Engineering Papers⌕ Search

SEARCH · Engineering Papers

Results for “Learning with errors”

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 991 records · Page 55
All sourcesNASA NTRSDOE OSTI
DOE OSTI2021

Apparatus and method for optimizing quantifiable behavior in configurable devices and systems

An apparatus and method are provided to perform constrained optimization of a constrained property of an apparatus, which is complex due to having several components, and these components are configurable in real-time. The optimization is achieved by detecting values of the constrained property and a plurality of other properties of the apparatus when the apparatus is configured in a first subset of the plurality of configurations. A model is learned using the detected values of the constrained property. The model represents the constrained property and can also represent other properties as a function of the configurations. The model can also include estimated uncertainties of the constrained property in the model. Then, using the d model and the estimated uncertainties, the optimal configuration can be selected to minimize an error value (e.g., the difference between a desired value and an observed value of the at least one constrained property).

Hoffmann, Henry↗
DOE OSTI2025

Uncertainty-Guided Prediction Horizon of Phase-Resolved Ocean Wave Forecasting Under Data Sparsity: Experimental and Numerical Evaluation

Accurate short-term wave forecasting is critical for the safe and efficient operation of marine structures that rely on real-time, phase-resolved ocean wave information for control and monitoring purposes (e.g., digital twins). These systems often depend on environmental sensors (e.g., waverider buoys, wave-sensing LIDAR). Challenges arise when upstream sensor data are missing, sparse, or phase-shifted due to drift. This study investigates the performance of two machine learning models, time-series dense encoder (TiDE) and long short-term memory (LSTM), for forecasting phase-resolved ocean surface elevations under varying degrees of data degradation. We introduce the τ-trimming algorithm, which adapts the prediction horizon based on uncertainty thresholds derived from historical forecasts. Numerical wave tank (NWT) and wave basin experiments are used to benchmark model performance under short- and long-term data masking, spatially coarse sensor grids, and upstream phase shifts. Results show under a 50% probability of upstream data loss, the τ-trimmed TiDE model achieves a 46% reduction in error at the most upstream target, compared to 22% for LSTM. Furthermore, phase misalignment in upstream data introduces a near-linear increase in forecast error. Under moderate model settings, a ±3 s misalignment increases the mean absolute error by approximately 0.5 m, while the same error is accumulated at ±4 s using the more conservative approach. These findings inform the design of resilient, uncertainty-aware wave forecasting systems suited for realistic offshore sensing environments.

42 ENGINEERING↗
DOE OSTI2023

Demonstration of machine-learning-enhanced Bayesian quantum state estimation

Machine learning (ML) has found broad applicability in quantum information science in topics as diverse as experimental design, state classification, and even studies on quantum foundations. Here, we experimentally realize an approach for defining custom prior distributions that are automatically tuned using ML for Bayesian quantum state estimation methods that generally better conform to the physical properties of the underlying system than standard fixed prior distributions. Previously, researchers have looked to Bayesian quantum state tomography for advantages like uncertainty quantification, the return of reliable estimates under any measurement condition, and minimal mean-squared error. However, practical challenges related to long computation times and conceptual issues concerning how to incorporate prior knowledge most suitably can overshadow these benefits. Using both simulated and experimental measurement results, we demonstrate that ML-defined prior distributions reduce net convergence times and provide a natural way to incorporate both implicit and explicit information directly into the prior distribution. These results constitute a promising path toward practical implementations of Bayesian quantum state tomography.

