127 records · Page 7
All sourcesNASA NTRSDOE OSTI
DOE OSTI2023

Application of a deep learning semantic segmentation model to helium bubbles and voids in nuclear materials

Imaging nanoscale radiation-induced defects using the transmission electron microscope (TEM) is a key factor in the successful implementation of materials for nuclear energy structural applications. Analyzing each defect in a TEM micrograph is currently a manual task. To identify the defects in a single image can take anywhere from 15 min to an hour and a project can require the analysis of anywhere from tens to ≥ 100 images. Here, we use artificial intelligence (AI) models to automate this task. For simplification, we evaluated images with only a single type of defect; helium bubbles. Additionally, we performed semantic segmentation of these helium bubble defects in electron microscopy images of irradiated FeCrAl alloys using a deep learning DefectSegNet model. This model, which was previously used to classify crystal defects, is inspired by the classic DenseNet and U-Net image segmentation models. It claims high spatial resolution, but has poor performance at object boundaries. Our paper improves the DefectSegNet model’s application by adding two new features. First, the DefectSegNet model is applied not only to perform calculation pixel-wise but also object (or feature) wise. Because object-wise metrics are directly relevant to our final goal of detecting bubbles, whereas pixel-wise classification is only an intermediate step, it is an important part of our study. Second, a distance map loss (DML) function has been added to increase its performance at object boundaries. It is crucial to accurately represent defects boundaries, especially bubbles, in order to track the bubble-induced swelling caused by irradiation. The boundary-focused DML function is also compared to other loss functions like Cross-entropy, Weighted Binary Cross Entropy (WBCE), Dice and Intersection over Union (IOU). Finally, by incorporating new features, we found a marked improvement on segmentation quality and better shape preservation at the boundaries and areas of the bubbles.

42 ENGINEERING↗
DOE OSTI2021

Risk-Averse Bi-Level Stochastic Network Interdiction Model for Cyber-Security Risk Management

Security of cyber networks is crucial; recent severe cyber-attacks have had a devastating effect on many large organizations. The attack graph, which maps the potential attack paths of a cyber network, is a popular tool for analyzing cyber system vulnerability. In this study, we propose a bi-level stochastic network interdiction model on an attack graph to enable a risk-averse, resource constrained cyber network defender to optimally deploy security countermeasures that protect against attackers with an uncertain budget. This risk- averse conditional-value-at-risk (CVaR) model minimizes a weighted sum of the expected maximum loss over all scenarios and the expected maximum loss from the most damaging attack scenarios. We develop a customized constraint and column generation algorithm to solve our model as well as several acceleration techniques to improve the computational efficiency. Numerical experiments demonstrate that the acceleration techniques enable the solution of relatively large problems within a reasonable amount of time: applying all the acceleration techniques also reduces the average computation time of the basic algorithm by 71% for 100-node graphs. Using metrics called mean-risk value of stochastic solution and value of risk-aversion, computational results suggest that our stochastic risk-averse model significantly outperforms deterministic and risk-neutral models when 1) the distribution of attacker budget is heavy-right-tailed and 2) the defender is highly risk-averse.

Bhuiyan, Tanveer H.↗
DOE OSTI2023

On the influence of over-parameterization in manifold based surrogates and deep neural operators

Constructing accurate and generalizable approximators (surrogate models) for complex physico-chemical processes exhibiting highly non-smooth dynamics is challenging. The main question is what type of surrogate models we should construct and should these models be under-parameterized or over-parameterized. In this work, we propose new developments and perform comparisons for two promising approaches: manifold-based polynomial chaos expansion (m-PCE) and the deep neural operator (DeepONet), and we examine the effect of over-parameterization on generalization. While m-PCE enables the construction of a mapping by first identifying low-dimensional embeddings of the input functions, parameters, and quantities of interest (QoIs), a neural operator learns the nonlinear mapping via the use of deep neural networks. Here, we demonstrate the performance of these methods in terms of generalization accuracy by solving the 2D time-dependent Brusselator reaction-diffusion system with uncertainty sources, modeling an autocatalytic chemical reaction between two species. We first propose an extension of the m-PCE by constructing a mapping between latent spaces formed by two separate embeddings of the input functions and the output QoIs. To further enhance the accuracy of the DeepONet, we introduce weight self-adaptivity in the loss function. We demonstrate that the performance of m-PCE and DeepONet is comparable for cases of relatively smooth input-output mappings. However, when highly non-smooth dynamics is considered, DeepONet shows higher approximation accuracy. We also find that for m-PCE, modest over-parameterization leads to better generalization, both within and outside of distribution, whereas aggressive over-parameterization leads to over-fitting. In contrast, an even highly over-parameterized DeepONet leads to better generalization for both smooth and non-smooth dynamics. Furthermore, we compare the performance of the above models with another recently proposed operator learning model, the Fourier Neural Operator, and show that its over-parameterization also leads to better generalization. Taken together, our studies show that m-PCE can provide very good accuracy at very low training cost, whereas a highly over-parameterized DeepONet can provide better accuracy and robustness to noise but at higher training cost. In both methods, the inference cost is negligible.

