SEARCH · Engineering Papers
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.
The potential for errors in engineering judgement arguably present the highest level of risk when applied to interplanetary spaceflight. JPL has been refining its lessons learned process to optimize the collection and transfer of critical success factors applicable to current and future spaceflight projects.
This study examined the improvement of microscopy segmentation accuracy by transfer learning from a large dataset of microscopy images called MicroNet. Many neural network encoder architectures, including VGG, Inception, and ResNet, were trained on over 100,000 labelled microscopy images from 54 classes. These pre-trained encoders were then embedded into multiple segmentation architectures including U-Net and DeepLabV3+ to evaluate segmentation performance on newly created benchmark microscopy datasets. Compared to ImageNet pre-training, models pre-trained on MicroNet generalized better to out-of-distribution micrographs taken under different imaging and sample conditions and were more accurate with less training data. When training with only a single Ni-superalloy image, pre-training on MicroNet produced a 72.2 percent reduction in relative segmentation error. These results suggest that transfer learning from large in-domain datasets generate models with learned feature representations that are more useful for downstream tasks and will likely improve any microscopy image analysis technique that can leverage pre-trained encoders.
Accurately forecasting wind and solar power output poses challenges for deeply decarbonized electricity systems. Grid operators must commit resources to provide reserves to ensure reliable operations in the face of forecast errors, a process which can increase fuel consumption and emissions. We apply neural network-based machine learning to expand the usefulness of median point forecast data by creating probabilistic distributions of short-term uncertainty in demand, wind, and solar forecasts that adapt to prevailing grid conditions. Machine learning derived estimates of forecast errors compare favorably to estimates based on incumbent methods. Reserves derived from machine learning are usually smaller than values derived using incumbent methods, which enables fuel savings during most hours. Machine learning reserves are generally larger than incumbent reserves during times of higher forecast error, potentially improving system reliability. Performance is tested using multi-stage production simulation modeling of the California Independent System Operator (CAISO) system. Machine learning reserves provide production cost and greenhouse gas (GHG) emission reductions of approximately 0.3% relative to historical 2019 requirements. Savings in the 2030 timeframe are highly dependent on battery storage capacity. At lower levels of battery capacity, savings of 0.4% from machine learning reserves are shown. Significant quantities of battery storage are expected to be added to meet California's resource adequacy needs and GHG reduction targets. Addition of these batteries saturate reserve needs and results in minimal within-hour balancing costs in 2030.
There were three major goals and objectives of this project: 1) Develop tools to understand density-driven errors in transition metal complexes, 2) Evaluate and minimize energy delocalization error and static correlation error through judicious functional choice, and 3) applying this workflow to machine-learning accelerated screening of redox couples. Over the reporting period, we developed a framework for eliminating flat plane errors. We introduced fully non-empirical coefficients. We demonstrated the approach on both molecules and solids. We investigated and eliminated density driven errors and demonstrated their impact on potential energy surfaces. We trained machine learning models both in a method-dependent fashion and to predict errors in method accuracy. We built large data sets of small molecule energetics and multi-reference character.
Fully-parallel hardware neural network implementations may be applied to high-speed recognition, classification, and mapping tasks in areas such as vision, or can be used as low-cost self-contained units for tasks such as error detection in mechanical systems (e.g. autos). Learning is required not only to satisfy application requirements, but also to overcome hardware-imposed limitations such as reduced dynamic range of connections.
A pulsed spallation target is subjected to very short (∼0.7μs) but intense loads (23.3 kJ) from repeated proton pulses, which knock away neutrons from the mercury atoms’ nuclei for a wide range application in physics, engineering, medicine, petroleum exploration, biology, chemistry, etc. The effect of this pulsed loading on the stainless-steel target module which contains the flowing mercury target material is difficult to predict not only due to its short but intense explosive-like physical reaction, but also the nonlinear material behavior of the liquid mercury in the structure. Injecting small helium bubbles in the mercury has been an efficient method of mitigating the pressure wave at high power level stage. However, prediction of the resultant loading on the target is more difficult when helium gas is intentionally injected into the mercury. A 2-phase material model that incorporates the Rayleigh-Plesset (R-P) model is expected to address this complex multi-physics dynamics problem by including the bubble dynamics in the liquid mercury. A parameter sensitivity study was firstly employed to understand their impact on the simulation strains. The investigated parameters included E, μ, γ, σ, n, VFgas, and gas cumulative volume curve control parameters a and b. Verification and validation results from sparse polynomial expansions (SPE) method and directional Gaussian smoothing (DGS) optimization show that the surrogate model had training error of ∼7% and validation error of ∼15%, indicating that machine learning methods and surrogate models can help optimize the uncertain parameters in the complex 2-phase material model. This approach is expected to fill the knowledge gap between unknown liquid-gas mixture material model and measured vessel strain responses.
