Engineering Papers⌕ Search

SEARCH · Engineering Papers

Results for “Transformer model”

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 253 records · Page 14
All sourcesNASA NTRSDOE OSTI
DOE OSTI2022

Integration of Omics into a New Comprehensive Rate Law for Competitive Terminal Electron-Accepting Processes in Reactive Transport Models: Application to N, Fe, S, and Contaminant Transformations in Stream and Wetland Sediments

Surface waters represent important sources of alternate energy and drinking water in the United States, and characterizing the biogeochemical processes that affect surface water quality is relevant to the DOE-BER mission. Sediment biogeochemical processes regulate the release of carbon (C), nutrients, and contaminants to surface waters and thus influence water quality. Sediment biogeochemical processes are dynamic and affected by the deposition and remobilization of solid material and changes in environmental conditions driven by water discharge variations. Wetlands are important natural filters of surface waters which may either trap, metabolize, or mobilize nutrients and contaminants. Despite their importance, biogeochemical processes regulating nutrient and contaminant release and C transformation in stream and wetland sediments cannot be predicted accurately by current mathematical models. These reactive transport models largely rely on detectable changes in geochemical conditions to activate metabolic processes, do not accurately account for the competition between microbial processes, and poorly constrain effects of hydrological perturbations on biogeochemical processes. In this BER-SBR exploratory project, metagenomic and geochemical signatures were combined to identify microbially-mediated redox processes in anaerobic stream and wetland sediments from the Savannah River Site (SRS, ANL SFA) and East Fork Poplar Creek (EFPC, ORNL SFA) that play important roles in C, uranium (U), and mercury (Hg) transformations. In addition, sediment incubations were conducted to examine the competition between anaerobic respiration processes Finally, new rate laws were developed for reactive transport models that rely on complementary metagenomic and geochemical signatures to identify the underlying anaerobic microbial processes in stream and wetland sediments, describe the competition between the dominant metabolic processes involved in nutrient release and U and Hg mobilization, and more accurately quantify carbon transformation and the response of microbial processes to changes in redox conditions associated with hydrological forcing. These rate laws were optimized in batch reactors with SRS wetland sediments, where iron and sulfate reduction dominate. Anaerobic carbon remineralization processes followed the expected thermodynamic sequence of microbial respiration with depth in both sediments, except that geochemical signals indicated that sulfate reduction was inactive in EFPC sediments and moderate in SRS wetland sediments. Estimates indicated that microbial iron reduction contributed to at least half of the production of reduced iron in these sediments. Incubations demonstrated that nitrate reduction, denitrification, and dissimilatory nitrate reduction to ammonium were active in the natural EFPC sediment and activated upon nitrate amendment in these nitrate-rich sediments. In turn, these processes were outcompeted by the addition of either iron oxides or sulfate as alternative terminal electron acceptors. Although geochemical products of sulfate reduction were not detected in the incubations, the abundance of sulfate reduction genes increased with depth in the sediment and was equally more pronounced in treatments amended with either iron oxides or sulfate. Simultaneously, anaerobic sulfide oxidizing bacteria coupling sulfide oxidation to DNRA (and not conventional denitrification) were apparently enriched over time, regardless of the treatments. These findings indicate that sulfate reduction is important in freshwater stream sediments and probably catalyzed by a cryptic sulfur cycle involving nitrogen species, in which the sulfur products from sulfate reduction are immediately removed by side reactions and not detectable by geochemical measurements alone. Similar experiments in SRS sediments, however, demonstrated little interaction between nitrogen and sulfur cycling microorganisms. Sulfate reduction was impacted by the addition of more thermodynamically favorable electron acceptors, suggesting either that iron-reducing microorganisms outcompeted sulfate-reducing microorganisms for organic substrate, depleting the stock of electron donor available for sulfate reduction, or that the cryptic sulfur cycle was shunted by the precipitation of FeS generated as a result of the abiotic reduction of iron oxides by dissolved sulfide. A diagnostic modeling exercise was conducted to further investigate the competition between terminal electron accepting processes. As conventional kinetic models typically do not account for cryptic cycles and used inaccurate formulations to describe competition between microbial communities, new metabolic rate laws were developed that explicitly express the electron acceptor-specific enzyme of each energetically favorable metabolic process based on gene abundance detected in the incubations. The model was tested with the sediment slurry incubation data to determine whether substrate competition could explain the decrease in sulfate reduction observed in the presence of iron oxide competitor. The model was able to reproduce geochemical concentrations really well in each treatment once the model was calibrated with the unamended control, suggesting that microbial competition was indeed driven by thermodynamic considerations. Overall, carbon remineralization processes and rates will be reproduced much more realistically with the new metabolic rate laws.