97 MATHEMATICS AND COMPUTING↗
DOE OSTI2025

SIDDA: SInkhorn Dynamic Domain Adaptation for Image Classification

SInkhorn Dynamic Domain Adaptation (SIDDA) supplements the experiments presented in 2501.14048, SIDDA: SInkhorn Dynamic Domain Adaptation for Image Classification with Equivariant Neural Networks. SIDDA introduces a semi-supervised, automatic domain adaptation method that leverages Sinkhorn divergences to dynamically adjust the regularization in the optimal transport plan and the weighting between classification and domain adaptation loss terms during training.

Pandya, Sneh [Fermi National Accelerator Laborator↗
DOE OSTI2026

Dynamic Disruption Resilience in Intermodal Transport Networks: Integrating Flow Weighting and Centrality Measures

Resilient intermodal freight networks are vital for sustaining supply chains amid increasing threats from natural hazards and cyberattacks. Transportation resilience has been widely studied; understanding how random and targeted disruptions affect structural connectivity and functional performance remains a key challenge. To address this, this study evaluates the robustness of the US intermodal freight network, which consists of rail and water modes, using a simulation-based framework that integrates graph-theoretic metrics with flow-weighted centrality measures. Disruption scenarios are examined, including random failures as well as targeted node and edge removals based on static and dynamically updated degree and betweenness centrality. To reflect more realistic conditions, flow-weighted degree centralities (WDC) and partial node degradation are considered. Two resilience indicators are used: (1) the size of the giant connected component to measure structural connectivity; and (2) flow-weighted network efficiency (NE) to assess freight mobility under disruption. The results show that progressively degrading nodes ranked by WDC to 60% of their original functionality causes a sharper decline in normalized NE, for up to approximately 45 affected nodes, than complete failure (100% loss of functionality) applied to nodes targeted by weighted betweenness centrality or selected at random. This highlights how partial degradation of high-tonnage hubs can produce disproportionately large functional losses. The findings emphasize the need for resilience strategies that go beyond network topology to incorporate freight flow dynamics.

42 ENGINEERING↗
DOE OSTI2022

Observations of gas-phase products from the nitrate-radical-initiated oxidation of four monoterpenes

Abstract. Chemical ionization mass spectrometry with the nitrate reagent ion (NO3- CIMS) was used to investigate the products of the nitrate radical (NO3) initiated oxidation of four monoterpenes in laboratory chamber experiments. α-Pinene, β-pinene, Δ-3-carene, and α-thujene were studied. The major gas-phase species produced in each system were distinctly different, showing the effect of monoterpene structure on the oxidation mechanism and further elucidating the contributions of these species to particle formation and growth. By comparing groupings of products based on the ratios of elements in the general formula CwHxNyOz, the relative importance of specific mechanistic pathways (fragmentation, termination, and radical rearrangement) can be assessed for each system. Additionally, the measured time series of the highly oxidized reaction products provide insights into the ratio of relative production and loss rates of the high-molecular-weight products of the Δ-3-carene system. The measured effective O:C ratios of reaction products were anticorrelated with new particle formation intensity and number concentration for each system; however, the monomer : dimer ratios of products had a small positive trend. Gas-phase yields of oxidation products measured by NO3- CIMS correlated with particle number concentrations for each monoterpene system, with the exception of α-thujene, which produced a considerable amount of low-volatility products but no particles. Species-resolved wall loss was measured with NO3- CIMS and found to be highly variable among oxidized reaction products in our stainless steel chamber.

