Using Bayesian deep learning approaches for uncertainty-aware building energy surrogate models

ATC provides a systematic approach in solving decomposed large scale systems that has solvable subsystems. However， complex engineering system usually has a high computational cost , which result in limiting real-life applications of ATC based on high-fidelity simulation models. To address these problems, this paper aims to develop an efficient approximation model building techniques under the analytical target cascading (ATC) framework, to reduce computational cost associated with multidisciplinary design optimization problems based on high-fidelity simulations. An approximation model building techniques is proposed: approximations in the subsystem level are based on variable-fidelity modeling (interaction of low- and high-fidelity models). The variable-fidelity modeling consists of computationally efficient simplified models (low-fidelity) and expensive detailed (high-fidelity) models. The effectiveness of the method for modeling under the ATC framework using variable-fidelity models is studied. Overall results show the methods introduced in this paper provide an effective way of improving computational efficiency of the ATC method based on variable-fidelity simulation models.

It is widely accepted that the concentration of people living in high-density city centers offers greater operational energy efficiency and lower greenhouse gas emissions than lower-density expanded suburbs. The prevailing assumption is that lower-density suburbs are dominated by larger low-rise buildings that have higher building energy use requirements and greater per-person automobile travel requirements than high-density city centers dominated by medium- and high-rise buildings located in close proximity to a variety of public transit systems. However, very few studies to date have utilized empirical data at an individual household scale to evaluate differences in the operational energy (OE) footprints for both building and transportation energy end-uses between high-rise urban and low-rise suburban households. Therefore, this work collects empirical data on building and transportation OE consumption by individuals and households living in two economically similar groups: existing high-rise residential buildings in downtown Chicago, IL, USA and existing low-rise residential buildings in suburban Oak Park, IL, USA. Data were collected from over 500 households via an online survey. We considered the following components of residential living: (1) building OE (BOE), which includes electricity and/or natural gas use for all building energy end-uses; and (2) transportation OE (TOE), which includes the OE for multiple modes of transportation (i.e., automobile, bus, subway, regional train, etc.) based on average travel behavior in each location, as well as the OE for supporting transportation infrastructure. We estimate that downtown high-rise living in this sample of residences in Chicago, IL accounts for approximately 427 GJ of primary OE per household per year, on average, which was 14% lower than the average for suburban low-rise living in the Oak Park, IL homes (499 GJ per household per year). However, on a per-person basis, downtown high-rise living accounts for approximately 246 GJ of primary OE per person per year, which was approximately 61% higher than suburban low-rise living (153 GJ per person per year). In both building types, building OE was the single largest contributor to total OE use. This study accurately captured the energy requirements associated with realistic behaviors and lifestyles of occupants of both low-rise suburban and high-rise urban households, and found that building OE dominates the total OE, which suggests that efforts to reduce building OE should be given high priority in building design and management as well as urban planning.

In the community of petroleum engineering, the use of surrogate modelling techniques have recently gained more and more popularity to improve the efficiency of history matching. However, it is still not possible to fully utilize their potential in realistic applications. One of the challenges is to retain high accuracy while increasing the computational efficiency using a surrogate model. This dissertation proposed a projection-based reduced-order model and a data-driven deep convolutional neural network. In the first part of the thesis, a non-intrusive subdomain POD-TPWL method for solving gradient-based reservoir history matching problems is presented. It is a projection-based reduced-order modelling approach wherein the adjoint model of the original high-dimensional non-linear model is approximated by a subdomain reduced-order linear model. Furthermore, by introducing domain decomposition for the reduced-order model and by restricting the number of uncertain parameter patterns to the subdomains, the number of full order simulations required for the derivation of this surrogate model is reduced drastically. In the second part of the thesis, we propose two kinds of deep-learning inversion frameworks for efficiently solving large-scale history matching problems. The first deep-learning deterministic inversion framework primarily explores the possibility of applying a DNN surrogate to approximate the gradient of the objective function by making use of auto-differentiation (AD). In combination with the DNN surrogate, the AD enables us to evaluate the gradients efficiently in a parallel manner and without the need of explicitly coding of the adjoint model. The second framework is the deep-learning stochastic inversion which constructs a deep-learning surrogate based on an image-oriented distance parameterization for ensemble-based seismic history matching. Instead of directly assimilating spatially dense seismic data, image-oriented distance parameterization is employed to extract valuable information from the water fronts. Inspired by the methodologies developed for image segmentation in the field of computer vision and image processing, we propose an advanced image segmentation network for accurately predicting water fronts with highly-complex spatial discontinuities. In comparison with the conventional workflows entirely based on high-fidelity simulation models, experimental results show that the proposed surrogate-supported workflow achieves an accuracy equal to or better than the conventional workflow at significantly lower cost.