54 ENVIRONMENTAL SCIENCES↗
DOE OSTI2021

Enriching Load Data Using Micro-PMUs and Smart Meters

In modern distribution systems, load uncertainty can be fully captured by micro-PMUs, which can record high-resolution data; however, in practice, micro-PMUs are installed at limited locations in distribution networks due to budgetary constraints. In contrast, smart meters are widely deployed but can only measure relatively low-resolution energy consumption, which cannot sufficiently reflect the actual instantaneous load volatility within each sampling interval. In this paper, we have proposed a novel approach for enriching load data for service transformers that only have low-resolution smart meters. The key to our approach is to statistically recover the high-resolution load data, which is masked by the low-resolution data, using trained probabilistic models of service transformers that have both high- and low-resolution data sources, i.e., micro-PMUs and smart meters. The overall framework consists of two steps: first, for the transformers with micro-PMUs, a Gaussian Process is leveraged to capture the relationship between the maximum/minimum load and average load within each low-resolution sampling interval of smart meters; a Markov chain model is employed to characterize the transition probability of known high-resolution load. Next, the trained models are used as teachers for the transformers with only smart meters to decompose known low-resolution load data into targeted high-resolution load data. The enriched data can recover instantaneous load uncertainty and significantly enhance distribution system observability and situational awareness. Here, we have verified the proposed approach using real high- and low-resolution load data.

24 POWER TRANSMISSION AND DISTRIBUTION↗
DOE OSTI2026

Surrogate model evaluation and building energy benchmarking for commercial buildings

Building energy consumption benchmarking involves challenges associated with various energy patterns for different building types; heating, ventilating, and air-conditioning (HVAC) system types; and climates. Given significant variation in energy use patterns, accurate prediction of long-term energy use using surrogate models remains challenging. Multiple linear regression (MLR) is commonly used for building energy benchmarking because of its simple structure; however, it lacks accuracy compared to other black-box models. Although many studies have compared surrogate models and offer guidance on model selection based on metrics, they do not provide detailed analysis on improving the surrogate model accuracy. In this paper, we implement a surrogate model using polynomial ridge regression (i.e., MLR with interaction terms combined with ridge regularization) for small office and retail strip mall buildings across six HVAC system types and all climate zones, for electricity and natural gas in baseline and proposed scenarios. A simulation workflow is developed using OpenStudio TM /EnergyPlus TM to generate simulation data using measures over a wide range of efficiency inputs. Enhancements based on statistical insights are used for improving the model accuracy using filters, input transformations, and change points. Surrogate models achieved average coefficient of variation of the root mean squared error (CVRMSE) values of 2.17, 1.06, 2.05, and 3.26 for proposed electricity, proposed natural gas, baseline electricity, and baseline natural gas, respectively, with enhancements reducing CVRMSE by an average of 14.9% across all combinations. We provide model interpretation via Shapley additive explanations to determine which input variables most influence energy consumption and provide supportive arguments for enhancements.