71 CLASSICAL AND QUANTUM MECHANICS, GENERAL PHYSIC↗
DOE OSTI2026

Advanced Semi-Supervised Learning with Uncertainty Estimation for Phase Identification in Distribution Systems

The integration of advanced metering infrastructure (AMI) into power distribution networks generates valuable data for tasks such as phase identification; however, the limited and unreliable availability of labeled data in the form of customer phase connectivity presents challenges. To address this issue, we propose a semi-supervised learning (SSL) framework that effectively leverages labeled and unlabeled data. Our approach incorporates self-training, label spreading, and Bayesian neural networks (BNNs) to enhance phase identification with AMI data. Our method uses an ensemble of multilayer perceptron classifiers in a self-training setup, iteratively adding high-confidence pseudo-labels to improve robustness. We also apply label spread to propagate labels based on data similarity, which enhances generalization across diverse distributions. In addition, we employ a BNNs with uncertainty estimation, boosting confidence in predictions and reducing phase identification errors. In our case study, we achieved approximately 98% +/- 0.08 accuracy with uncertainty using minimal and unreliable labeled data from a real U.S. utility, Duquesne Light Company. Our SSL approach, combined with uncertainty estimation, provides an efficient solution for phase identification in AMI data, ultimately improving the reliability of smart grid applications.

24 POWER TRANSMISSION AND DISTRIBUTION↗
DOE OSTI2024

Snow Distribution Patterns Revisited: A Physics-Based and Machine Learning Hybrid Approach to Snow Distribution Mapping in the Sub-Arctic

Snowpack distribution in Arctic and alpine landscapes often occurs in repeating, year-to-year patterns due to local topographic, weather, and vegetation characteristics. Previous studies have suggested that with years of observational data, these snow distribution patterns can be statistically integrated into a snow process modeling workflow. Recent advances in snow hydrology and machine learning (ML) have increased our ability to predict snowpack distribution using in-situ observations, remote sensing data sets, and simple landscape characteristics that can be easily obtained for most environments. Here, we propose a hybrid approach to couple a ML snow distribution pattern (MLSDP) map with a physics-based, snow process model. We trained a random forest ML algorithm on tens of thousands of snow survey observations from a subarctic study area on the Seward Peninsula, Alaska, collected during peak snow water equivalent (SWE). We validated hybrid model outputs using in-situ snow depth and SWE observations, as well as a light detection and ranging data set and a distributed temperature profiling sensor data set. When the hybrid results were compared with the physics-based method, the hybrid method more accurately depicted the spatial patterns of the snowpack, areas of drifting snow, and years when no in-situ observations were used in the random forest ML training data set. The hybrid method also showed improvements in root mean squared error at 61% of locations where time-series estimations of snow depth were observed. These results can be applied to any physics-based model to improve the snow distribution patterning to reflect observed conditions in high latitude and high elevation cold region environments.

54 ENVIRONMENTAL SCIENCES↗
DOE OSTI2025

Inference of response functions with the help of machine-learning algorithms

Response functions are a key quantity to describe the near-equilibrium dynamics of strongly interacting many-body systems. Recent techniques that attempt to overcome the challenges of calculating these ab initio have employed expansions in terms of orthogonal polynomials. We employ a neural network prediction algorithm to reconstruct a response function 𝑆⁡(𝜔) defined over a range in frequencies 𝜔. Here, we represent the calculated response function as a truncated Chebyshev series whose coefficients can be optimized to reduce the representation error. We compare the quality of response functions obtained using coefficients calculated using a neural network (NN) algorithm with those computed using the Gaussian integral transform (GIT) method. In the regime where only a small number of terms in the Chebyshev series are retained, we find that the NN scheme outperforms the GIT method.

Kurkcuoglu, Doga Murat [Fermi National Accelerator↗
DOE OSTI2026

Sensor Reduction for Diversion Detection in a Realistic Heat Pipe Microreactor Using Supervised Machine Learning

Microreactors are designed as a smaller, cheaper, and safer alternative to traditional nuclear power plants. Their non-traditional characteristics and prospect of mass production and deployment will likely require new approaches to nuclear safeguards. The primary proliferation concern with microreactors is the diversion of fuel material. Such diversion may produce measurable defects in key physical attributes like neutron flux, which may in turn be detectable using machine learning models. Preliminary work has demonstrated this ability for modeled nominal and diversion scenarios using large quantities of energy integrated neutron flux data. In practice, the number of available sensors for such measurements will be limited and energy integrated flux information will not be available. This work explores the ability of tree-based gradient boosted ensemble models to classify a given microreactor core is nominal or diversion, and determine the number of fuel pins diverted in the case of diversion with reduced numbers of sensors and more realistic detector responses. Classification accuracy of greater than 98% and regression errors as low as 5% of the total number of fuel pins were achieved with as few as 15 sensors, compared to 99% and 4.1% with a maximum of 240 sensors.

