Session J28: Low-Order Modeling: Design

Data-driven optimisation of two-phase problems using composite fidelities

Design optimisation plays a crucial role across industries, including chemical engineering, to enhance performance, sustainability, and energy efficiency. It is necessary to develop computationally inexpensive methods that can quickly evaluate and select optimal designs, particularly for processes involving multiphase flows. In this study, we present a framework that addresses this challenge by employing multi-fidelity Bayesian optimisation. This framework is integrated with the OpenFOAM solver using the PyFoam library, allowing for low to high-fidelity simulations that balance cost and accuracy in exploring the design parameter space. The fidelities considered in this work include variations in mesh sizes in the continuous space and different multiphase flow models, such as Eulerian-Eulerian, geometric VOF, and algebraic VOF, in the discrete space. To demonstrate the effectiveness of our framework, we apply it to a ‘toy problem’ involving two immiscible fluids (water and silicon oil) in a channel. The goal is to shape-parameterise the channel to maximise mixing efficiency. By formulating a cost-based acquisition function, our framework automatically selects the combination of fidelities for function evaluations based on the expected value of the objective function. This approach leads to the achievement of near-optimal designs in a computationally efficient manner. We anticipate that this low-cost modelling framework can be extended to a wide range of industrial problems involving multiphase flows to address design optimisation challenges.   

Application of Proper Orthogonal Decomposition to Variable in Computational Space for Reduced Order Modeling

A reduced order model (ROM) has been widely used as a new approach in computational fluid dynamics (CFD). The ROM estimates a flow field roughly and quickly while the conventional CFD approach obtains the flow field with high accuracy. One of the advantages of CFD is the ability to modify the shape of an object and perform calculations. This enables the optimization of object shape, aiming to increase the lift force and reduce the drag force. However, in conventional CFD, it is necessary to perform a vast number of computations for flow fields around each specific shape. To reduce computational time, one approach is to develop an ROM using a limited number of flow fields under representative conditions. The developed ROM can estimate the flow field with unknown shapes quickly. There are various methods for the development of ROM. An ROM based on the proper orthogonal decomposition (POD) has been studied. The POD is a method to construct a reduced-order space that best represents the snapshots. However, The POD cannot decompose snapshots of flow fields including various computational grids. Therefore, at present, POD-based ROM cannot be utilized in shape optimization. To utilize shape optimization, we propose a POD method that performs decomposition with respect to the velocity in the generalized coordinate system rather than in the Cartesian coordinate system. To be based on the proposed POD, an ROM for flow fields around an object with different shapes is developed.   

An End-to-End Framework Coupling CFD and Active Machine Learning Optimizer (ActivO) for Rapid Simulation-driven Design of Turbulent Jet Mixers

Chemical process efficiency is a critical part of sustainability efforts in the chemical industry and often relies on effective mixing of multiple components. These types of mixing equipment (known as "mixers") can have complex geometries and internal features that pose optimization and scale-up challenges, particularly when no guiding principles or empirical correlations exist for designing a new type of mixer. In this context, computational fluid dynamics (CFD) can be leveraged to guide mixer design. However, it is difficult and time-consuming to perform design optimization using manual design-of-experiments type approaches or conventional evolutionary optimizers. In this work, an end-to-end simulation-AI framework is developed for optimizing the design of a gas-liquid turbulent jet mixer wherein a CFD model of the mixer is coupled with a novel active ensemble machine learning optimizer (ActivO) developed at Argonne National Laboratory. The workflow is structured as a loop of the following steps: (i) multi-parameter morphing of the mixer geometry with the CAD tool CAESES, (ii) automated grid and case definition and CFD simulation with open-source software OpenFOAM, (iii) multi-dimensional optimization guided by ActivO to modify the geometry towards the maximization of mixing performance and energy efficiency.   

Bayesian design optimization of a BLI propulsion fan with regenerative air brake for electric passenger aircraft

In our previous work, we conducted an optimal design of a propulsion fan considering regenerative air braking and found that high-efficiency performance tends to be obtained when both regenerative and propulsion are in the high tip speed ratio. An effective means of achieving a high tip speed ratio is the use of Boundary Layer Injection (BLI), in which the slow velocity boundary layer is directly sucked in by the fan. Therefore, this study focuses on a regenerative airbrake using a BLI propulsion fan with a non-variable angle of attack for electric aircraft. The objective is to realize a fan shape with high efficiency during propulsion and regeneration with BLI at different rotation speeds and to find its characteristics. To realize the fan shape, it is necessary to conduct a global exploration of multivariable problems. But it is not realistic to conduct many costly evaluations, such as CFD. Therefore, Bayesian optimization is used as an efficient and global optimization method using approximate functions, and an attempt is made to find a fan shape with high performance in both efficiencies. The relationship between each efficiency and fan shape is then visually analyzed using self-organizing maps. As a result, we will find a fan shape that exhibits positive efficiency in both cases and shows aerodynamically appropriate shape characteristics for a BLI propulsion fan with regenerative air brake function. The results will lead to a reduction in total power consumption per flight in the future as electric fan engines.   