37 INORGANIC, ORGANIC, PHYSICAL, AND ANALYTICAL CH↗
DOE OSTI2023

Highly durable platinum group metal-free catalyst fiber cathode MEAs for proton exchange membrane fuel cells

For this work, Fe-based platinum group metal (PGM)-free catalysts were incorporated into electrospun fiber mat or powder cathode membrane-electrode-assemblies (MEAs) with a Nafion 211 membrane and a Pt/C powder anode. Fabrication and characterization tests were performed on MEAs with: (1) a conventional powder cathode with a neat Nafion binder, (2) a fiber mat cathode with a Nafion/polyethylene oxide (PEO) binder, where PEO was extracted before MEA testing, (3) a powder cathode with a blended binder of Nafion and polyvinylidene fluoride (PVDF), and (4) a series of fiber mat cathodes with different Nafion/PVDF binder weight ratios. Cathode degradation occurred, with a loss in power output, in MEAs with a neat Nafion powder cathode or with a Nafion fiber cathode. In contrast, little or no power loss was observed for powder or fiber cathodes when the binder was a blend of Nafion and PVDF. The presence of hydrophobic PVDF drove water and electrogenerated peroxide out of the cathode, away from catalyst particles, which improved cathode durability, but PVDF also decreased the binder conductivity and slowed oxygen reduction kinetics, resulting in lower power densities. A 75:25 w:w Nafion:PVDF fiber cathode MEA was the best compromise for maximizing power and minimizing catalyst degradation. For such a cathode, with a PGM-free cathode catalyst loading of 3.0 mg cm –2 , a power density of 88 mW cm –2 at 0.5 V, 80 °C, and 200 kPa abs pressure was maintained for 80 h of continuous operation.

25 ENERGY STORAGE↗
DOE OSTI2023

Lithium-Ion Battery Diagnostics Using Electrochemical Impedance via Machine-Learning

Diagnosing battery states such as health, state-of-charge, or temperature is crucial for ensuring the safety and reliability of electrochemical energy storage systems. While some states, such as temperature, may be measured using cheap sensors, accurate diagnosis of battery health metrics usually requires time-consuming performance measurements, making them infeasible for use in real-world operation. These health metrics can be measured during lab-testing and then estimated on-line using predictive life models or via state observer algorithms such as Kalman filters, but these predictive methods should be supplemented by actual measurement of battery health whenever possible to ensure reliability. Rapid measurement of battery health may be done by various types of fast diagnostic techniques such as electrochemical impedance spectroscopy (EIS), which can be performed in only a few minutes and require only a fraction of the energy and power needed for a full charge and discharge measurement. But there is a substantial challenge for estimating battery health using EIS data, as EIS is sensitive to cell temperature, state-of-charge, current, and resting time in addition to health. Thus, utilizing EIS data to predict battery capacity requires correcting for all these additional variables, a task that is extremely difficult to handle analytically. This talk utilizes machine-learning methods to estimate the effectiveness of battery capacity prediction from EIS data, leveraging a data set of hundreds of EIS measurements recorded at varying temperature and state-of-charge throughout a 500-day aging study of 32 commercial, large-format NMC-Graphite lithium-ion batteries. Using EIS as input to machine-learning models is complicated by the nonlinear response of impedance to battery health, temperature, and state-of-charge, as well as the collinearity between the impedance response at neighboring frequencies, which can easily lead to overfit models. To train robust models, features from EIS data need to be extracted from the data or some subset of critical frequencies selected. Many approaches for extracting and selecting features from EIS data from electrochemical analysis and machine-learning fields were identified for analysis: using the entire raw spectra; selection of one, two, or many frequencies from the entire spectra; selecting interesting points from the EIS measurement using domain knowledge; fitting EIS with an equivalent-circuit model; calculating statistics on the raw impedance values; and reducing the dimensionality of the data using unsupervised linear (principal component analysis) and non-linear (uniform manifold approximation and projection) methods. These approaches were rigorously compared using a machine-learning pipeline approach, training linear, Gaussian process, and random forest regression models and quantifying performance using cross-validation as well as a held-out test set. An artificial neural network model trained on the raw spectra was also tested. Promising pipelines were fine-tuned via Bayesian hyperparameter optimization using cross-validation loss and training with class-specific weights to counter data set imbalance. The most reliable method for utilizing impedance in this work was the selection of two optimal frequencies through an exhaustive search, resulting in about 2% mean absolute error on test data for both Gaussian process and random forest model architectures. Interrogation of a variety of models reveals critical frequencies of 100 Hz and 103 Hz for this data set, though the optimal set of frequencies is not necessarily intuitive, i.e., the best performing models are not simply those that use impedance at frequencies that have the highest correlation to the relative discharge capacity. The best performing model is an ensemble model, which is able to predict battery capacity with 1.9% mean absolute error for unseen cells using impedance recorded at a variety of temperatures and states-of-charge.

