Robust Multi-fidelity Bayesian Optimization with Deep Kernel and Partition
Explore the source record for details and available documents.
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.
Explore the source record for details and available documents.
As machine learning models continue to increase in complexity, so does the potential number of free model parameters commonly known as hyperparameters. While there has been considerable progress toward finding optimal configurations of these hyperparameters, many optimization procedures are treated as black boxes. We believe optimization methods should not only return a set of optimized hyperparameters, but also give insight into the effects of model hyperparameter settings. To this end, we present HyperSpace, a parallel implementation of Bayesian sequential model-based optimization. HyperSpace leverages high performance computing (HPC) resources to better understand unknown, potentially non-convex hyperparameter search spaces. We show that it is possible to learn the dependencies between model hyperparameters through the optimization process. By partitioning large search spaces and running many optimization procedures in parallel, we also show that it is possible to discover families of good hyperparameter settings over a variety of models including unsupervised clustering, regression, and classification tasks.
There has been an increase in extreme weather events, posing a threat to power grid systems, potentially influenced by factors such as population growth, changes in ecosystems, land cover, and land use in the service area, as well as the growth of certain vegetation types. This research seeks to develop a predictive model to mitigate potential damages caused by future winter storms. This research utilizes the Light Gradient Boosting Machine (LightGBM), incorporating the number of power outages experienced at the county level, geographic details, weather information, and lagged outage and lagged weather data. The developed models were broadly divided into two groups, with six models in each group - one group without optimization and another with optimization, totaling 12 trained models. For model optimization, Bayesian optimization was employed using Root Mean Squared Error (RMSE) as the objective function. In results, when comparing Group 2 (the optimized group) with Group 1 (the non-optimized group), it was found that optimization did not always lead to a reduction in RMSE and Mean Absolute Error (MAE). However, in terms of Mean Directional Accuracy (MDA), while all results in Group 1 were below the baseline accuracy of 0.33, all results in Group 2 exceeded 0.33, with some cases showing an increase of more than three times the baseline. The results indicated that, in the optimized model group, Population and Pressure were the most influential factors when using current weather data and geographical information. When using lagged data, lagged recorded outages and lagged Pressure emerged as the most significant factors. Among the 12 developed models, the L-1-2-O model showed the lowest RMSE and MAE, as well as the highest accuracy, with values of 390.62 households and 168.13 households, respectively. To normalize the RMSE and MAE values, each metric was divided by the average number of households among the counties in Texas. For the L-1-2-O model, the scaled RMSE was 0.88% and the scaled MAE was 0.38%. In terms of MDA, which indicates the accuracy of the prediction direction, the L-1-O model achieved the highest score of 0.41. Although this study focused on Texas, which suffered the greatest impact from the winter storms in 2021, with additional validation, the methodology used in this research could be applied to other regions.
Bayesian optimization has shown promise for the design optimization of inertial confinement fusion targets. Specifically, in Vazirani et al. [Phys. Plasmas 28 , 122709 (2021)], optimal designs for double shell capsules with graded inner shells were identified using one-dimensional xRAGE simulation yield calculations. While the machine learning models were able to accurately learn and predict one-dimensional simulation target performance, using simulations with higher fidelity would improve design optimization and better match with the expected experimental performance. However, higher fidelity physics modeling, i.e., two-dimensional xRAGE simulations, requires significantly larger computational time/cost, usually at least an order of magnitude, in comparison with one-dimensional simulations. This study presents a multi-fidelity Bayesian optimization, in which the machine learning model leverages low-fidelity (one-dimensional xRAGE) and high-fidelity (two-dimensional xRAGE) simulations to more accurately predict “pre-shot” target performance with respect to the expected experimental performance. By building a multi-fidelity Bayesian optimization framework coupled with xRAGE, the low-fidelity and high-fidelity simulations are able to inform one another, such that we have: (1) improved physics modeling in comparison with using low-fidelity simulations alone, (2) reduced computational time/cost in comparison with using high-fidelity simulations alone, and (3) more confidence in the expected performance of optimized targets during real-world experiments. In the future, we plan to use this robust multi-fidelity Bayesian optimization methodology to expedite the design of graded inner shells further and eventually full capsules as a part of the current double shell campaign at the National Ignition Facility.
