Engineering Papers⌕ Search

SEARCH · Engineering Papers

Results for “network performance evaluation”

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 505 records · Page 28
All sourcesNASA NTRSDOE OSTI
DOE OSTI2021

Verification and validation of developed short-term forecasting models

Recent advancements in machine learning (ML) and artificial intelligence (AI) technologies provide an opportunity for leveraging data-driven algorithms to predict future nuclear power plant (NPP) operating conditions by using recorded plant process data. Successfully implementing these models can lead to cost-reducing, conditioned-based predictive maintenance through optimized maintenance schedules and a reduction of unnecessary maintenance activities. This report discusses the verification and validation of short-term forecasting processes (i.e., data cleaning, feature selection, model optimization, and forecasting) developed in previous reports. The verification and validation (V&V) process demonstrates the expected precision and accuracy when the ML model encounters new datasets from different systems. Shapley additive explanations were used as the primary means of feature selection across these different data set. Individual models were trained for each data set, then validated through a cross-validation procedure. In this report, two different ML models were tasked to predict variables from three different plant process data sets with varying prediction horizons. The results indicate that support vector regression (SVR) outperformed long short-term memory (LSTM) neural networks in regard to each data set and each prediction horizon in this study, but further tuning and optimization could improve long short-term memory results. However, each forecasting model showed reduced performance as the prediction horizon was extended from 1 hour to 1 day ahead. Research is ongoing to evaluate the optimal input variable space, which is based on a given set of process parameters, to further improve forecasting accuracy.

46 INSTRUMENTATION RELATED TO NUCLEAR SCIENCE AND ↗
DOE OSTI2025

Web-Based Tools for Data-Informed Remedy Optimization: Software Theory and User Guide

This report documents the development and application of two web-based decision-support tools for pump-and-treat (P&T) groundwater remediation systems: PTOLEMY (Pump-and-Treat Optimized Location Evaluation to Maximize Yields) and OPTIMA (Optimization for Pump-and-Treat Implementation, Management, & Assessment). These tools enhance remedy design and management by leveraging advanced computational methods – specifically deep learning and multi-objective optimization – within a user-friendly platform. By integrating data-driven models with established hydrogeological knowledge, PTOLEMY and OPTIMA enable more efficient evaluation of well placement and operational strategies, helping site managers balance multiple remediation objectives under complex conditions. Both tools are implemented as modules within the SOCRATES (Suite Of Comprehensive Rapid Analysis Tools for Environmental Sites) web platform, which provides data access, visualization, and analytics to support remedy optimization across sites in the U.S. Department of Energy Office of Environmental Management complex. PTOLEMY is a rapid screening module designed to identify promising locations for new extraction wells. It employs a multi-channel three-dimensional convolutional neural network (MC3D-CNN) trained on high-fidelity simulation data to predict the relative performance (in terms of contaminant mass recovery) of potential well sites. Through an interactive web interface, PTOLEMY visualizes the probability of high performance across a site, highlighting areas where an extraction well is likely to yield above-threshold contaminant removal over a multi-year period. PTOLEMY’s map-based displays and exportable results support transparent communication of screening analyses. By focusing attention on the most favorable candidate locations, the tool augments traditional engineering judgment and physics-based modeling, providing a data informed basis for subsequent detailed evaluations. OPTIMA is a multi objective optimization module designed to find wellfield layouts and operating schedules that meet various cleanup goals. It quickly evaluates thousands of candidate setups – combinations of well locations, timing, and rates – and returns a small set of best trade-off options for comparison. At its core, OPTIMA uses a U-Net-based surrogate model – a deep-learning emulator of a groundwater flow and transport simulator – to dramatically accelerate scenario evaluations. Coupling this fast surrogate with the NSGA-II (Non-dominated Sorting Genetic Algorithm II) evolutionary algorithm, OPTIMA explores a wide decision space of well locations and schedules to identify Pareto-optimal solutions that trade off key objectives (e.g., minimizing cleanup time, maximizing contaminant mass removal, and minimizing plume extent). The tool outputs a family of optimal configurations and visualizes their trade-offs (Pareto frontiers of cleanup metrics and maps of optimized well placements). Site managers can use these results to understand the range of viable strategies and to select candidate designs for more detailed verification. OPTIMA is currently under active development and not yet fully released; this guide provides early documentation to support planning and gather user feedback.