22 - GENERAL STUDIES OF NUCLEAR REACTORS↗
DOE OSTI2023

Dimension Reduction and Redundancy Removal through Successive Schmidt Decompositions

Quantum computers are believed to have the ability to process huge data sizes, which can be seen in machine learning applications. In these applications, the data, in general, are classical. Therefore, to process them on a quantum computer, there is a need for efficient methods that can be used to map classical data on quantum states in a concise manner. On the other hand, to verify the results of quantum computers and study quantum algorithms, we need to be able to approximate quantum operations into forms that are easier to simulate on classical computers with some errors. Motivated by these needs, in this paper, we study the approximation of matrices and vectors by using their tensor products obtained through successive Schmidt decompositions. We show that data with distributions such as uniform, Poisson, exponential, or similar to these distributions can be approximated by using only a few terms, which can be easily mapped onto quantum circuits. The examples include random data with different distributions, the Gram matrices of iris flower, handwritten digits, 20newsgroup, and labeled faces in the wild. Similarly, some quantum operations, such as quantum Fourier transform and variational quantum circuits with a small depth, may also be approximated with a few terms that are easier to simulate on classical computers. Furthermore, we show how the method can be used to simplify quantum Hamiltonians: In particular, we show the application to randomly generated transverse field Ising model Hamiltonians. The reduced Hamiltonians can be mapped into quantum circuits easily and, therefore, can be simulated more efficiently.

97 MATHEMATICS AND COMPUTING↗
DOE OSTI2022

Robust Carbon Dioxide Plume Imaging Using Joint Tomographic Inversion of Seismic Onset Time and Distributed Pressure and Temperature Measurements (Final Report)