Toward automatic CFD of flow through a blade passage using deep reinforcement learning

A method for generating an optimal mesh for a passage of an arbitrary blade is developed using deep reinforcement learning (DRL). Despite advances in the automation of mesh generation using empirical approaches or optimization algorithms, repeated tuning of meshing parameters remains necessary for each new configuration. The present method employs a DRL-based multi-condition optimization technique to define optimal meshing parameters as functions of the blade geometry and the flow condition, attaining automation, minimization of human intervention, and computational efficiency. An elliptic mesh generator, which requires manual tuning of meshing parameters, is developed to produce a structured mesh for a blade passage. Among these parameters, those controlling the mesh shape are optimized to maximize the geometric mesh quality. Subsequently, resolution-related meshing parameters are optimized by integrating the results from computational fluid dynamics simulations. After training, the mesh generator can create an optimal mesh for a new arbitrary configuration in a single try, eliminating repetitive parameter tuning. The effectiveness and robustness of the proposed method are demonstrated through the generation of meshes for various blade passages.   

Data-driven inference of adjoint sensitivities without adjoint solvers: An application to thermoacoustics

Adjoint methods offer a computationally cheap and accurate way to calculate the sensitivity of a quantity of interest with respect to all the system parameters. However, adjoint methods require the implementation of an adjoint solver, which can be cumbersome. In this work, we infer the adjoint solution from data via data-driven stability analysis with reservoir computing. First, we derive the adjoint of a reservoir computer, and compute the sensitivity of the acoustic energy of a prototypical thermoacoustic system with respect to its design parameters. Second, we improve generalizability and robustness by embedding the physical knowledge about the nonlinearity and the time-delayed nature of the thermoacoustic dynamics. Third, we employ the data-driven sensitivity provided by the adjoint of the trained network within a parameter optimization framework to minimise the acoustic energy. This work opens possibilities for data-driven gradient-based design optimization.   

CNN Embedded Topology Optimization framework for Designing Channel Flow Layouts

Topology optimization, as a non-heuristic and highly versatile design tool, has attracted considerable interest and extensive investigation across diverse fields. However, when dealing with multivariate problems such as optimizing channel flow layout, the increased computational cost due to the large-scale finite element solution process restricts its practicality. In this study, we proposed a novel machine learning embedded topology optimization framework to achieve a high-efficient procedure for optimizing channel flow problems. The framework incorporates a convolutional neural network that proficiently learns and replace the conventional finite element derivation process. Moreover, it can autonomously adjust and update its network parameters online, ensuring a continuous improvement in prediction accuracy. In the case study of a 90-degree bending design, conducted without any offline training data, the proposed method outperforms the original topology optimization method: over 100 iterations, the total time required for calculations is reduced by 60%; Simultaneously, the optimized result exhibits a decrease of 4.8% in the pressure drop value. The further improvement of the optimized results is due to the introduction of the prediction error to address the local optimization trend of the original method. The proposed method has the potential to replace existing techniques and facilitate more efficient and superior new topology optimization processes.   

Design Methodology for Smart Structures with Wrinkle-Induced Shape Transformation

As comprehension of materials’ behaviors under specific conditions advances, the demand for materials capable of altering their shape under given circumstances, referred to as smart structures, has risen due to their ability to provide the multiple functionality. This research endeavors to present a novel design methodology for smart structures that introduce unique surface wrinkles, enabling them to undergo shape transformation in response to external conditions. Our approach focuses on understanding the deformation behavior of structures made from hyperelastic materials with surface wrinkling patterns, like Turing patterns. By optimizing design parameters, we aim to incorporate these wrinkling patterns, obtaining distinctive characteristics of hyperelastic behavior in genetic materials and exploring their influence on the deformation behavior of structures.

In pursuit of our research objective, a specific approach that involves confining the design space and representing it as a collection of repetitive structures, called unit cells. Within this method, the discretized internal space of each unit cell is considered as pixelated space, so that a thorough analysis is performed to explore potential configurations of pixel arrays. The outcome of this analysis enhances understanding of the deformation behavior of pixelated unit cell structures, which subsequently serves as the foundation for generating and validating smart structures made by unit cells.

Through the integration of strategically introduced wrinkle patterns and the meticulous analysis of pixel arrays, our research endeavors to elucidate valuable insights into the intricate behavior of materials and their dynamic responses to external stimuli. This innovative approach holds the promise of encouraging a paradigm shift in the realm of smart structures. By leveraging this approach, considerable improvements in performance, durability, and predictability can be expected, thereby enhancing the efficacy and applicability of materials across a wide array of practical applications.