54 ENVIRONMENTAL SCIENCES↗
DOE OSTI2025

Monte Carlo Dropout Uncertainty Quantification of Long Short-Term Memory Autoencoder Anomaly Detection in a Liquid Sodium Cold Trap

Advanced high-temperature fluid reactors, such as sodium-cooled fast reactors (SFRs) and molten salt–cooled reactors (MSCRs), require coolant purification systems to prevent fluid contamination and local freezing that can lead to plugging. Liquid sodium purification can be achieved with a cold trap, where the sodium temperature is reduced to a near-freezing point to precipitate out impurities. Automation of monitoring of the cold trap performance with machine learning algorithms can aid in early detection of incipient anomalies. An efficient approach to loss-of-coolant–type anomaly detection in a cold trap monitored with more than two dozen thermal-hydraulic sensors consists of a long short-term memory (LSTM) autoencoder. This work develops the uncertainty quantification of the LSTM autoencoder performance for cold trap anomaly detection using the Monte Carlo (MC) dropout method. The MC dropout methodology creates a distribution of sister distributions that all slightly differ from each other because of random neurons being turned off for testing. The variances of the sister network distributions are used to make an uncertainty interval. Our analysis shows that the uncertainty in the autoencoder performance is largest near the peak of the anomaly signal. Using the MC dropout method, we investigate the uncertainty in the anomaly detection with missing sensor inputs. This capability allows the reactor operator to evaluate resilience of the anomaly detection system and to make informed decisions about continuity of operation in the event of sensor failure.

22 GENERAL STUDIES OF NUCLEAR REACTORS↗
DOE OSTI2026

Community detection robustness of graph neural networks

Graph neural networks (GNNs) are increasingly widely used for community detection in attributed networks. They combine structural topology with node attributes through message passing and pooling. However, their robustness or lack thereof with respect to different perturbations and targeted attacks in conjunction with community detection tasks is not well understood. To shed light on latent mechanisms behind GNN sensitivity on community detection tasks, we conduct a systematic computational evaluation of six widely adopted GNN architectures graph convolutional network, graph attention network, graph sample and aggregate (GraphSAGE), differentiable pooling (DiffPool), minimum cut pooling (MinCUT), and deep modularity networks (DMoN). The analysis covers three perturbation categories: node attribute manipulations, edge topology distortions, and adversarial attacks. We use element-centric similarity as the evaluation metric on synthetic benchmarks and real-world citation networks. Our findings indicate that supervised GNNs tend to achieve higher baseline accuracy, while unsupervised methods, particularly DMoN, maintain stronger resilience under targeted and adversarial perturbations. Furthermore, robustness appears to be strongly influenced by community strength, with well-defined communities reducing performance loss. Across all models, node attribute perturbations associated with targeted edge deletions and shifts in attribute distributions tend to cause the largest degradation in community recovery. These findings highlight important trade-offs between accuracy and robustness in GNN-based community detection and offer insights into selecting architectures resilient to noise and adversarial attacks.

Goel, Jaidev [Virginia Polytechnic Inst. and State↗
DOE OSTI2022

Control of a Three-Phase Grid-Connected Voltage-Sourced Converter Using Long Short-Term Memory Networks

With the rise of inverter-based resources (IBRs) within the power system, the control of grid-connected converters (GCCs) has become pertinent due to the fact they interface IBRs to the grid. The conventional method of control for a GCC such as the voltage-sourced converter (VSC) is through a decoupled control loop in the synchronous reference frame. However, this model-based control method is sensitive to parameter changes causing deterioration in controller performance. Data-driven approaches such as machine learning can be utilized to design controllers that are capable of operating GCCs in various system conditions. This work explores a deep learning-based control method for a three-phase grid-connected VSC, specifically utilizing a long short-term memory (LSTM) network for robust control. Simulations of a conventional controlled VSC are conducted using Simulink to collect data for training the LSTM-based controller. The LSTM model is built and trained using the Keras and TensorFlow libraries in Python and tested in Simulink. The performance of the LSTM-based controller is evaluated under different case studies and compared to the conventional method of control. Simulation results demonstrate the effectiveness of this approach by outperforming the conventional controller and maintaining stability under different system parameter changes.