Device sizing is a challenging problem for analog circuit design. Traditional methods depend on domain knowledge and intensive simulations to search for feasible parameters. Recent studies apply the Bayesian optimization (BO) and a Gaussian process (GP) model in analog circuit synthesis to improve efficiency. The BO framework automatically selects the parameter candidates by inferring the surrogate GP model. However, naive BO employs a sequential updating strategy which is inefficient in a multicore environment. Besides, the widely used GP model requires costly high fidelity data, which are obtained from fine simulations. In this article, we propose a constrained batch BO approach with a multifidelity (MF) model to solve the above difficulties. The batch BO exploits parallel computing and selects promising parameters by multiple acquisition function ensemble. In addition, the MF GP model adapts the low fidelity data obtained from coarse simulations. Specifically, the proposed method incorporates information gain in a weighted clustering algorithm to refine the parameter candidates. As a result, the proposed method maintains the candidates’ quality and diversity, which speeds up the optimization convergence. In the experiments, we demonstrate the efficiency of the proposed approach on three real-world circuits. The results show that our approach reduces the simulation costs by at least 54.6% compared to the state-of-the-art baselines.

Research on multi-fidelity aerodynamic optimization methods

A Randomized Control Trial Focused on Breast Core Needle Biopsy Skill Acquisition Using High Fidelity vs Low Fidelity Simulation Models in Rwanda: A Baseline Analysis

Robotic patients show great potential for helping to improve medical palpation training, as they can provide feedback that cannot be obtained in a real patient. They provide information about internal organ deformation that can significantly enhance palpation training by giving medical trainees visual insight based on the pressure they apply for palpation. This can be achieved by using computational models of abdomen mechanics. However, such models are computationally expensive, and thus unable to provide real-time predictions. In this work, we proposed an innovative surrogate model of abdomen mechanics by using machine learning (ML) and finite element (FE) modelling to virtually render internal tissue deformation in real time. We first developed a new high-fidelity FE model of the abdomen mechanics from computerized tomography (CT) images. We performed palpation simulations to produce a large database of stress distribution on the liver edge, an area of interest in most examinations. We then used artificial neural networks (ANNs) to develop the surrogate model and demonstrated its application in an experimental palpation platform. Our FE simulations took 1.5 h to predict stress distribution for each palpation while this only took a fraction of a second for the surrogate model. Our results show that our artificial neural network (ANN) surrogate has an accuracy of 92.6%. We also showed that the surrogate model is able to use the experimental input of palpation location and force to provide real-time projections onto the robotics platform. This enhanced robotics platform has the potential to be used as a training simulator for trainees to hone their palpation skills.

The number of aerodynamic analysis in the design optimization of propeller has been reduced significantly by using the surrogate model, such as kriging. In this paper, a more efficient aerodynamic design optimization method of propeller is proposed by using the HK (hierarchical kriging) model. The high-fidelity model is defined as a RANS (Reynolds- Averaged Navier-Stokes) simulation for propeller and the low-fidelity model is defined as the results by blade-element/vortex theory. Initial samples are selected for two levels of fidelity via LHS (Latin Hypercube Sampling). The aerodynamic performance of varying fidelity is used to build variable-fidelity surrogate models for functions of objective (e.g. thrust) and constraint (e.g. shaft power). Infill sampling criteria, including MSP (minimizing of surrogate prediction), EI (expected improvement), PI (probability of improvement), LCB (lower-confidence bounding) and MSE (mean-squared error) are used to obtain new samples, and the surrogate models are repetitively updated until a global optimum is found. High-altitude propellers are optimized by the kriging model and hierarchical kriging model, respectively. Compared to kriging model, the number of RANS solving by the hierarchical kriging model is reduced 37.5%, and the optimization time is reduced 24.3%. The results have shown that the proposed design method for propeller can significantly improve the optimization efficiency.