The increasing integration of renewable energy and rising electricity demand highlight the importance of battery energy storage systems for peak shaving and demand response. Unlike prior approaches that overlook operational impacts on degradation, this study proposes a Bayesian Optimization–Mixed Integer Linear Programming framework for optimal battery energy storage system sizing. In this framework, Mixed Integer Linear Programming determines short-term scheduling while a calibrated electrochemical model iteratively evaluates degradation. The central hypothesis is that the framework can efficiently identify optimal sizes that yield realistic and economically robust outcomes. The method is tested across three scenarios: peak shaving, peak shaving with energy-reduction demand response, and peak shaving with power-reduction demand response. Results show that the framework converge to the optimum within 20 iterations out of 150 possible sizes. Under baseline conditions, the framework consistently selects the smallest feasible system, minimizing unnecessary degradation costs from oversized storage. Sensitivity analyses reveal that larger systems are favored as demand rates or incentives increase. Comparisons of demand response programs indicate that power-reduction demand response offers greater economic benefits than energy-reduction demand response, although demand savings from peak shaving remain the dominant contributor to overall performance. This study demonstrates that the proposed framework balances computational tractability with degradation fidelity, identifies critical economic thresholds for investment, and offers a practical, flexible tool to guide industrial stakeholders in cost-effective battery energy storage system deployment.
DeepOpt is a simple and easy-to-use library for performing Bayesian optimization, leveraging the powerful capabilities of BoTorch. Its key feature is the ability to use neural networks as surrogate functions during the optimization process, allowing Bayesian optimization to work smoothly even on large datasets and in many dimensions. DeepOpt also provides simplified wrappers for BoTorch fitting and optimization routines.
This work introduces automated machine learning workflows that address critical bottlenecks in surrogate model development for Ion Cyclotron Range of Frequencies (ICRF) heating applications. The automated framework includes data analysis tools that transform raw datasets into actionable insights in seconds, replacing weeks of manual exploratory effort and ensuring consistent, reproducible dataset characterization. By integrating advanced hyperparameter optimization (HPO) methods including Bayesian optimization via BoTorch and Tree-structured Parzen Estimators (TPE), the framework significantly reduces model development time from weeks to hours, decreasing computational cost and required expertise, while enabling high-accuracy surrogate models. Compared to traditional hyperparameter scanning (HPS) techniques such as methodical, randomized, and grid searches, HPO methods achieve superior convergence and predictive performance, even when compared to already well-tuned reference models. On NSTX High Harmonic Fast Wave (HHFW) heating datasets, both Random Forest Regressor (RFR) and neural network surrogates demonstrate improved accuracy, achieving R 2 values beyond 0.97 and 0.98, respectively. The results show that while HPO gains are modest for robust architectures like RFR, they become essential for more sensitive models such as neural networks, highlighting the trade-offs across optimization strategies. Through automated workflows that eliminate manual hyperparameter tuning and require minimal ML expertise, this work enables widespread adoption of high-fidelity surrogate models across the fusion community for real-time plasma control, uncertainty quantification, rapid experimental scenario development, and integrated system optimization.