24 POWER TRANSMISSION AND DISTRIBUTION↗
DOE OSTI2024

Performance Comparison of Machine Learning Models for Ultrasonic Nondestructive Evaluation of Alkali-Silica Reaction in Concrete

Alkali-silica reaction (ASR) causes concrete degradation, leading to cracking, rebar corrosion, and reduced structural integrity, which raises safety concerns. Ultrasonic nondestructive evaluation (NDE) effectively assesses concrete properties and monitors ASR progression. However, its deployment and analysis require specialized expertise and subjective interpretation. As computational power increases, artificial intelligence (AI) and machine learning (ML) algorithms are increasingly being used to automate NDE data analysis across various industries for AI-assisted automation. Regulatory agencies are adapting to this technological shift, prompting a need to evaluate current ML technologies’ capabilities and limitations in assessing concrete material properties and damage. This report presents a comparative analysis of four ML regression models for predicting concrete material damage induced by ASR expansion using long-term ultrasonic data monitoring. The models investigated include linear regression (LR), support vector regression (SVR), shallow neural networks (NN), and deep neural networks (DNN). LR, SVR, and shallow NN models use features extracted from ultrasonic signals, whereas the DNN model processes time-domain ultrasonic signals and frequency spectra directly. The study systematically compared the models’ performance from various perspectives, including model input, prediction performance, and generalization ability. The findings indicate significant variability in model performance, with some ML algorithms achieving very high or very low prediction accuracy depending on the preprocessing and feature engineering (extraction and selection) applied. Key insights include the observation that shallow ML models (LR, SVR, and shallow NNs) require meticulous preprocessing and feature extraction to achieve high accuracy. In contrast, the DNN model, although it bypasses the need for feature engineering, necessitates extensive preprocessing to mitigate noise and computational demands. The SVR model emerged as the top performer among the shallow models, and the DNN model exhibited superior performance on specific datasets but struggled with generalization across specimens from different batches. Additionally, the SVR model is sensitive to temperature variations, whereas the DNN model is robust in this regard. Using recurrent neural networks is recommended for future ASR expansion prediction studies. Recurrent neural networks’ inherent ability to capture temporal dependencies and long-term patterns makes them well suited for analyzing sequential ultrasonic monitoring data. Overall, the results and conclusions of this study could provide insights into the capabilities and effectiveness of ML when applied to ultrasonic NDE data and help identify best practices for using ML for ultrasonic NDE of concrete material properties.

36 MATERIALS SCIENCE↗
DOE OSTI2022

Evaluation and Optimization of Well Completion Options for the Utah FORGE Site

Orientation and completion for well pairs that have been subjected to multi-zonal stimulation play a critical role in the long-term performance of an Enhanced Geothermal Reservoir. Enhanced geothermal systems often rely on preferential flow along fractures between well injection and production locations. Modeling this preferential flow using discrete fracture networks (DNF) relies on stochastic realizations of the DFN based on geological sampling. Here we present the development of a stochastic optimization methodology to determine well completion options in a discrete fracture network based on using parallel subset simulation. Stochastic optimization will provide insight into regions where placements of the injection and production wells are optimal. An example optimization of well-pair location optimization based on a deterministic-stochastic DFN model representing FORGE follows a discussion of the theory.

15 GEOTHERMAL ENERGY↗
DOE OSTI2022

Classification of Nuclear Reactor Operations Using Spatial Importance and Multisensor Networks

Distributed multisensor networks record multiple data streams that can be used as inputs to machine learning models designed to classify operations relevant to proliferation at nuclear reactors. The goal of this work is to demonstrate methods to assess the importance of each node (a single multisensor) and region (a group of proximate multisensors) to machine learning model performance in a reactor monitoring scenario. This, in turn, provides insight into model behavior, a critical requirement of data-driven applications in nuclear security. Using data collected at the High Flux Isotope Reactor at Oak Ridge National Laboratory via a network of Merlyn multisensors, two different models were trained to classify the reactor’s operational state: a hidden Markov model (HMM), which is simpler and more transparent, and a feed-forward neural network, which is less inherently interpretable. Traditional wrapper methods for feature importance were extended to identify nodes and regions in the multisensor network with strong positive and negative impacts on the classification problem. These spatial-importance algorithms were evaluated on the two different classifiers. The classification accuracy was then improved relative to baseline models via feature selection from 0.583 to 0.839 and from 0.811 ± 0.005 to 0.884 ± 0.004 for the HMM and feed-forward neural network, respectively. While some differences in node and region importance were observed when using different classifiers and wrapper methods, the nodes near the facility’s cooling tower were consistently identified as important—a conclusion further supported by studies on feature importance in decision trees. Node and region importance methods are model-agnostic, inform feature selection for improved model performance, and can provide insight into opaque classification models in the nuclear security domain.