32 ENERGY CONSERVATION, CONSUMPTION, AND UTILIZATI↗
DOE OSTI2025

HydraGNN v4.0

The new version of HydraGNN v4.0 provides additional core capabilities, such as: Inclusion of multi-body atomistic cluster expansion MACE, polarizable atom interaction neural network PAINN, and equivariant principal neighborhood aggregation (PNAEq) among the message passing layers supported -Inclusion of graph transformers to directly model long-range interactions between nodes that are distant in the graph topology Integration of graph transformers with message passing layers by combining the graph embedding generated by the two mechanisms, which allows for an improved expressivity of the HydraGNN architecture Improved re-implementation of multi-task learning (MTL) to allow its use for stabilized training across imbalanced, multi-source, multi-fidelity data Introduction of multi-task parallelism, a newly proposed type of model parallelism specifically for MTL architectures, which allows to dispatch different output decoding heads to different GPU devices Integration of multi-task parallelism with pre-existing distributed data parallelism to enable a 2D parallelization for distributed training Improved portability of the distributed training across Intel GPUs, which has been testes on ALCF exascale supercomputer Aurora Inclusion of 2-level fine-grained energy profilers portable across NVIDIA, AMD, and Intel GPUs to monitor the power and energy consumption associated with different functions executed by the HydraGNN code during data pre-load and training Restructuring of previous examples and inclusion of new sets of examples to illustrate the download, preprocess, and training of HydraGNN models on new large-scale open-source datasets for atomistic materials modeling (e.g., Alexandria, Transition1x, OMat24, OMol25)

Lupo Pasini, Massimiliano [Oak Ridge National Labo↗
DOE OSTI2026

wa-hls4ml: A Benchmark and Surrogate Models for hls4ml Resource and Latency Estimation

As machine learning (ML) is increasingly implemented in hardware to address real-time challenges in scientific applications, the development of advanced toolchains has significantly reduced the time required to iterate on various designs. These advancements have solved major obstacles, but also exposed new challenges. For example, processes that were not previously considered bottlenecks, such as hardware synthesis, are becoming limiting factors in the rapid iteration of designs. To mitigate these emerging constraints, multiple efforts have been undertaken to develop an ML-based surrogate model that estimates resource usage of ML accelerator architectures. We introduce wa-hls4ml, a benchmark for ML accelerator resource and latency estimation, and its corresponding initial dataset of over 680,000 fully connected and convolutional neural networks, all synthesized using hls4ml and targeting Xilinx FPGAs. The benchmark evaluates the performance of resource and latency predictors against several common ML model architectures, primarily originating from scientific domains, as exemplar models, and the average performance across a subset of the dataset. Additionally, we introduce GNN- and transformer-based surrogate models that predict latency and resources for ML accelerators. We present the architecture and performance of the models and find that the models generally predict latency and resources for the 75% percentile within several percent of the synthesized resources on the synthetic test dataset.

Hawks, Benjamin [Fermilab] (ORCID:0000000157000288↗
DOE OSTI2025

wa-hls4ml: A Benchmark and Surrogate Models for hls4ml Resource and Latency Estimation

As machine learning (ML) is increasingly implemented in hardware to address real-time challenges in scientific applications, the development of advanced toolchains has significantly reduced the time required to iterate on various designs. These advancements have solved major obstacles, but also exposed new challenges. For example, processes that were not previously considered bottlenecks, such as hardware synthesis, are becoming limiting factors in the rapid iteration of designs. To mitigate these emerging constraints, multipleefforts have been undertaken to develop an ML-based surrogate model that estimates resource usage of synthesized ML accelerator architectures. We introduce wa-hls4ml, a benchmark for ML accelerator resource and latency estimation, and its corresponding initial dataset of over 680 000 fully connected and convolutional neural networks, all synthesized using hls4ml and targeting Xilinx FPGAs. The benchmark evaluates the performance of resource and latency predictors against several common ML model architectures, primarily originating from scientific domains, as exemplar models, and the average performance across a subset of the dataset. Additionally, we introduce GNN- and transformer-based surrogate models that predict latency and resources for ML accelerators. We present the architecture and performance of the models and find that the models generally predict latency and resources for the 75% percentile within several percent of the synthesized resources on the synthetic test dataset.