Access to reliable, clean energy sources is a major concern for national security. Much research is focused on the “grand challenge” of producing energy via controlled fusion reactions in a laboratory setting. For fusion experiments, specifically inertial confinement fusion (ICF), to produce sufficient energy, the fusion reactions in the ICF fuel need to become self-sustaining and burn deuterium-tritium (DT) fuel efficiently. The recent record-breaking NIF ignition shot was able to achieve this goal as well as produce more energy than used to drive the experiment. This achievement brings self-sustaining fusion-based power systems closer than ever before, capable of providing humans with access to secure, renewable energy. In order to further progress toward the actualization of such power systems, more ICF experiments need to be conducted at large laser facilities such as the United States's National Ignition Facility (NIF) or France's Laser Mega-Joule. The high cost per shot and limited number of shots that are possible per year make it prohibitive to perform large numbers of experiments. As such, experimental design relies heavily on complex predictive physics simulations for high-fidelity “preshot” analysis. These multidimensional, multi-physics, high-fidelity simulations have to account for a variety of input parameters as well as modeling the extreme conditions (pressures and densities) present at ignition. Such simulations (especially in 3D) can become computationally prohibitive to turn around for each ICF experiment. In this work, we explore using Bayesian optimization with Gaussian processes (GPs) to find optimal designs for ICF double shell targets, while keeping computational costs to manageable levels. These double shell targets have an inner shell that grades from beryllium on the outer surface to the higher Z material molybdenum, as opposed to the nominally used tungsten, on the inside in order to trade off between the high performance associated with high density inner shells and capsule stability. We describe our results for “capsule-only” xRAGE simulations to study the physics between different capsule designs, inner shell materials, and potential for future experiments.
The ever increasing computational cost of Deep Neural Networks (DNN) and the demand for energy efficient hardware for DNN acceleration has made accuracy and hardware cost co-optimization for DNNs tremendously important, especially for edge devices. Owing to the large parameter space and cost of evaluating each parameter in the search space, manually tuning of DNN hyperparameters is impractical. Automatic joint DNN and hardware hyperparameter optimization is indispensable for such problems. Bayesian optimization-based approaches have shown promising results for hyperparameter optimization of DNNs. However, most of these techniques have been developed without considering the underlying hardware, thereby leading to inefficient designs. Further, the few works that perform joint optimization are not generalizable and mainly focus on CMOS-based architectures. In this work, we present a novel pseudo agent-based multiobjective hyperparameter optimization (PABO) for maximizing the DNN performance while obtaining low hardware cost. Compared to the existing methods, our work poses a theoretically different approach for joint optimization of accuracy and hardware cost and focuses on memristive crossbar based accelerators. PABO uses a supervisor agent to establish connections between the posterior Gaussian distribution models of network accuracy and hardware cost requirements. The agent reduces the mathematical complexity of the co-optimization problem by removing unnecessary computations and updates of acquisition functions, thereby achieving significant speed-ups for the optimization procedure. PABO outputs a Pareto frontier that underscores the trade-offs between designing high-accuracy and hardware efficiency. Our results demonstrate a superior performance compared to the state-of-the-art methods both in terms of accuracy and computational speed (~100x speed up).
Herein we present a new modeling paradigm for optimization that we call random field optimization. Random fields are a powerful modeling abstraction that aims to capture the behavior of random variables that live on infinite-dimensional spaces (e.g., space and time) such as stochastic processes (e.g., time series, Gaussian processes, and Markov processes), random matrices, and random spatial fields. This paradigm involves sophisticated mathematical objects (e.g., stochastic differential equations and space-time kernel functions) and has been widely used in neuroscience, geoscience, physics, civil engineering, and computer graphics. Despite of this, however, random fields have seen limited use in optimization; specifically, existing optimization paradigms that involve uncertainty (e.g., stochastic programming and robust optimization) mostly focus on the use of finite random variables. This trend is rapidly changing with the advent of statistical optimization (e.g., Bayesian optimization) and multi-scale optimization (e.g., integration of molecular sciences and process engineering). Our work extends a recently-proposed abstraction for infinite-dimensional optimization problems by capturing more general uncertainty representations. Moreover, we discuss solution paradigms for this new class of problems based on finite transformations and sampling, and identify open questions and challenges.
optimas is a Python library that can coordinate a large number of numerical simulations on high-performance computing resources, with the aim of optimizing a given simulation result. More specifically, this library is currently used for design optimization of laser-plasma particle accelerators. In this context, the performance of particle accelerator designs are often evaluated with large-scale simulation codes, and many separate simulations need to be run - with different design parameters - in order to find the most performant design. optimas facilitates this process by providing a convenient interface to advanced optimizers (e.g. Bayesian optimization), and by coordinating the execution of the different simulations on HPC resources. (This is done by leveraging the library libensemble.) Compared to other open-source optimization libraries (e.g. Ax), optimas is more tailored towards execution on DOE HPC resources (e.g. Perlmutter, Summit, etc.) and is specialized for the type of simulation codes and workflows that are used in the community of laser-plasma acceleration.