Tibbetts, Jake↗
DOE OSTI2023

Influence of oxygen on performance of multi-principal element alloy as braze filler for Ni-base alloys

In manufacturing of multi-principal element alloys (MPEAs), the casting process is often a primary point of oxygen introduction to the material system, which was demonstrated by the laboratory scale button arc-melting process. Oxygen introduction raises concerns for the mechanical performance of these alloys if employed as structural materials alone or in engineering applications, such as serving as filler materials to enable joining of similar or dissimilar alloy pairs that are conventionally considered difficult to join. In this work, oxide inclusions in an as-cast off-equiatomic MnFeCoNiCu MPEA and Ni-base alloy braze joints made with this MPEA were evaluated by synchrotron x-ray diffraction mapping and electron microscopy. MnO was found to be prevalent throughout the cast MPEA and was chemically reduced during brazing by trace Al from the base material, Ni-base Alloy 600. Incomplete reaction of the MnO with Al, and some alternative reaction with Cr, left a mixture of MnO, Al 2 O 3 , and Cr 2 O 3 in the as-brazed joint, with the oxides concentrated near the braze centerline due to directional solidification. Further, in the braze joints, the non-oxide constituent phases detected were an FCC matrix and particles of Cr 23 C 6 , Cr 7 C 3 , and TiN which are all native to Alloy 600. The evolution of oxides within the brazed joints at elevated service temperatures of 600-800°C was also evaluated. While MnO and Al 2 O 3 were stable during high-temperature service, Cr 2 O 3 particles grew to several hundred microns, with dissolved oxygen in the MPEA providing a significant source for oxide growth. Comparing tensile performance of the as-brazed condition with post-service conditions showed that the evolution of oxide particles contributed to an increase in strength and decrease in ductility of the brazed joints. Evaluating particle morphologies among fracture surfaces of individual specimens demonstrated that large Cr 2 O 3 particles and continuous cluster networks of Al 2 O 3 were the features most detrimental to total elongation. Hence, unmitigated evolution of oxygen-containing species during brazing and service conditions is detrimental to the ductility of the brazed joints, which highlights the importance of oxygen control in initial manufacturing of MPEAs for engineering applications.

36 MATERIALS SCIENCE↗
DOE OSTI2020

Comparing the performance of density functionals in describing the adsorption of atoms and small molecules on Ni(111)

Reaction networks calculated using density functional theory (DFT) play a crucial role in understanding and predicting heterogeneous catalyst performance. However, results obtained using different density functionals can vary significantly. Accordingly, it is important to assess the accuracy of available functionals in capturing adsorption energetics derived from single-crystal adsorption microcalorimetry experiments. Here, we evaluate how different density functionals describe the molecular and dissociative adsorption of CH 3 I, the adsorption of CH 3 , I, and H, and the dissociative adsorption of CH 4 on Ni(111). First, we identify energetically-preferred adsorbed states for each adsorbate at three surface coverages and calculate the corresponding enthalpies of adsorption at 160 K. Then, we compare the results to experimental measurements published in the literature. For each adsorption process we find at least one functional that is quantitatively accurate within experimental error, but no functional is accurate (within experimental error) for all adsorption processes studied. Even when assuming an additional DFT error of ±20 kJ/mol on top of the experimental error, only the PBE-D3 and RPBE-D3 functionals are accurate for all considered adsorption systems. Overall, these results indicate that quantitative agreement between density functional theory calculations and experimental measurements is both system- and functional-dependent. Importantly, single-crystal adsorption microcalorimetry experiments will continue to play an important role for benchmarking density functionals.