We develop and demonstrate rapid and cost-effective methodologies for spatiotemporal tracking of CO2 plumes during geologic sequestration using joint inversion of seismic data and distributed pressure and temperature measurements. Key elements of our methodology are: (a) a computationally efficient approach to pressure and temperature propagation, (b) analysis of time lapse seismic data using a novel ‘seismic onset time’ approach to detect fluid front propagation, and (c) data assimilation and uncertainty assessment via joint inversion of pressure, temperature and time lapse seismic data, and (d) validating the numerical tomographic inversion using a CO2 injection demonstration projects, specifically data collected from the from the Petra Nova Parish Holdings CCUS project in the West Ranch Field, Texas and the Chester-16 reef CO2 injection site in Northern Michigan which is part of the DOE Midwestern Carbon Sequestration Project. The research team is led by Texas A&M University and includes Battelle as a subcontractor with support from Shell, Anadarko, Chevron and JX Nippon. A carbon dioxide (CO2) water-alternating-gas (WAG) pilot was conducted to gain insights into tertiary oil recovery potential via CO2 flood in the West Ranch Field as part of the Petra Nova project, the world’s largest post-combustion CO2 capture and utilization initiative. With a fluvial formation geology and large contrasts in permeability, this is a challenging and novel application of CO2 enhanced oil recovery (EOR). We build a predictive dynamic model of the subsurface that incorporates the multiphase and compositional data acquired during the pilot operation. The calibrated model is used for the carbon dioxide plume imaging. The study began with an initialization of the pilot sector model extracted from a calibrated full-field model. The pilot model calibration follows a two-step hierarchical workflow. First, we performed a large-scale update of the permeability distribution by integrating available bottomhole pressure and multiphase production data. In the second step, local permeability field is fine-tuned using a streamline-based method to match CO2 breakthrough times at the producers. The predictive capability of the calibrated model was verified through two blind validation tests: (1) the model showed good agreement with saturation logs acquired at two observation wells; and (2) the model reproduced the CO2 recovery as a fraction of the injected CO2. The use of seismic onset times has shown great promise for integrating near-continuous seismic surveys for updating geologic models. In this study, we analyze the impact of seismic survey frequency on the onset time approach aiming to extend the application of onset time to infrequent seismic surveys. In addition, we quantitatively examine the nonlinearity of the onset time method and compare it to the commonly used amplitude inversion method. We carry out a sensitivity analysis of seismic survey frequency based on the complete seismic survey data (over 175 surveys) of steam injection in a heavy oil reservoir (Peace River Unit) in Canada. Our results show that an adequate onset time map can be obtained from the infrequent seismic surveys by interpolation between seismic surveys as long as there is no change in the dominant underlying physics between the successive surveys. The study also shows that nonlinearity of the onset time method can be -smaller than that of the amplitude inversion method by several orders of magnitude. Application to the Brugge benchmark case shows that the onset time method obtains comparable permeability update as the traditional seismic amplitude inversion method with faster computation and improved convergence characteristics. We extend the streamline-based data integration approach to incorporate distributed temperature sensor (DTS) data using the concept of thermal tracer travel time. Then, a hierarchical workflow composed of evolutionary and streamline methods is employed to jointly history match the DTS and pressure data. Finally, CO2 saturation and streamline maps are used to visualize the CO2 plume movement during the sequestration process. The hierarchical workflow is applied to a carbon sequestration project in a carbonate reef reservoir within the Northern Niagaran Pinnacle Reef Trend in Michigan, USA. The monitoring data set consists of distributed temperature sensing (DTS) data acquired at the injection well and a monitoring well, flowing bottom-hole pressure data at the injection well, and time-lapse pressure measurements at several locations along the monitoring well. The history matching results indicate that the CO2 movement is mostly restricted to the intended zones of injection which is consistent with an independent warm-back analysis of the temperature data. In addition to employing simulation models and inverse methods for CO2 plume imaging, we also initialized a data-driven technology for detecting inter-well connectivity based on production and pressure data. Our machine-learning framework is built on the statistical recurrent unit (SRU) model and interprets well-based injection/production data into inter-well connectivity without relying on a geologic model. We test it on synthetic and field-scale CO2 EOR projects utilizing the water-alternating-gas (WAG) process. The validation of the proposed data-driven inter-well connectivity assessment is performed using synthetic data from simulation models where inter-well connectivity can be easily measured using the streamline-based flux allocation. The SRU model is shown to offer excellent prediction performance on the synthetic case. Despite significant measurement noise and frequent well shut-ins imposed in the field-scale case, the SRU model offers good prediction accuracy, the overall relative error of the phase production rates at most producers ranges from 10% to 30%. It is shown that the dominant connections identified by the data-driven method and streamline method are in close agreement. Texas A&M University, the lead organization in the project, was primarily responsible for the development of tomographic approaches for CO2 plume mapping in conjunction with distributed pressure, temperature and seismic onset time data. Battelle, as a subcontractor, was primarily responsible for the development of analytical and empirical methods for analyzing transient injection rate and pressure data from point/line sources such as injection and monitoring wells. An additional area of emphasis for Battelle was the use of machine learning for such tasks as inferring reservoir connectivity information from injection-production data, and identifying variable importance for machine learning-based proxy models developed from full-physics simulations. The two organizations also collaborated on the application of the tomographic inversion methodology for a field data set.

02 PETROLEUM↗
DOE OSTI2026

Advanced Semi-Supervised Learning With Uncertainty Estimation for Phase Identification in Distribution Systems

The integration of advanced metering infrastructure (AMI) into power distribution networks generates valuable data for tasks such as phase identification; however, the limited and unreliable availability of labeled data in the form of customer phase connectivity presents challenges. To address this issue, we propose a semi-supervised learning (SSL) framework that effectively leverages labeled and unlabeled data. Our approach incorporates self-training, label spreading, and Bayesian neural networks (BNNs) to enhance phase identification with AMI data. Our method uses an ensemble of multilayer perceptron classifiers in a self-training setup, iteratively adding high-confidence pseudo-labels to improve robustness. We also apply label spread to propagate labels based on data similarity, which enhances generalization across diverse distributions. In addition, we employ a BNNs with uncertainty estimation, boosting confidence in predictions and reducing phase identification errors. In our case study, we achieved approximately 98% +/- 0.08 accuracy with uncertainty using minimal and unreliable labeled data from a real U.S. utility, Duquesne Light Company. Our SSL approach, combined with uncertainty estimation, provides an efficient solution for phase identification in AMI data, ultimately improving the reliability of smart grid applications.