The critical task of inferring anomalous cross-field transport coefficients is addressed in simulations of boundary plasmas with fluid models. A workflow for parameter inference in the UEDGE fluid code is developed using Bayesian optimization with parallelized sampling and integrated uncertainty quantification. In this workflow, transport coefficients are inferred by maximizing their posterior probability distribution, which is generally multidimensional and non-Gaussian. Uncertainty quantification is integrated throughout the optimization within the Bayesian framework that combines diagnostic uncertainties and model limitations. As a concrete example, we infer the anomalous electron thermal diffusivity $\chi_\perp$ from an interpretive 2D model describing electron heat transport in the conduction-limited region with radiative power loss. The workflow is first benchmarked against synthetic data and then tested on H-, L-, and I-mode discharges to match their midplane temperature and divertor heat flux profiles. We demonstrate that the workflow efficiently infers diffusivity and its associated uncertainty, generating 2D profiles that match 1D measurements. Future efforts will focus on incorporating more complicated fluid models and analyzing transport coefficients inferred from a large database of experimental results.
Adjoint-based optimization methods are attractive for aerodynamic shape design primarily due to their computational costs being independent of the dimensionality of the input space and their ability to generate high-fidelity gradients that can then be used in a gradient-based optimizer. This makes them very well suited for high-fidelity simulation based aerodynamic shape optimization of highly parametrized geometries such as aircraft wings. However, the development of adjoint-based solvers involve careful mathematical treatment and their implementation require detailed software development. Furthermore, they can become prohibitively expensive when multiple optimization problems are being solved, each requiring multiple restarts to circumvent local optima. In this work, we propose a machine learning enabled, surrogate-based framework that replaces the expensive adjoint solver, without compromising on predicting predictive accuracy. Specifically, we first train a deep neural network (DNN) from training data generated from evaluating the high-fidelity simulation model on a model-agnostic design of experiments on the geometry shape parameters. The optimum shape may then be computed by using a gradient-based optimizer coupled with the trained DNN. Subsequently, we also perform a gradient-free Bayesian optimization, where the trained DNN is used as the prior mean. We observe that the latter framework (DNN-BO) improves upon the DNN-only based optimization strategy for the same computational cost. Overall, this framework predicts the true optimum with very high accuracy, while requiring far fewer high-fidelity function calls compared to the adjoint-based method. Furthermore, we show that multiple optimization problems can be solved with the same machine learning model with high accuracy, to amortize the offline costs associated with constructing our models. Our methodology finds applications in the early stages of aerospace design. (C) 2021 Published by Elsevier Masson SAS.
Computer experiments often employ space-filling strategies to create surrogate models with strong predictive performance. The impact of model parameter estimation for Gaussian process surrogates, however, is often overlooked. Obtaining a better initial estimate of the covariance lengthscale parameter, θ, can greatly improve the resulting Gaussian process fit through more effective sequential acquisitions during active learning. In this work, we propose a novel initial design maximizing the Bayesian D-optimality criterion of the Gaussian process lengthscale parameter. Previously published results have shown the emphasis on lengthscale estimation to be promising, but relied on an empirically driven design creation process. Our Bayesian D-optimal designs are rooted in information theory and lead to more informative sequential acquisitions by improving lengthscale estimation. In many cases, these gains eventually result in better surrogates than those seeded with space-filling initial designs. Furthermore, Bayesian D-optimal designs can be tailored to either isotropic or anisotropic covariance structures, and the Bayesian framework enables the inclusion of prior knowledge in the design process, offering greater flexibility and adaptability. Through several simulation studies, we demonstrate the advantages of Bayesian D-optimal designs in terms of both lengthscale estimation accuracy and predictive performance during active learning.