Residual-based adaptive strategies are widely used in scientific machine learning yet remain largely heuristic. We introduce a variational framework that formalizes these methods through convex transformations of the residual, where different transformations correspond to distinct objective functionals. For instance, exponential weights target uniform error minimization, while linear weights recover quadratic error minimization. This perspective reveals adaptive weighting as a means of selecting sampling distributions that optimize a primal objective, directly linking discretization choices to error metrics. This principled approach yields three key benefits: it enables systematic design of adaptive schemes, reduces discretization error by lowering estimator variance, and enhances learning dynamics by improving gradient signal-to-noise ratio. Extending the framework to operator learning, we demonstrate substantial performance gains across diverse optimizers and architectures. Our results provide a theoretical perspective for residual-based adaptivity and establish a foundation for principled discretization and training.
Curtailed photovoltaic (PV) generation is a zero-marginal-cost spinning reserve that can be used for a number of active power control services. Unlike traditional spinning reserve providers, however, i.e., fossil-fueled generators, which have well-defined operating characteristics, e.g., available headroom or potential high limit (PHL), PV plants have by nature variable and uncertain operating characteristics. To ensure the effective coordination between PV plants and the system operator during an active power control event, accurate knowledge of the PV PHL is essential. It ensures that enough headroom is reserved by the PV plants to deliver the award services in real time and informs feasible dispatch decisions made by the market operator. To tackle this challenge, a novel reference-control grouping-based PV plant reserve estimation method has been proposed by the National Renewable Energy Laboratory under past projects funded by the U.S. Department of Energy Office of Energy Efficiency and Renewable Energy Solar Energy Technologies Office. The estimation method separates inverters within a plant into two groups: a control group and a reference group. While the reference group is reserved to operate at its PHL, the control group can be curtailed to provide the grid services. Real-time outputs from the reference inverters are used to estimate the PHL for the whole plant based on the ratio between capacities of the reference group and of the plant. This work further enhances the methodology by (1) improving the model accuracy through machine learning; (2) automating the reference inverter selection through correlation analysis; (3) considering estimation look-ahead windows; and (4) applying to regional spinning reserve estimation. Significant performance improvement has been observed based on real-world data collected by CAISO, Southern Company, and Terabase Energy. Compared with the original scaling method, the newly proposed machine learning-based approach reduces the estimation errors by 30% and 13% at the plant level and region level, respectively. Results obtained from this project are intended to be used by grid operators, market operators, balancing authorities, and PV plant owners and operators to facilitate PV participation in ancillary service markets. Regulators, policymakers, and system planners can also consider the results of this work in their decision-making processes. In addition to the performance improvement on the existing reference-control based grouping method, we also investigated how the variability of PV generation from a single PV inverter can be used to represent the variability of PV generation at the plant level.
The astrophysical r process plays a vital role in the production of heavy elements. Modeling of the r process is sensitive to masses and further requires knowledge of masses beyond current experimental reach. Therefore, simulations of the r process offer a unique test bed for predicting mass extrapolations. We take a Machine-Learning (ML) approach to model the masses across the entire chart of nuclides. For the first time, we simulate r-process nucleosynthesis with a physics-based ML mass model. We compare simulated abundances to solar data in order to evaluate the model's performance far from stability. The resulting r-process abundances up to thorium and uranium qualitatively match those of the observed solar system abundance pattern, with the characteristic peaks well positioned. We propagate the mass uncertainties obtained from the ML model to r-process abundance yields to estimate an uncertainty band associated with our approach. The size of the uncertainty band is approximately one order of magnitude which aligns with the uncertainty reported using alternative techniques.
To realize efficient computational fluid dynamics (CFD) prediction of two-phase flow, a multi-scale framework was proposed in this paper by applying a physics-guided data-driven approach. Instrumental to this framework, Feature Similarity Measurement (FSM) technique was developed for error estimation in two-phase flow simulation using coarse-mesh CFD, to achieve a comparable accuracy as fine-mesh simulations with fast-running feature. In this work, by defining physics-guided parameters and variable gradients as physical features, FSM has the capability to capture the underlying local patterns in the coarse-mesh CFD simulation. Massive low-fidelity data and respective high-fidelity data are used to explore the underlying information relevant to the main simulation errors and the effects of phenomenological scaling. By learning from previous simulation data, a surrogate model using deep feedforward neural network (DFNN) can be developed and trained to estimate the simulation error of coarse-mesh CFD. In a demonstration case of two-phase bubbly flow, the DFNN model well captured and corrected the unphysical “peaks” in the velocity and void fraction profiles near the wall in the coarse-mesh configuration, even for extrapolative predictions. The research documented supports the feasibility of the physics-guided deep learning methods for coarse mesh CFD simulations which has a potential for the efficient industrial design.