Urbanization and the increasing demand for accommodation and activities have led to significant energy consumption and CO2 emissions in building construction. This thesis investigates urban building energy performance (BEP) through a multi-faceted approach integrating data-driven methods and physical modeling to enhance energy efficiency while considering environmental impacts and climate change. This thesis is structured with four major objectives: (1) identifying the driving forces behind BEP; (2) predicting and optimizing urban building energy usage with geographical information and urban heat island (UHI) effects; (3) developing a Gradient-Weighted LightGBM (GW-LGBM) model to evaluate BEP considering urban morphology; and (4) conducting an optimization process to identify optimal solutions for improving BEP and thermal comfort under climate change with a hybrid method that incorporates physical modeling in Grasshopper and a data-driven approach. Key findings can be summarized: (1) the geographically and temporally weighted regression (GTWR) model significantly improves the prediction of building energy performance by capturing spatio-temporal heterogeneity, identifying key factors such as the number of floors and energy star ratings, and offering a reliable tool for sustainable urban planning; (2) integrating urban morphology and building geometry significantly enhances the accuracy of predicting building energy consumption and GHG emissions, with total gross floor area and natural gas usage identified as the most influential factors, demonstrating the superiority of the LightGBM model over traditional methods; (3) integrating a geographically weighted Light Gradient Boosting Machine (GW-LGBM) model with multi-objective optimization significantly improves the prediction accuracy of building energy performance and GHG emissions, identifying natural gas usage, total gross floor area, Energy Star score, and shape form as the most influential factors, achieving up to 89.91% improvement in energy optimization scenarios. (4) a novel framework combining physical simulation modeling, explainable machine learning, and multi-objective optimization to predict and optimize BEP considering urban heat island effects and climate change scenarios. The results highlight significant improvements in energy use intensity and indoor thermal comfort, with the hybrid Bayesian optimization-LightGBM model yielding high accuracy and identifying key factors influencing BEP, thereby offering actionable strategies for sustainable urban development.

The interest in probabilistic methodologies to demonstrate structural fire safety has increased significantly in recent times. However, the evaluation of the structural behavior under fire loading is computationally expensive even for simple structural models. In this regard, machine learning-based surrogate modeling provides an appealing way forward. Surrogate models trained to simulate the behavior of structural fire engineering (SFE) models predict the response at negligible computational expense, thereby allowing for rapid probabilistic analyses and design iterations. Herein, a framework is proposed for the probabilistic analysis of fire exposed structures leveraging surrogate modeling. As a proof-of-concept a simple (analytical) non-linear model for the capacity of a concrete slab and an advanced (numerical) model for the capacity of a concrete column are considered. First, the procedure for training surrogate models is elaborated. Subsequently, the surrogate models are developed, followed by a probabilistic analysis to evaluate the probability density functions for the capacity. The results show that fragility curves developed based on the surrogate model agree with those obtained through direct sampling of the computationally expensive model, with the 10–2 capacity quantile predicted with an error of less than 5%. Moreover, the computational cost for the probabilistic studies is significantly reduced by the adoption of surrogate models.

Background: The inevitable mistuning of a bladed disk breaks the cyclic symmetry of the structure and can cause vibration localization and amplitude magnification, thereby seriously reducing the structural reliability. Although numerical simulations based on high-fidelity finite element models can precisely calculate the response under some specific mistuning patterns, repeated calculations of random mistuning cases still bring a significant computational burden, even with ad-hoc reduced-order modeling technologies. Objective: In this work, we aim to capture the dynamics of randomly mistuned bladed disks using data-driven surrogate models. Methods: To do that, a physics-informed surrogate model of a designed structural neural network (DS) is proposed, where a “shift” feature from the cyclic symmetry in physics is taken into account for data augmentation. The finite element model of a dummy bladed disk with blade stiffness mistuning is established and taken as an example to test the effectiveness of the surrogate model. Under different training set sizes and excited modes, the performance of the proposed model is compared systematically with a fully connected structural neural network model (FC) and a state-of-the-art deep neural network framework (MS-DNN). Results: The DS model can accurately predict blade displacement for blade-dominated modes with only 600 training data. The R2 value of the displacement response amplitude reaches 0.99956, and the error of the amplitude magnification factor is only 0.68%. The DS model converges faster and has higher prediction accuracy than the FC model. The DS model is superior to the MS-DNN model concerning the generalization ability.