Hawks, Benjamin G. [Fermilab]↗
DOE OSTI2024

Automated Framework for Groundwater Monitoring Using DWT with LSTM and Transformers

Environmental monitoring is critical for safeguarding public health and ecological well-being. Traditional data structuring and workflow monitoring methods consume significant time and effort, hindering timely insights and effective decision-making. Our study addresses this challenge by presenting an AI framework that automates data cleaning, structuring, and modeling processes, specifically targeting applications in groundwater monitoring. By leveraging automation for data processing and model training, our framework establishes a novel and efficient paradigm for environmental monitoring, with its potential application to the vast network of over a hundred Department of Energy Environmental Management (DoE-EM) cleanup sites across the country. It analyzes data streams from a network of groundwater Internet-of-Things (IoT) sensors deployed at the Savannah River Site (SRS) for prediction modeling. This allows human experts to focus on analysis and decision-making, ultimately leading to better environmental outcomes.The framework employs multivariate time-series forecasting methods to study and model the behavior of varying chemical analytes. The continuous learning process is enabled by utilizing deep learning techniques. It allows the framework to become more nuanced in its analysis over time, adapting to the specific characteristics of the environmental site and the evolving nature of contaminant behavior. Deep learning models known for sequence modeling, LSTM, and Transformers are employed for time series forecasting. Data processing and structuring are essential components significantly impacting the final model's performance. This hypothesis was proven by presenting a comparative analysis of model performance with processed and unprocessed data. The feature engineering approach utilized was the Discrete Wavelet Transform, which works well with time series data.

Discrete Wavelet Transform (DWT)↗
DOE OSTI2024

Distribution Transformer Demand: Understanding Demand Segmentation, Drivers, and Management Through 2050

The National Renewable Energy Laboratory (NREL) has been working closely with the U.S. Department of Energy's Office of Electricity (OE) to understand the critical drivers and potential means of managing distribution transformer demand through 2050. This effort has consulted with utility representative organizations and transformer manufacturers to understand the problem, characterized the in-service assets, and modeled future demand. Distribution transformers, or service transformers, range from 10 to 5,000 kilovolt-amperes (kVA), have a high-side voltage of less than 34.5 kilovolts, and have step-down power delivery for customer end use. This research will help the manufacturing sector understand production requirements and better inform utility strategies for managing their demand.

24 POWER TRANSMISSION AND DISTRIBUTION↗
DOE OSTI2022

A Length Adaptive Algorithm-Hardware Co-design of Transformer on FPGA Through Sparse Attention and Dynamic Pipelining

Transformers are considered one of the most important deep learning models since 2018, in part because it establishes state-of-the-art (SOTA) records and could potentially replace existing Deep Neural Networks (DNNs). Despite the remarkable triumphs, the prolonged turnaround time of Transformer models is a widely recognized roadblock. The variety of sequence lengths imposes additional computing overhead where inputs need to be zero-padded to the maximum sentence length in the batch to accommodate the parallel computing platforms. This paper targets the field-programmable gate array (FPGA) and proposes a coherent sequence length adaptive algorithm–hardware co-design for Transformer acceleration. Particularly, we develop a hardware-friendly sparse attention operator and a length-aware hardware resource scheduling algorithm. The proposed sparse attention operator brings the complexity of attention-based models down to linear complexity and alleviates the off-chip memory traffic. The proposed length-aware resource hardware scheduling algorithm dynamically allocates the hardware resources to fill up the pipeline slots and eliminates bubbles for NLP tasks. Experiments show that our design has very small accuracy loss and has 80.2 × and 2.6 × speedup compared to CPU and GPU implementation, and 4 × higher energy efficiency than state-of-the-art GPU accelerator optimized via CUBLAS GEMM.