Powder metallurgy (PM)–hot isostatic pressing (PM-HIP) has long been recognized as a powerful route for producing fully dense, near net shape metallic components. By consolidating powders under high temperature and pressure, HIP provides isotropic properties, uniform microstructures, and scalability to complex geometries that are vital for sectors such as aerospace, energy, and nuclear power. Yet despite these advantages, the technology has remained constrained by costly trial and error canister fabrication, limitations of conventional forging, and incomplete knowledge about how the canister design influences final part properties. Additive manufacturing (AM), by contrast, thrives on design freedom and geometric flexibility but struggles with speed, scalability, and cost when applied to very large structures. The research presented in this report investigated how a convergent manufacturing approach, combining AM with PM-HIP, can merge the strengths of both technologies, leveraging AM’s flexibility for canister design and HIP’s consolidation capability to deliver reliable, large, and complex parts. The work progressed through three case studies that built on one another in scale and complexity. Small cylindrical canisters fabricated by conventional methods, laser powder bed fusion, and directed energy deposition were filled with stainless steel powders and subjected to HIP. The resulting parts demonstrated near-full density and mechanical properties on par with wrought stainless steel, showing for the first time that AM canisters can be a direct substitute for conventional ones without sacrificing quality. The next step involved a medium-scale, noncentrosymmetric T-valve, which is an enclosed, multibranch geometry that tested the limits of AM + PM-HIP integration. The T-valve achieved predictable shrinkage and uniform densification, confirming feasibility for enclosed designs. However, this study also revealed oxide inclusions and interfacial challenges at the AM + HIP boundary, underscoring the critical importance of controlling interface chemistry and employing robust, in situ strategies, such as melt pool monitoring and thermal monitoring, coupled with nondestructive evaluation techniques such as x-ray computed tomography. Finally, the effort culminated in fabricating a large-scale impeller weighing nearly 2000 lb and spanning 5 ft in diameter. Produced via multirobot wire arc AM and hot isostatic pressed to near-full density, the impeller validated industrial-scale feasibility. Predictive models closely matched experimental shrinkage, tensile properties were spatially uniform across the component, and the AM + PM-HIP interface proved mechanically sound despite the presence of oxide-decorated prior particle boundaries. This large-scale demonstration is a major milestone, showing that hybrid AM + PM‑HIP can reliably deliver components at reactor-relevant scales. Collectively, these studies charted a logical pathway: small-scale work built scientific confidence, medium-scale work highlighted opportunities and challenges, and large-scale work proved industrial impact. The overarching conclusion of this report is that AM + PM-HIP should not be seen as a replacement for forging but as a complementary pathway that provides the US with flexibility, resilience, and new options for manufacturing nuclear-grade components. Looking ahead, several directions emerge as critical to sustaining progress. Predictive modeling must become faster, more accessible, and more accurate, with digital twins and machine learning reducing reliance on trial and error. Powders and alloys must be optimized for HIP, with improved cleanliness, reduced oxides, and tailored chemistries that enhance creep, fatigue, and irradiation resistance. Interfaces between AM and HIP regions must be better engineered through coatings, machining strategies, and surface treatments to mitigate oxide formation and ensure reliable bonding to explore opportunities for HIP of targeted compositional parts, as well as multimaterial HIP cladding applications. Monitoring and nondestructive evaluation need to expand, incorporating multimodal sensors, x-ray computed tomography, and real-time data integration through platforms such as Pelican. At the same time, the pathway to industrial adoption requires techno-economic analysis, machinability studies, and qualification frameworks aligned with industry and regulatory standards. Finally, workforce and academic engagement must be strengthened. Programs that train technicians and engineers for US Navy and US Department of Energy manufacturing challenges should be paired with academic partnerships to support fundamental research, with open sharing of non-export-controlled data to accelerate innovation and build the next generation of experts. In conclusion, this report demonstrates that hybrid AM + PM-HIP is scientifically viable and strategically important. By combining the design agility of AM with the consolidation strength of HIP and embedding modeling, monitoring, and workforce development, this approach provided a transformative new capability for US manufacturing. The path forward is clear: hybrid AM + PM-HIP is not just a promising research direction but is also potentially an industrially relevant pathway that can reshape how nuclear-grade components are designed, qualified, and deployed.