Neuromorphic computing is emerging as a promising Beyond Moore computing paradigm that employs event-triggered computation and non-von Neumann hardware. Spike Timing Dependent Plasticity (STDP) is a well-known bio-inspired learning rule that relies on activities of locally connected neurons to adjust the weights of their respective synapses. In this work, we analyze a basic STDP rule and its sensitivity on the different hyperparameters for training spiking neural networks (SNNs) with supervision, customized for a neuromorphic hardware implementation with integer weights. We compare the classification performance on four UCI datasets (iris, wine, breast cancer and digits) that depict varying levels of complexity. We perform a search for optimal set of hyperparameters using both grid search and Bayesian optimization. Through the use of Bayesian optimization, we show the general trends in hyperparameter sensitivity in SNN classification problem. With the best sets of hyperparameters, we achieve accuracies comparable to some of the best performing SNNs on these four datasets. With a highly optimized supervised STDP rule we show that these accuracies can be achieved with just 20 epochs of training.
This project develops a Bayesian optimization approach to extracting insights from Z Machine experimental data to determine if and how these insights can be used to extrapolate to a larger facility. The primary goal is to address the scientific challenge of informing how confidently experimental conditions can be predicted on a next generation facility, the design of which requires the reliable extrapolation of current high energy density technologies to regimes yet unobserved, except by costly high-fidelity computational models. Maximizing the use of presently available data and understanding how it informs future endeavors is critically important to enable transformative pulsed power and the science of extreme conditions. We explore a Bayesian optimization approach to experimental design which combines information theory, experimental data, and computational modeling to explore how information gain can be maximized.
The SAGUARO (Scalable Algorithms for Groundwater Uncertainty Analysis and Robust Optimization) Project focused on the development of scalable numerical algorithms for large-scale Bayesian inversion in complex systems that capitalize on advances in large-scale simulation-based optimization and inversion methods. The project was a collaborative effort among MIT, the University of Texas at Austin, Georgia Institute of Technology, and Sandia National Laboratories. The research was directed in three complementary areas: efficient approximations of the Hessian operator, reductions in complexity of forward simulations via stochastic spectral approximations and model reduction, and employing large-scale optimization concepts to accelerate sampling. The MIT₋Sandia component of the SAGUARO Project addressed the intractability of conventional sampling methods for large-scale statistical inverse problems by devising reduced-order models that are faithful to the full-order model over a wide range of parameter values; sampling then employs the reduced model rather than the full model, resulting in very large computational savings. Results indicate little effect on the computed posterior distribution. On the other hand, in the Texas₋Georgia Tech component of the project, we retain the full-order model, but exploit inverse problem structure (adjoint-based gradients and partial Hessian information of the parameter-to-observation map) to implicitly extract lower dimensional information on the posterior distribution; this greatly speeds up sampling methods, so that fewer sampling points are needed. We can think of these two approaches as "reduce then sample" and "sample then reduce." In fact, these two approaches are complementary, and can be used in conjunction with each other. Moreover, they both exploit deterministic inverse problem structure, in the form of adjoint-based gradient and Hessian information of the underlying parameter-to-observation map, to achieve their speedups.
ABSTRACT Vertical hetero-structures made from stacked monolayers of transition metal dichalcogenides (TMDC) are promising candidates for next-generation optoelectronic and thermoelectric devices. Identification of optimal layered materials for these applications requires the calculation of several physical properties, including electronic band structure and thermal transport coefficients. However, exhaustive screening of the material structure space using ab initio calculations is currently outside the bounds of existing computational resources. Furthermore, the functional form of how the physical properties relate to the structure is unknown, making gradient-based optimization unsuitable. Here, we present a model based on the Bayesian optimization technique to optimize layered TMDC hetero-structures, performing a minimal number of structure calculations. We use the electronic band gap and thermoelectric figure of merit as representative physical properties for optimization. The electronic band structure calculations were performed within the Materials Project framework, while thermoelectric properties were computed with BoltzTraP. With high probability, the Bayesian optimization process is able to discover the optimal hetero-structure after evaluation of only ~20% of all possible 3-layered structures. In addition, we have used a Gaussian regression model to predict not only the band gap but also the valence band maximum and conduction band minimum energies as a function of the momentum.