37 INORGANIC, ORGANIC, PHYSICAL, AND ANALYTICAL CH↗
DOE OSTI2021

Extracting low energy signals from raw LArTPC waveforms using deep learning techniques -- A proof of concept

We investigate the feasibility of using deep learning techniques, in the form of a one-dimensional convolutional neural network (1D-CNN), for the extraction of signals from the raw waveforms produced by the individual channels of liquid argon time projection chamber (LArTPC) detectors. A minimal generic LArTPC detector model is developed to generate realistic noise and signal waveforms used to train and test the 1D-CNN, and evaluate its performance on low-level signals. We demonstrate that our approach overcomes the inherent shortcomings of traditional cut-based methods by extending sensitivity to signals with ADC values below their imposed thresholds. This approach exhibits great promise in enhancing the capabilities of future generation neutrino experiments like DUNE to carry out their low-energy neutrino physics programs.

46 INSTRUMENTATION RELATED TO NUCLEAR SCIENCE AND ↗
DOE OSTI2024

Computational Approaches for Clean Energy Materials

Currently, 80% of the global final energy consumption occurs in form of fuels and only 20% as electricity. On the other hand, renewable energy additions come almost exclusively in the form of electricity (dominantly photovoltaics and wind). Thus, a successful energy transition will require enormous growth in renewables, sufficient to convert excess electricity into fuels, as well as the development of non-electricity based solar fuel technologies. As much as photovoltaic capacities have grown over the past 20 years, it is far from clear that current technologies and materials are up to the task to grow from here by yet another factor 100 until 2050. Therefore, sustained research efforts on emerging inorganic semiconductors for solar electricity and fuels are essential for facing the double challenge of climate change and energy security. Computational materials science can make important contributions, guiding and supporting research activities through both materials search and discovery and through detailed studies that help to develop a mechanistic understanding of materials performance and bottlenecks. This presentation will highlight three recent computational projects with relevance for photovoltaics and solar fuels (1) Defect graph neural networks (dGNN) for materials discovery in solar thermochemical hydrogen (STCH) [1]. The dGNN approach facilitates broad and fast materials screening for defect properties. (2) Modeling highly off-stoichiometric systems by evaluating the free energy of defect interaction [2]. This approach allows quantitative prediction of H2 production in complex STCH oxides. (3) First-principles atomic structure prediction for interfaces [3]. This work showed how an atomically thin CdCl2 interlayer phase enables in principle ideal electron transport across the incommensurate SnO2/CdTe interface. [1] M.D. Witman, A. Goyal, T. Ogitsu, A.H. McDaniel, S. Lany, Nat. Comput. Sci. (2023). https://doi.org/10.1038/s43588-023-00495-2. [2] A. Goyal, M.D. Sanders, R.P. O'Hayre, S. Lany, PRX Energy 3, 013008 (2024). https://doi.org/10.1103/PRXEnergy.3.013008. [3] A. Sharan, M. Nardone, D. Krasikov, N. Singh, S. Lany, Appl. Phys. Rev. 9, 041411 (2022). https://doi.org/10.1063/5.0104008.

density functional theory↗
DOE OSTI2020

Deep learning-based model for progress variable dissipation rate in turbulent premixed flames

A deep neural network (DNN) based large eddy simulation (LES) model for progress variable dissipation rate in turbulent premixed flames is presented. The DNN model is trained using filtered data from direct numerical simulations (DNS) of statistically planar turbulent premixed flames with n-heptane as fuel. Training data was comprised of flames with varying turbulence levels leading to a range of Karlovitz numbers. Through a-priori tests the DNN model is shown to predict the subfilter contribution to progress variable dissipation rate accurately over a range of filter widths and for all Karlovitz numbers examined in this study. Superior performance of the DNN model relative to an established physics-based model is also demonstrated. Additionally, transferability of the DNN model is highlighted by a-priori evaluation of the model using filtered DNS data from multiple cases with different Karlovitz numbers and fuel species than those that were used for training the model.