Any system with a human component is inherently risky. Studies in human factors and psychology have repeatedly shown that human operators will inevitably make errors, regardless of how well they are trained. Onboard the International Space Station (ISS) where crew time is arguably the most valuable resource, errors by the crew or ground operators can be costly to critical science objectives. Operations experts at the ISS Payload Operations Integration Center (POIC), located at NASA's Marshall Space Flight Center in Huntsville, Alabama, have learned that from payload concept development through execution, there are countless opportunities to introduce errors that can potentially result in costly losses of crew time and science. To effectively address this challenge, we must approach the design, testing, and operation processes with two specific goals in mind. First, a systematic approach to error and human centered design methodology should be implemented to minimize opportunities for user error. Second, we must assume that human errors will be made and enable rapid identification and recoverability when they occur. While a systematic approach and human centered development process can go a long way toward eliminating error, the complete exclusion of operator error is not a reasonable expectation. The ISS environment in particular poses challenging conditions, especially for flight controllers and astronauts. Operating a scientific laboratory 250 miles above the Earth is a complicated and dangerous task with high stakes and a steep learning curve. While human error is a reality that may never be fully eliminated, smart implementation of carefully chosen tools and techniques can go a long way toward minimizing risk and increasing the efficiency of NASA's space science operations.
We explore using neural operators, or neural network representations of nonlinear maps between function spaces, to accelerate infinite-dimensional Bayesian inverse problems (BIPs) with models governed by nonlinear parametric partial differential equations (PDEs). Neural operators have gained significant attention in recent years for their ability to approximate the parameter-to-solution maps defined by PDEs using as training data solutions of PDEs at a limited number of parameter samples. The computational cost of BIPs can be drastically reduced if the large number of PDE solves required for posterior characterization are replaced with evaluations of trained neural operators. However, reducing error in the resulting BIP solutions via reducing the approximation error of the neural operators in training can be challenging and unreliable. We provide an a priori error bound result that implies certain BIPs can be ill-conditioned to the approximation error of neural operators, thus leading to inaccessible accuracy requirements in training. To reliably deploy neural operators in BIPs, we consider a strategy for enhancing the performance of neural operators: correcting the prediction of a trained neural operator by solving a linear variational problem based on the PDE residual. We show that a trained neural operator with error correction can achieve a quadratic reduction of its approximation error, all while retaining substantial computational speedups of posterior sampling when models are governed by highly nonlinear PDEs. The strategy is applied to two numerical examples of BIPs based on a nonlinear reaction–diffusion problem and deformation of hyperelastic materials. We demonstrate that posterior representations of the two BIPs produced using trained neural operators are greatly and consistently enhanced by error correction.
This paper presents the first converged frequency-dependent HF polarizability results for general molecules, on a set of 89 closed-shell atoms and molecules. The solver employs multiresolution analysis (MRA) in a multiwavelet basis to compute both ground and response states to a guaranteed precision, which are validated against independent numerical grid calculations on atoms and linear molecules. The MRA ground-state energies and response properties are used to evaluate results in correlation-consistent basis sets up to 5Z augmented with either single or double diffuse functions and core-polarization functions. Systematic trends are revealed through consideration of chemical composition as well as the use of machine learning to cluster convergence trends, the latter suggesting the possibility of learning and correcting basis-set error.
Here, we build upon recent work on the use of machine-learning models to estimate Hamiltonian parameters using continuous weak measurement of qubits as input. We consider two settings for the training of our model: (1) supervised learning, where the weak-measurement training record can be labeled with known Hamiltonian parameters, and (2) unsupervised learning, where no labels are available. The first has the advantage of not requiring an explicit representation of the quantum state, thus potentially scaling very favorably to a larger number of qubits. The second requires the implementation of a physical model to map the Hamiltonian parameters to a measurement record, which we implement using an integrator of the physical model with a recurrent neural network to provide a model-free correction at every time step to account for small effects not captured by the physical model. We test our construction on a system of two qubits and demonstrate accurate prediction of multiple physical parameters in both the supervised context and the unsupervised context. We demonstrate that the model benefits from larger training sets, establishing that it is “learning,” and we show robustness regarding errors in the assumed physical model by achieving accurate parameter estimation in the presence of unanticipated single-particle relaxation.
Not provided.
Radial basis functions (RBFs) are a powerful tool for constructing high-order accurate reduced representations of scattered data in arbitrary dimension and on manifolds. We present a method of constructing data approximations in which we utilize a functional tail to capture a global background profile and a RBF neural network (NN) to capture the smaller-scale features. In the RBF NN the RBF centers, matrix shape parameters were selected adaptively for each RBF. We also utilized a geodesic notion of distance on the manifold on which the data lies, e.g., the spherical geodesic for data on the sphere. Although each of these ideas have been been investigated separately in previous works, their combination into a single algorithm is novel. We defined a machine learning problem in which these properties are learned to minimize the data reduction error. We demonstrate the algorithm for applications of scattered data reduction in the plane and on the sphere.
Statistical communication theory for learning and adaptive system programs, and signal design