Yield estimation and optimization is ubiquitous in modern circuit design but remains elusive for large-scale chips. This is largely due to the mounting cost of transistor-level simulation and one's often limited resources. In this study, we propose a novel framework to estimate and optimize yield using Bayesian Neural Network (BNN-YEO). By coupling machine learning method with Bayesian network, our approach can effectively integrate prior knowledge and is unaffected by the overfitting problem prevalent in most surrogate models. With the introduction of a smooth approximation of the indicator function, it incorporates gradient information to facilitate global yield optimization. We examine its effectiveness via numerical experiments on 6T SRAM and found that BNN-YEO provides 100x speedup (in terms of SPICE simulations) over standard Monte Carlo in yield estimation, and 20x faster than the state-of-the-art method for total yield estimation and optimization with improved accuracy.

Many computer models of large complex systems are time consuming to experiment on. Even when surrogate models are developed to approximate the computer models, estimating an appropriate surrogate model can still be computationally challenging. In this article, we propose an Additive Global and Local Gaussian Process (AGLGP) model as a flexible surrogate for stochastic computer models. This model attempts to capture the overall global spatial trend and the local trends of the responses separately. The proposed additive structure reduces the computational complexity in model fitting, and allows for more efficient predictions with large data sets. We show that this metamodel form is effective in modelling various complicated stochastic model forms.

Summary Model-based reservoir management workflows rely on the ability to generate predictions for large numbers of model and decision scenarios. When suitable simulators or models are not available or cannot be evaluated in a sufficiently short time frame, surrogate modeling techniques can be used instead. In the first part of this paper, we describe extensions of a recently developed open-source framework for creating and training flow network surrogate models, called FlowNet. In particular, we discuss functionality to reproduce historical well rates for wells with arbitrary trajectories, multiple perforated sections, and changing well type or injection phase, as one may encounter in large and complex fields with a long history. Furthermore, we discuss strategies for the placement of additional network nodes in the presence of flow barriers. Despite their flexibility and speed, the applicability of flow network models is limited to phenomena that can be simulated with available numerical simulators. Prediction of poorly understood physics, such as reservoir souring, may require a more data-driven approach. We discuss an extension of the FlowNet framework with a machine learning (ML) proxy for the purpose of generating predictions of H2S production rates. The combined data-physics proxy is trained on historical liquid volume rates, seawater fractions, and H2S production data from a real North Sea oil and gas field, and is then used to generate predictions of H2S production. Several experiments are presented in which the data source, data type, and length of the history are varied. Results indicate that, given a sufficient number of training data, FlowNet is able to produce reliable predictions of conventional oilfield quantities. An experiment performed with the ML proxy suggests that, at least for some production wells, useful predictions of H2S production can be obtained much faster and at much lower computational cost and complexity than would be possible with high-fidelity models. Finally, we discuss some of the current limitations of the approach and options to address them.

A novel multiobjective adaptive surrogate modeling-based optimization (MO-ASMO) framework is proposed to utilize a minimal number of training samples efficiently for sequential model updates. All the sample points are enforced to be feasible, and to provide coverage of sparsely explored sparse design regions using a new optimization subproblem. The MO-ASMO method only evaluates high-fidelity functions at feasible sample points. During an exploitation sample phase, samples are selected to enhance solution accuracy rather than the global exploration. Sampling tasks are especially challenging for multiobjective optimization; for an n-dimensional design space, a strategy is required for generating model update sample points near an (n − 1)-dimensional hypersurface corresponding to the Pareto set in the design space. This is addressed here using a force-directed layout algorithm, adapted from graph visualization strategies, to distribute feasible sample points evenly near the estimated Pareto set. Model validation samples are chosen uniformly on the Pareto set hypersurface, and surrogate model estimates at these points are compared to high-fidelity model responses. All high-fidelity model evaluations are stored for later use to train an updated surrogate model. The MO-ASMO algorithm, along with the set of new sampling strategies, are tested using two mathematical and one realistic engineering problems. The second mathematical test problems is specifically designed to test the limits of this algorithm to cope with very narrow, non-convex feasible domains. It involves oscillatory objective functions, giving rise to a discontinuous set of Pareto-optimal solutions. Also, the third test problem demonstrates that the MO-ASMO algorithm can handle a practical engineering problem with more than 10 design variables and black-box simulations. The efficiency of the MO-ASMO algorithm is demonstrated by comparing the result of two mathematical problems to the results of the NSGA-II algorithm in terms of the number of high fidelity function evaluations, and is shown to reduce total function evaluations by several orders of magnitude when converging to the same Pareto sets.