battery↗
DOE OSTI2022

Silicon nitride stress-optic microresonator modulator for optical control applications

Modulation-based control and locking of lasers, filters and other photonic components is a ubiquitous function across many applications that span the visible to infrared (IR), including atomic, molecular and optical (AMO), quantum sciences, fiber communications, metrology, and microwave photonics. Today, modulators used to realize these control functions consist of high-power bulk-optic components for tuning, sideband modulation, and phase and frequency shifting, while providing low optical insertion loss and operation from DC to 10s of MHz. In order to reduce the size, weight and cost of these applications and improve their scalability and reliability, modulation control functions need to be implemented in a low loss, wafer-scale CMOS-compatible photonic integration platform. The silicon nitride integration platform has been successful at realizing extremely low waveguide losses across the visible to infrared and components including high performance lasers, filters, resonators, stabilization cavities, and optical frequency combs. Yet, progress towards implementing low loss, low power modulators in the silicon nitride platform, while maintaining wafer-scale process compatibility has been limited. Here we report a significant advance in integration of a piezo-electric (PZT, lead zirconate titanate) actuated micro-ring modulation in a fully-planar, wafer-scale silicon nitride platform, that maintains low optical loss (0.03 dB/cm in a 625 µm resonator) at 1550 nm, with an order of magnitude increase in bandwidth (DC - 15 MHz 3-dB and DC - 25 MHz 6-dB) and order of magnitude lower power consumption of 20 nW improvement over prior PZT modulators. The modulator provides a >14 dB extinction ratio (ER) and 7.1 million quality-factor (Q) over the entire 4 GHz tuning range, a tuning efficiency of 162 MHz/V, and delivers the linearity required for control applications with 65.1 dB·Hz 2/3 and 73.8 dB·Hz 2/3 third-order intermodulation distortion (IMD3) spurious free dynamic range (SFDR) at 1 MHz and 10 MHz respectively. We demonstrate two control applications, laser stabilization in a Pound-Drever Hall (PDH) lock loop, reducing laser frequency noise by 40 dB, and as a laser carrier tracking filter. This PZT modulator design can be extended to the visible in the ultra-low loss silicon nitride platform with minor waveguide design changes. This integration of PZT modulation in the ultra-low loss silicon nitride waveguide platform enables modulator control functions in a wide range of visible to IR applications such as atomic and molecular transition locking for cooling, trapping and probing, controllable optical frequency combs, low-power external cavity tunable lasers, quantum computers, sensors and communications, atomic clocks, and tunable ultra-low linewidth lasers and ultra-low phase noise microwave synthesizers.

Wang, Jiawei (ORCID:0000000257965220)↗
DOE OSTI2024

Graph network heterogeneity predicts interplant wake losses

Wind plants generate large-scale wakes, which can affect the performance of neighboring installations. Such wakes are challenging to estimate due to the inherent complexity in modeling wake interactions between large quantities of turbines at various distances. Weighted directed graph networks can inform complex models by linking turbine pairs into chains of upstream and downstream neighbors for a given wind direction. A novel interpretation of the graph network adjacency matrix is proposed where each element of the matrix represents the cumulative impact of upstream turbines on an individual. In this study, wake losses were estimated with an engineering wake model across a range of inflow conditions for nine parametric variations of a system containing two neighboring wind plants. The parametric nature of the study isolates turbine spacing within the plant, separation distance between plants, and wind direction as the main drivers of wake losses. Spatial heterogeneity is computed from the weighted average adjacency matrix of each plant arrangement. The proposed method is orders of magnitude faster than wake modeling and does not require detailed turbine information or atmospheric conditions. Furthermore, the weighted average adjacency matrix provides insight on the spatial organization of wake losses at various scales. Plant heterogeneity is correlated with wake losses within and among plants. Framing wind plant wake interaction in terms of graph network spatial heterogeneity provides an efficient approach for predicting wake losses within and among neighboring wind plants with applications to other complex systems where wake interactions are key factor.