24 POWER TRANSMISSION AND DISTRIBUTION↗
DOE OSTI2021

Encoder–decoder neural network for solving the nonlinear Fokker–Planck–Landau collision operator in XGC

An encoder–decoder neural network has been used to examine the possibility for acceleration of a partial integro-differential equation, the Fokker–Planck–Landau collision operator. This is part of the governing equation in the massively parallel particle-in-cell code XGC, which is used to study turbulence in fusion energy devices. The neural network emphasizes physics-inspired learning, where it is taught to respect physical conservation constraints of the collision operator by including them in the training loss, along with the ℓ 2 loss. In particular, network architectures used for the computer vision task of semantic segmentation have been used for training. A penalization method is used to enforce the ‘soft’ constraints of the system and integrate error in the conservation properties into the loss function. During training, quantities representing the particle density, momentum and energy for all species of the system are calculated at each configuration vertex, mirroring the procedure in XGC. This simple training has produced a median relative loss, across configuration space, of the order of 10 –4 , which is low enough if the error is of random nature, but not if it is of drift nature in time steps. The run time for the current Picard iterative solver of the operator is O(n 2 ), where n is the number of plasma species. As the XGC1 code begins to attack problems including a larger number of species, the collision operator will become expensive computationally, making the neural network solver even more important, especially since its training only scales as O(n). Here, a wide enough range of collisionality has been considered in the training data to ensure the full domain of collision physics is captured. An advanced technique to decrease the losses further will be subject of a subsequent report. Eventual work will include expansion of the network to include multiple plasma species.

70 PLASMA PHYSICS AND FUSION TECHNOLOGY↗
DOE OSTI2025

Deep Learning Approaches for Predicting the Surface Tension of Ionic Liquids

Ionic liquids (ILs) are a novel class of solvents that have attracted significant attention due to their unique and tunable properties. Among their physiochemical characteristics, surface tension plays a critical role in various industrial applications including electrolytes, heat transfer fluids, and separation processes. However, because of the exploratory nature of IL design and the vast combinatorial space of possible anion–cation pairs, the experimental determination of these properties is often impractical, being both time-consuming and costly. To overcome these challenges, computational approaches are increasingly employed to develop accurate predictive models that can accelerate IL discovery and design. In this study, we present two deep learning (DL) models for predicting the surface tension of ILs across a broad temperature range at a constant pressure. The models use simplified molecular input line entry system, SMILES, representations of ILs to extract molecular features as inputs. Both DL models demonstrate excellent agreement with experimental data, achieving an R 2 value of 0.990 and a root-mean-square error of 0.792 mN/m. In conclusion, these results offer valuable insights for the rapid screening and rational design of ILs with tailored surface tension values.

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

Unsupervised acoustic detection of fatigue-induced damage modes from wind turbine blades

This paper proposes a new in-situ damage detection approach for wind turbine blades, which leverages blade-internal non-stationary acoustic pressure fluctuations caused by the mechanical loading as the main source of excitation. This acoustic excitation was leveraged for the detection of fatigue-related damage modes on a full-scale wind turbine blade undergoing edgewise fatigue testing. An unsupervised, data-driven structural health monitoring strategy was developed to learn the normal cavity-internal acoustic sequences generated by the blade’s load cycles and to detect damage-related anomalies in the context of those sequences. A linear cepstral-coefficient based feature set was used to characterize the cavity-internal acoustics and LSTM-autoencoders were trained to accurately reconstruct healthy-case sequences. The reconstruction error was then used to characterize anomalous acoustic patterns within the blade cavity. The technique was able to detect a damage event earlier than a strain-based system by 120,000 load cycles.