Peng, Hongwu↗
DOE OSTI2025

Ground-state-based model reduction with unitary circuits

Here, we present a method to numerically obtain low-energy effective models based on a unitary transformation of the ground state. The algorithm finds a unitary circuit that transforms the ground state of the original model to a projected wavefunction with only the low-energy degrees of freedom. The effective model can then be derived using the unitary transformation encoded in the circuit. We test our method on the one-dimensional and two-dimensional square-lattice Hubbard model at half-filling, and obtain more accurate effective spin models than the standard perturbative approach.

Hubbard model↗
DOE OSTI2025

Physics-informed transformation toward improving the machine-learned NLTE models of ICF simulations

The integration of machine-learning techniques into inertial confinement fusion (ICF) simulations has emerged as a powerful approach for enhancing computational efficiency. By replacing the costly nonlocal thermodynamic equilibrium (NLTE) model with machine-learning models, significant reductions in calculation time have been achieved. However, determining how to optimize machine-learning-based NLTE models in order to match ICF simulation dynamics remains challenging, underscoring the need for physically relevant error metrics and strategies to enhance model accuracy with respect to these metrics. Thus, we propose novel physics-informed transformations designed to emphasize energy transport, use these transformations to establish new error metrics, and demonstrate that they yield smaller errors within reduced principal-component spaces compared to conventional transformations. Published by the American Physical Society 2025

70 PLASMA PHYSICS AND FUSION TECHNOLOGY↗
DOE OSTI2025

Rapid wavefield forecasting for earthquake early warning via deep sequence to sequence learning

We propose a deep learning model, WaveCastNet, to forecast high-dimensional wavefields. WaveCastNet integrates a convolutional long expressive memory architecture into a sequence-to-sequence forecasting framework, enabling it to model long-term dependencies and multiscale patterns in both space and time. By sharing weights across spatial and temporal dimensions, WaveCastNet requires significantly fewer parameters than more resource-intensive models such as transformers, resulting in faster inference times. Crucially, WaveCastNet also generalizes better than transformers to rare and critical seismic scenarios, such as high-magnitude earthquakes. Here, we show the ability of the model to predict the intensity and timing of destructive ground motions in real time, using simulated data from the San Francisco Bay Area. Furthermore, we demonstrate its zero-shot capabilities by evaluating WaveCastNet on real earthquake data. Our approach does not require estimating earthquake magnitudes and epicenters, steps that are prone to error in conventional methods, nor does it rely on empirical ground-motion models, which often fail to capture strongly heterogeneous wave propagation effects.

Geophysics↗
DOE OSTI2025

Model form and sensitivity analysis of CALPHAD-based nucleation models in b-stabilized Ti alloys

Accurate prediction of α-phase nucleation and growth in β-stabilized titanium alloys is crucial for designing heat treatments to optimize mechanical properties in additively manufactured lightweight components. Ideally, predictions of nucleation and growth would incorporate both top-down observations of past experimental heat treatments and bottom-up modeling of phase transformations; however, the appropriate method of combining these information sources is not self-evident. Combining top-down and bottom-up information requires a unified form of model that can connect between spatiotemporal scales, as well as sets of fitting parameters that can be identified by each data source. The selection of which parameters to fit to which data source can be made based on expert opinion, or by performing a sensitivity analysis. In solid-solid nucleation, direct observation of the nucleation and growth process is challenging. Most data on the heat treatment-controlled phase transformations are not in-situ. To predict the process and outcome of the nucleation, growth and coarsening of precipitates, theoretical models of the nucleation pathway are used to bridge the gap. Many sources of uncertainty affect the modeling of this nucleation process. It can be influenced by small variations in the thermomechanical processing history, chemical composition, and initial microstructure. If molecular dynamics (MD) simulations are used to determine thermodynamic quantities and inform CALPHAD modeling, additional uncertainty can be introduced and accounted for using Bayesian methods. Top-down uncertainties require additional steps to quantify. The influence of nucleation model form on the sensitivity of predictions to input parameters and physical conditions is the focus of this study. Classical nucleation theory (CNT) allows modeling to formulate the nucleation as homogeneous or, more commonly, heterogeneous. Non-classical nucleation models are also increasingly explored as a means of reconciling top-down and bottom-up data. In this study, the sensitivity of the intragranular nucleation of α in a β-annealed, slow-cooled aging (BASCA) heat treatment of β-stabilized Ti5553 alloy is explored using CNT and both heterogeneous and homogeneous assumptions. The Kampmann-Wagner Numerical model of precipitate nucleation and growth is employed. Using open-source tools (pyCalphad and thermodynamic modeling of TiMo as a surrogate system, a sensitivity analysis is performed to measure variations in key parameters, including chemical driving force, interfacial energy, and diffusivity, as they relate to predictions of precipitate number density. The inclusion of top-down and bottom-up data in selection of nucleation model form is discussed.