17 WIND ENERGY↗
DOE OSTI2026

4D Multimodal Co-attention Fusion Network with Latent Contrastive Alignment for Alzheimer’s Diagnosis

Multimodal neuroimaging provides complementary structural and functional insights into both human brain organization and disease-related dynamics. Recent studies demonstrate enhanced diagnostic sensitivity for Alzheimer’s disease (AD) through synergistic integration of neuroimaging data (e.g., sMRI, fMRI) with tabular data (e.g., behavioral and cognitive tests). However, the intrinsic heterogeneity across modalities (e.g., 4D spatiotemporal fMRI dynamics vs. 3D anatomical sMRI structure) presents critical challenges for discriminative feature fusion, often leading to information loss or biased fusion. To bridge this gap, we propose M2M-AlignNet: a multimodal co-attention network with latent alignment for early AD diagnosis using sMRI and fMRI. At the core of our approach is a multi-patch-to-multi-patch (M2M) contrastive loss function that quantifies and reduces representational discrepancies via weighted patch correspondence, explicitly aligning fMRI components across brain regions with their sMRI structural substrates without one-to-one constraints. Additionally, we propose a latent-as-query co-attention module to autonomously discover fusion patterns, circumventing modality prioritization biases while minimizing feature redundancy. We conduct extensive experiments to confirm the effectiveness of our method and highlight the correspondence between fMRI and sMRI as AD biomarkers.

Wei, Yuxiang [Georgia Institute of Technology]↗
DOE OSTI2026

Gearbox bearing crack growth prognostics and uncertainty quantification with physics-informed machine learning

This paper introduces the extreme theory of functional connections (X-TFC), a physics-informed machine learning algorithm, and tailors it to estimate the remaining useful life (RUL) of wind turbine gearbox bearings experiencing fatigue crack growth. Unlike purely data-driven methods, X-TFC embeds a physics model, based on Head's theory in this work, into its training objective. The core of X-TFC is a random-projection single-layer neural network trained via an extreme learning machine, which requires only limited damage progression data and solves for output weights with a least-squares optimization algorithm. A composite loss function balances the network's fit to observed degradation data against the residuals of the governing crack growth differential equation, ensuring the learned damage trajectory remains physically plausible. When applied to a vibration-based health-index (HI) dataset measured during the growth of a crack on the inner ring of a high-speed bearing in a wind turbine gearbox (Bechhoefer and Dubé, 2020), X-TFC achieves near-zero prediction bias. Even when trained on only the first 10 %–20 % of the damage progression data, with sufficient physics weighting its predictions remain monotonic and smooth, delivering high prognosability and trendability. To quantify the epistemic uncertainty, we employ a Monte Carlo ensemble of independently initialized X-TFC models trained on noise-perturbed data, which yields confidence intervals around each RUL estimate and captures both model-parameter and epistemic uncertainty. In addition to a vibration-based HI, we demonstrate that the proposed framework can be directly applied to a supervisory control and data acquisition (SCADA) data-based HI (Eftekhari Milani et al., 2026) measured during similar wind turbine gearbox bearing crack faults, preserving its accuracy and interpretability. This extension shows the versatility of our approach, which is applicable to bearings of multiple gearbox manufacturers, models, and ratings using only SCADA data. By integrating domain knowledge with machine learning, X-TFC offers a rapid, reliable tool for crack prognostics. Its adaptability to other bearing failure modes, such as pitch bearing ring cracks, positions X-TFC as a powerful enabler of data-driven, physics-informed asset management in the wind energy sector and beyond.