17 WIND ENERGY↗
DOE OSTI2021

Quantifying the structural uncertainty of the aerosol mixing state representation in a modal model

Abstract. Aerosol mixing state is an important emergent property that affects aerosol radiative forcing and aerosol–cloud interactions, but it has not been easy to constrain this property globally. This study aims to verify the global distribution of aerosol mixing state represented by modal models. To quantify the aerosol mixing state, we used the aerosol mixing state indices for submicron aerosol based on the mixing of optically absorbing and non-absorbing species (χo), the mixing of primary carbonaceous and non-primary carbonaceous species (χc), and the mixing of hygroscopic and non-hygroscopic species (χh). To achieve a spatiotemporal comparison, we calculated the mixing state indices using output from the Community Earth System Model with the four-mode version of the Modal Aerosol Module (MAM4) and compared the results with the mixing state indices from a benchmark machine-learned model trained on high-detail particle-resolved simulations from the particle-resolved stochastic aerosol model PartMC-MOSAIC. The two methods yielded very different spatial patterns of the mixing state indices. In some regions, the yearly averaged χ value computed by the MAM4 model differed by up to 70 percentage points from the benchmark values. These errors tended to be zonally structured, with the MAM4 model predicting a more internally mixed aerosol at low latitudes and a more externally mixed aerosol at high latitudes compared to the benchmark. Our study quantifies potential model bias in simulating mixing state in different regions and provides insights into potential improvements to model process representation for a more realistic simulation of aerosols towards better quantification of radiative forcing and aerosol–cloud interactions.

54 ENVIRONMENTAL SCIENCES↗
DOE OSTI2024

Prediction of stability constants of metal–ligand complexes by machine learning for the design of ligands with optimal metal ion selectivity

The new LOGKPREDICT program integrates HostDesigner molecular design software with the machine learning (ML) program Chemprop. By supplying HostDesigner with predicted log K values, LOGKPREDICT enhances the computer-aided molecular design process by ranking ligands directly by metal–ligand binding strength. Harnessing reliable experimental data from a historic National Institute of Standards and Technology (NIST) database and data from the International Union of Pure and Applied Chemistry (IUPAC), we train message passing neural net algorithms. The multi-metal NIST-based ML model has a root mean square error (RMSE) of 0.629 ± 0.044 (R 2 of 0.960 ± 0.006), while two versions of lanthanide-only IUPAC-based ML models have, respectively, RMSE of 0.764 ± 0.073 (R 2 of 0.976 ± 0.005) and 0.757 ± 0.071 (R 2 of 0.959 ± 0.007). For relative log K predictions on an out-of-sample set of six ligands, demonstrating metal ion selectivity, the RMSE value reaches a commendably low 0.25. Here we showcase the use of LOGKPREDICT in identifying ligands with high selectivity for lanthanides in aqueous solutions, a finding supported by recent experimental evidence. We also predict new ligands yet to be verified experimentally. Therefore, our ML models implemented through LOGKPREDICT and interfaced with the ligand design software HostDesigner pave the way for designing new ligands with predetermined selectivity for competing metal ions in an aqueous solution.

37 INORGANIC, ORGANIC, PHYSICAL, AND ANALYTICAL CH↗
DOE OSTI2022

Modeling Spatial Distribution of Snow Water Equivalent by Combining Meteorological and Satellite Data with Lidar Maps

