Data Acquisition and Synthesis in Simulation-Based Design with Stochastic Models

van Beek, Anton

doi:https://doi.org/10.21985/n2-v68r-3773

Work

Data Acquisition and Synthesis in Simulation-Based Design with Stochastic Models

Public

Download PDF

Download All Files (.zip)

Engineering design is a systematic process of identifying needs and their translation into functional systems. This is a cyclic process that alternates between the acquisition of data and the synthesis of said data to inform design decisions. Conventionally, data from physical experiments are used to explore the efficacy of alternative design decisions but are increasingly supplanted by typically more monetary and time-efficient simulation models. In this dissertation, we present a study of statistical methods for simulation-based design with stochastic models (i.e., models that have variability in their outcome even when run with the same inputs). Without loss of generality, the statistical tools brought forth by this work have been leveraged for the design of new materials. Stochastic models in this domain are particularly prevalent as material properties dependent on uncontrollable variabilities that manifest at different length scales and among different part instances. Designing materials requires inference into the dependence of material performance criteria on a set of controllable design variables. Data synthesis provides designers with the means to leverage information from heterogeneous data sources to elucidate this functional relation. Although, computer simulations provide a cheaper/faster data source than physical experiments they often require the calibration of a set of unknown scalars or functions to be of sufficient fidelity. In this dissertation, we introduce an uncertainty quantification decision support framework for semi-parametric functional calibration. The contribution of this framework is that it enables designers to select the functional form of the calibration functions and optimize their parameters for problem dimensions intractable with existing methods. Using calibrated models in multiscale simulations enables designers to explore how part performance depends on fine-scale structural variations. For this purpose, designers are concerned with the investigation of data synthesis tools that can characterize part-to-part variation from a manageable number of experimental observations. In this dissertation, we introduce an uncertainty quantification scheme for multiscale materials with non-stationary distributed agglomeration anomalies by hierarchically sampling descriptors at progressively smaller length scales. The contribution of this work is a sampling and emulation framework that enables a designer to explore how discretely distributed microstructure features, such as agglomerations, influence part performance. Although simulations models (e.g., multi-scale simulations, and molecular dynamics) greatly reduce the cost of exploring vast material design spaces, they can still be prohibitively costly when evaluated many times. This is where response surface modeling plays an important role as they can use a relatively small set of training data to predict the functional response of unobserved designs. To improve prediction accuracy, the response surface model fidelity can be enhanced by sequentially adding new training samples at spatial locations informed by past observations. In this dissertation, we introduce a global emulation approach that uses normative decision-making and thrifty adaptive batch sampling. The contribution of this work is that the autonomous sampling agent considers the influence of future sampling decisions and adaptive selection of batch size. Finally, to design a new material we need both data synthesis and data acquisition to efficiently explore the material design space with a minimal number of simulation data. We conclude this dissertation by combining the stochastic Kriging data synthesis model with a novel data acquisition function for stochastic models to design a new organic photovoltaic cell with optimal efficiency. The contribution of this work is twofold: (i) the sampling agent adaptively weights the trilemma of exploration, exploitation, and replication, and (ii) it enables designers to run batches of samples to optimally benefit from parallel computing capabilities. Through an extensive benchmark test, we show that our method uniformly outperforms all available methods.

Creator