Rodriguez Negron, A. M.↗
DOE OSTI2023

Mitigation of spatial nonstationarity with vision transformers

Spatial nonstationarity, the location variance of features’ statistical distributions, is ubiquitous in many natural settings. For example, in geological reservoirs rock matrix porosity varies vertically due to geomechanical compaction trends, in mineral deposits grades vary due to sedimentation and concentration processes, in hydrology rainfall varies due to the atmosphere and topography interactions, and in metallurgy crystalline structures vary due to differential cooling. Conventional geostatistical modeling workflows rely on the assumption of stationarity to be able to model spatial features for geostatistical inference. Nevertheless, this is often not a realistic assumption when dealing with nonstationary spatial data and this has motivated a variety of nonstationary spatial modeling workflows such as trend and residual decomposition, cosimulation with secondary features, and spatial segmentation and independent modeling over stationary subdomains. The advent of deep learning technologies has enabled new workflows for modeling spatial relationships. However, there is a paucity of demonstrated best practice and general guidance on mitigation of spatial nonstationarity with deep learning in the geospatial context. We demonstrate the impact of two common types of geostatistical spatial nonstationarity on deep learning model prediction performance and propose the mitigation of such impacts using self-attention (vision transformer) models. We demonstrate the utility of vision transformers for the mitigation of nonstationarity with relative errors as low as 10%, exceeding the performance of alternative deep learning methods such as convolutional neural networks. We establish best practice by demonstrating the ability of self-attention networks for modeling large-scale spatial relationships in the presence of commonly observed geospatial nonstationarity.

58 GEOSCIENCES↗
DOE OSTI2024

Short-Term Probabilistic Solar Forecasting via Reinforcement Learning over ECMWF

In this paper, we present an innovative reinforcement learning approach for short-term solar forecasting, leveraging data from the European Centre for Medium-Range Weather Forecasts (ECMWF). The methodology begins with the application of the System Advisor Model (SAM) to transform various ECMWF numerical weather prediction members into predictive photovoltaic power generation. To enhance the precision of deterministic forecasting, we introduce a dynamic model selection algorithm based on Q-learning. This algorithm dynamically identifies and utilizes the most accurate ensemble member for forecasting purposes. Furthermore, we employ a support vector regression surrogate model with a Gaussian distribution to generate probabilistic forecasts, providing a holistic view of solar energy generation uncertainty. To expedite the training process and make it more practical for real-world applications, we integrate a rolling update workflow. This innovative workflow reduces the training period from months to a mere 19 days, making our method highly efficient. Numerical results of the case study show that in comparison to benchmark models, the proposed method improves the deterministic and probabilistic solar forecasting accuracy by up to 40.84% and 48.42%, respectively.