17 WIND ENERGY↗
DOE OSTI2024

Using the Metropolis algorithm to explore the loss surface of a recurrent neural network

In the limit of small trial moves the Metropolis Monte Carlo algorithm is equivalent to gradient descent on the energy function in the presence of Gaussian white noise. This observation was originally used to demonstrate a correspondence between Metropolis Monte Carlo moves of model molecules and overdamped Langevin dynamics, but it also applies in the context of training a neural network: making small random changes to the weights of a neural network, accepted with the Metropolis probability, with the loss function playing the role of energy, has the same effect as training by explicit gradient descent in the presence of Gaussian white noise. We explore this correspondence in the context of a simple recurrent neural network. We also explore regimes in which this correspondence breaks down, where the gradient of the loss function becomes very large or small. In these regimes the Metropolis algorithm can still effect training, and so can be used as a probe of the loss function of a neural network in regimes in which gradient descent struggles. We also show that training can be accelerated by making purposely-designed Monte Carlo trial moves of neural-network weights.

Casert, Corneel↗
DOE OSTI2022

Formulating energy density for designing practical lithium–sulfur batteries

We report the lithium–sulfur (Li–S) battery is one of the most promising battery systems due to its high theoretical energy density and low cost. Despite impressive progress in its development, there has been a lack of comprehensive analyses of key performance parameters affecting the energy density of Li–S batteries. Here, we analyse the potential causes of energy loss during battery operations. We identify two key descriptors (R weight and R energy ) that represent the mass- and energy-level compromise of the full-cell energy density, respectively. A formulation for energy density calculations is proposed based on critical parameters, including sulfur mass loading, sulfur mass ratio, electrolyte/sulfur ratio and negative-to-positive electrode material ratio. The current progress of Ah-level Li–S batteries is also summarized and analysed. Finally, future research directions, targets and prospects for designing practical high-performance Li–S batteries are proposed.

36 MATERIALS SCIENCE↗
DOE OSTI2026

PowerModelsGAT-AI: Physics-Informed Graph Attention for Multi-System Power Flow With Continual Learning

Solving the alternating current power flow equations in real time is essential for secure grid operation, yet classical Newton–Raphson solvers can be slow under stressed conditions. Existing graph neural networks for power flow are typically trained on a single system and often degrade on different systems. We present PowerModelsGAT-AI, a physics-informed graph attention network that predicts bus voltages and generator injections. The model uses bus-type-aware masking to handle different bus types and balances multiple loss terms, including a power-mismatch penalty, using learned weights. We evaluate the model on 14 benchmark systems (4 to 6,470 buses) and train a unified model on 13 of these under contingency conditions with up to two branch outages, achieving an average normalized mean absolute error of 0.89% for voltage magnitudes and R 2 >0.99 for voltage angles. We also show continual learning: when adapting a base model to a new 1,354-bus system, standard fine-tuning causes severe forgetting with error increases exceeding 1000% on base systems, while our experience replay and elastic weight consolidation strategy keeps error increases below 2% and in some cases improves base-system performance. Interpretability analysis shows that learned attention weights correlate with physical branch parameters (susceptance: r=0.38 ; thermal limits: r=0.22 ), and feature importance analysis supports that the model captures established power flow relationships.

24 POWER TRANSMISSION AND DISTRIBUTION↗
DOE OSTI2021

Degradation of Platinum-Cobalt Alloy PEMFC Cathode Catalysts in Catalyst-Ionomer Inks

In this paper we report on studies of the effects of catalyst-ionomer ink composition: ionomer equivalent weight (EW), ink solvent, and ink mixing on a commercial PtCo alloy PEMFC cathode catalyst. X-ray absorption spectroscopy (XAS) and wide-angle X-ray scattering (WAXS) are utilized to determine catalyst atomic structure, catalyst crystallite composition, and extent of loss of Co into the ionomer-solvent phase. Three different n-propanol to water weight ratios (7:3, 5:5, and 3:7) and two different ionomers (3 M 800 EW and Nafion 1000 EW) were studied. Cobalt loss from the PtCo particles was found to increase with increasing water content in the inks and to be more extensive for the 800 EW ionomer inks, causing loss of contraction of the PtCo lattice and formation of a Pt shell-PtCo alloy core intraparticle structure.

25 ENERGY STORAGE↗
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)↗