Abstract An accurate characterization of the water content of snowpack, or snow water equivalent (SWE), is necessary to quantify water availability and constrain hydrologic and land surface models. Recently, airborne observations (e.g., lidar) have emerged as a promising method to accurately quantify SWE at high resolutions (scales of ∼100 m and finer). However, the frequency of these observations is very low, typically once or twice per season in the Rocky Mountains of Colorado. Here, we present a machine learning framework that is based on random forests to model temporally sparse lidar-derived SWE, enabling estimation of SWE at unmapped time points. We approximated the physical processes governing snow accumulation and melt as well as snow characteristics by obtaining 15 different variables from gridded estimates of precipitation, temperature, surface reflectance, elevation, and canopy. Results showed that, in the Rocky Mountains of Colorado, our framework is capable of modeling SWE with a higher accuracy when compared with estimates generated by the Snow Data Assimilation System (SNODAS). The mean value of the coefficient of determination R 2 using our approach was 0.57, and the root-mean-square error (RMSE) was 13 cm, which was a significant improvement over SNODAS (mean R 2 = 0.13; RMSE = 20 cm). We explored the relative importance of the input variables and observed that, at the spatial resolution of 800 m, meteorological variables are more important drivers of predictive accuracy than surface variables that characterize the properties of snow on the ground. This research provides a framework to expand the applicability of lidar-derived SWE to unmapped time points. Significance Statement Snowpack is the main source of freshwater for close to 2 billion people globally and needs to be estimated accurately. Mountainous snowpack is highly variable and is challenging to quantify. Recently, lidar technology has been employed to observe snow in great detail, but it is costly and can only be used sparingly. To counter that, we use machine learning to estimate snowpack when lidar data are not available. We approximate the processes that govern snowpack by incorporating meteorological and satellite data. We found that variables associated with precipitation and temperature have more predictive power than variables that characterize snowpack properties. Our work helps to improve snowpack estimation, which is critical for sustainable management of water resources.

54 ENVIRONMENTAL SCIENCES↗
DOE OSTI2026

Machine learning based prediction of airflow maldistribution in air-to-refrigerant heat exchangers

Flow maldistribution is a common challenge in heat exchanger (HX) design and particularly important for air-to-refrigerant geometries where capacity losses can approach 65%. This has a major impact on central air conditioning systems, as compact duct design motivates the use of A-type HXs which are known to be affected by airflow maldistribution. Because velocity profiles are difficult to predict, components are often oversized leading to increased material cost, system footprint, and refrigerant charge. Several studies detail airflow maldistribution for individual HXs and packages, but findings cannot always be extrapolated to new designs. In this work, a machine learning (ML) based flow profile prediction framework is developed and applied to two common package configurations: (i) A-type and (ii) U-type HXs, across a broad range of HX geometries and flow rates. Porous media CFD simulations are validated against independent data for both package types as well as comprehensive in house measurements for a finless geometry with shape optimized non-round tubes, which validates the framework for new heat transfer surfaces. The ML models are trained on the porous media CFD simulations, predicting volumetric flow rate (VFR) within 1.1% and 1.9% with maximum relative L 2 norm errors of 0.48 and 0.65, respectively, while also delivering 10 5 speed up factor compared to full porous media CFD. HX level simulations show an up to 9% reduction in heat transfer from flow maldistribution, with greater losses occurring at smaller half apex angles. This framework enables rapid and highly accurate prediction of airflow maldistribution induced capacity degradation.

42 ENGINEERING↗
DOE OSTI2025

Enhancing Autonomous Control of Microreactors Using Multi-Agent Reinforcement Learning

In order for microreactors to be economically competitive, operation costs will need to be minimized through some degree of autonomous control. Previous work has demonstrated the effectiveness of reinforcement learning (RL) for load-following control in a drum-controlled microreactor. This study extends that work by exploring the potential of RL to independently control each of the reactor’s drums. We compare a single-agent RL approach with a multi-agent RL (MARL) framework, testing them for generalization across different load-following power profiles and control timescales, and for robustness in cases of randomly disabled control drums. Since the point kinetics simulation environment used in this study cannot resolve spatial effects, we assume that in the absence of spatially localized disturbances, optimal drum movements should be symmetrical. We demonstrate that single-agent RL is able to achieve accurate performance only when symmetric actions are ignored; otherwise, it fails to train a useful controller. Meanwhile, the MARL framework performs symmetric actions by design and trains a robust, accurate agent, as evidenced by mean absolute errors in power matching of 0.41% for the training power profile, 0.68% for a profile with half the drums disabled, and 0.21% for a profile on a realistic load-following time horizon.

22 - GENERAL STUDIES OF NUCLEAR REACTORS↗