33 ADVANCED PROPULSION SYSTEMS↗
DOE OSTI2026

Explaining System-Level Prognostics with Established Machine Learning Methods

System-level prognostics is crucial for ensuring reliability and enabling predictive maintenance in complex systems with interconnected components. This study presents a framework that integrates data-driven methods to predict the remaining useful life (RUL) of a subsystem under multiple and concurrent faults within a nuclear power plant system with explainable artificial intelligence (XAI). A nuclear power plant (NPP) operation was simulated to model the degradation behavior of NPP components, and four machine learning models—Gradient Boosting Regressor (GBR), Support Vector Regressor (SVR), Fully Connected Neural Network (FCNN), and Long Short-Term Memory (LSTM)—were evaluated for prognostics with a novel system RUL parameter. The LSTM model demonstrated potential superior repeatability, while SHAP (SHapley Additive exPlanations) for explainability provided consistent and trustworthy global explanations. In contrast, LIME (Local Interpretable Model-agnostic Explanations) offered localized interpretability but showed reduced stability for sequential data. Key findings include the interplay between component-level degradation and system-wide performance, with LSTM effectively capturing these dynamics through sequence-level predictions. The XAI techniques enhanced transparency by identifying critical features influencing model predictions and aligning with domain knowledge. Furthermore, this framework has significant implications for improving trust and understanding in predictive maintenance, particularly in safety-critical industries like nuclear energy.

22 GENERAL STUDIES OF NUCLEAR REACTORS↗
DOE OSTI2021

Computational and numerical analysis of AC optimal power flow formulations on large-scale power grids

We report that alternating current optimal power flow (AC-OPF) is a fundamental tool in electric utilities to determine optimal operation of the various resources. Typically, the AC-OPF problem uses power balance formulation containing voltages and power equations. Yet, there is no comprehensive comparison of the different AC-OPF formulations, especially for large-scale networks. This paper presents a detailed comparative evaluation of different formulations of the AC-OPF problem on networks ranging from 9-bus to 25,000 buses. Three different formulations: 1) power balance with polar voltages, 2) power balance with Cartesian voltages, and 3) current balance with Cartesian voltages are discussed in detail by comparing their characteristics, and numerical and computational performance.

24 POWER TRANSMISSION AND DISTRIBUTION↗
DOE OSTI2025

Accelerating high-order continuum kinetic plasma simulations using multiple GPUs

Kinetic plasma simulations solve the Vlasov-Poisson or Vlasov-Maxwell equations to evolve scalar-variable distribution functions in position-velocity phase space and vector-variable electromagnetic fields in configuration space. The immense computational cost of evolving high-dimensional variables, and their large number of degrees of freedom, often limits the utility of continuum kinetic simulations and presents a challenge when it comes to accurately simulating real-world physical phenomena. To address this challenge, we present techniques that accelerate and minimize the computational work required for a scalable Vlasov-Poisson solver. We show theoretical hardware compute and communication bounds for solving a fourth-order finite-volume Vlasov-Poisson system. These bounds are then used to inform and evaluate the design of performance portable algorithms for a multiple graphics processing unit (GPU) accelerated version of the Vlasov-Poisson solver VCK-CPU [1]. We demonstrate that the multi-GPU Vlasov solver implementation, VCK-GPU, simultaneously minimizes required inter-process data transfer while also being bounded by the machine network performance limits. This results in an overall strong scaling speedup per timestep of up to 40x in three-dimensional phase space (one position, two velocity coordinates) and 54x in four dimensional phase space (two position, two velocity coordinates) and a 341x increase in simulation throughput of the GPU accelerated code over the existing CPU code. The GPU code is also able to weak scale up to 256 compute nodes and 1024 GPUs. In conclusion, we demonstrate that the improved compute performance enables exploring configurations which were previously computationally infeasible, including resolving fine-scale distribution function filamentation and multi-species dynamics with realistic electron-proton mass ratios.

Continuum kinetics↗
DOE OSTI2026

ChemGraph as an agentic framework for computational chemistry workflows

Atomistic simulations are essential in chemistry and materials science but remain challenging to run due to the expert knowledge required for the setup, execution, and validation stages of these calculations. We present ChemGraph, an agentic framework powered by artificial intelligence and state-of-the-art simulation tools to streamline and automate computational chemistry and materials science workflows. ChemGraph leverages graph neural network-based foundation models for accurate yet computationally efficient calculations and large language models (LLMs) for natural language understanding, task planning, and scientific reasoning to provide an intuitive and interactive interface. We evaluate ChemGraph across 13 benchmark tasks and demonstrate that smaller LLMs (GPT-4o-mini, Claude-3.5-haiku, Qwen-2.5-14B) perform well on simple workflows, while more complex tasks benefit from using larger models. Importantly, we show that decomposing complex tasks into smaller subtasks through a multi-agent framework enables GPT-4o to reach perfect accuracy and smaller LLMs to match or exceed single-agent GPT-4o's performance in these benchmarks.

Computational chemistry↗