ensemble forecasting↗
DOE OSTI2022

AbBERT: Learning Antibody Humanness via Masked Language Modeling

Understanding the degree of humanness of antibody sequences is critical to the therapeutic antibody development process to reduce the risk of failure modes like immunogenicity or poor manufacturability. We introduce AbBERT, a transformer-based language model trained on up to 20 million unpaired heavy/light chain sequences from the Observed Antibody Space database. We first validate AbBERT using a novel “multi-mask” scoring procedure to demonstrate high accuracy in predicting complementary determining regions—including the challenging hypervariable H3 region. We then demonstrate several uses of AbBERT at various points along the antibody design process. AbBERT enhances in silico antibody optimization via deep reinforcement learning by utilizing its learned embeddings as additional observations during optimization. Within a larger computational antibody design platform, AbBERT has been successfully applied as an additional design objective, where it displays strong correlations with computational tools predicting antibody structural stability. Finally, mutant antibody sequences that have been scored as unfavorable by AbBERT have shown corresponding low yields when expressed in cells. These use cases demonstrate the power of language modeling within computational antibody design.

Bioinformatics↗
DOE OSTI2025

Sub-microsecond Transformers for Jet Tagging on FPGAs

We present the first sub-microsecond transformer implementation on an FPGA achieving competitive performance for state-of-the-art high-energy physics benchmarks. Transformers have shown exceptional performance on multiple tasks in modern machine learning applications, including jet tagging at the CERN Large Hadron Collider (LHC). However, their computational complexity prohibits use in real-time applications, such as the hardware trigger system of the collider experiments up until now. In this work, we demonstrate the first application of transformers for jet tagging on FPGAs, achieving $\mathcal{O}(100)$ nanosecond latency with superior performance compared to alternative baseline models. We leverage high-granularity quantization and distributed arithmetic optimization to fit the entire transformer model on a single FPGA, achieving the required throughput and latency. Furthermore, we add multi-head attention and linear attention support to hls4ml, making our work accessible to the broader fast machine learning community. This work advances the next-generation trigger systems for the High Luminosity LHC, enabling the use of transformers for real-time applications in high-energy physics and beyond.

Laatu, Lauri [Imperial Coll., London]↗
DOE OSTI2022

VWC-BERT: Scaling Vulnerability–Weakness–Exploit Mapping on Modern AI Accelerators

Defending cybersystems needs accurate mapping of software and hardware vulnerabilities to generalized descriptions of weaknesses, and weaknesses to exploits. These mappings enable cyber defenders to build plans for effective defense and assessment of potential risks to a cybersystem. With close to 170k vulnerabilities, manual mapping is not a feasible option. However, automated mapping is challenging due to limited training data, computational intractability, and limitations in computational natural language processing. Tools based on breakthroughs in Transformer-based language models have been demonstrated to classify vulnerabilities with high accuracy. We make three key contributions in this paper: (1) We present a new framework, \VWCBERT, that augments the Transformer-based hierarchical multi-class classification framework of Das et al. (\textsc{V2W-BERT}) with the ability to map weaknesses to exploits. (2) We implement \VWCBERT~ on modern AI accelerator platforms using two data parallel techniques for the pre-training phase and demonstrate nearly linear speedups across NVIDIA and Graphcore accelerator platforms. We observe nearly linear speedups for up to 16 V100 and 8 A100 GPUs, and about 3.4$\times$ speedup for A100 relative to V100 GPUs. We also observe excellent speedups on Graphcore, with $5.7\times$ speedup on 128 IPUs relative to 16 IPUs. Enabled by scaling, we also demonstrate higher accuracy using a larger language model, RoBERTa-Large. We show up to 87\% accuracy for strict and up to 98\% accuracy for relaxed classification. (3) We develop a novel parallel link manager for the link prediction phase and demonstrate up to 21$\times$ speedup with 16 V100 GPUs relative to one V100 GPU, and thus reducing the runtime from 2.5 hours to 10 minutes. We believe that generalizability and scalability of \VWCBERT~ will benefit both the theoretical development and practical deployment of novel cyberdefense solutions and vulnerability classification.

Das, Siddhartha Shankar↗