In this work, an in house topology optimization (TO) solver is developed to optimize a conjugate heat transfer problem: realizing more complex and efficient coolant systems by minimizing pressure losses and maximizing the heat transfer. The TO method consists in an idealized sedimentation process in which a design variable, in this case impermeability, is iteratively updated across the domain. The optimal solution is the solidified region uniquely defined by the final distribution of impermeability. Due to the geometrical complexity of the optimal solutions obtained, this design method is not always suitable for classic manufacturing methods (molding, stamping....) On the contrary, it can be thought as an approach to better and fully exploit the flexibility offered by additive manufacturing (AM), still often used on old and less efficient design techniques. In the present article, the proposed method is developed using a Lagrangian optimization approach to minimize stagnation pressure dissipation while maximizing heat transfer between fluid and solid region. An impermeability dependent thermal conductivity is included and a smoother operator is adopted to bound thermal diffusivity gradients across solid and fluid. Simulations are performed on a straight squared duct domain. The variability of the results is shown on the basis of different weights of the objective functions. The solver builds automatically three-dimensional structures enhancing the heat transfer level between the walls and the flow through the generation of pairs of counter rotating vortices. This is consistent to solution proposed in literature like v-shaped ribs, even if the geometry generated is more complex and more efficient. It is possible to define the desired level of heat transfer and losses and obtain the closest optimal solution. It is the first time that a conjugate heat transfer optimization problem, with these constraints, has been tackled with this approach for three-dimensional geometries.

Uncertainty quantification (UQ) has recently become an important part of the design process of countless engineering applications. However, up to now in computational fluid dynamics (CFD) the errors introduced by the turbulent viscosity models in Reynolds-Averaged Navier Stokes (RANS) models have often been neglected in UQ studies. Although Direct Numerical Simulations (DNS) are physically correct, obtaining a large enough set of DNS data for UQ studies is currently computationally intractable.UQ based only on RANS simulations or on DNS often leads to physical and statistical inaccuracies in the output probability distribution functions (PDF). Therefore, three hybrid methods combining both RANS simulations and DNS to perform non-intrusive UQ are suggested in this work. Low-fidelity RANS simulations and high-fidelity DNS are combined to give an approximation of an output PDF using the advantages of both data sets: the physical accuracy via the DNS and the statistical accuracy via the RANS simulations. The hybrid methods are applied to the flow over 2D periodically arranged hills. It is shown that the Gaussian CoKriging (GCK) method is the best hybrid method and that a non-intrusive hybrid UQ approach combining both DNS and RANS simulations is possible, with both physically more accurate and statistically better PDF.

This paper shows the application of Deep Neural Network algorithms for Fluid-Structure Topology Optimization. The strategy offered is a new concept which can be added to the current process used to study Topology Optimization with Cellular Automata, Adjoint and Level-Set methods. The design space is described by a computational grid where every cell can be in two states: fluid or solid. The system does not require human intervention and learns through an algorithm based on Deep Neural Network and Monte Carlo Tree Search. In this work the objective function for the optimization is an incompressible fluid solver but the overall optimization process is independent from the solver. The test case used is a standard duct with back facing step where the optimizer aims at minimizing the pressure losses between inlet and outlet. The results obtained with the proposed approach are compared to the solution via a classical adjoint topology optimization code.

This work presents an innovative design method to obtain valves without moving parts that can be built using additive manufacturing and applied to gas turbines. Additive manufacturing offers more flexibility than traditional manufacturing methods, which implies less constraints on the manufacture of engineering parts and it is possible to build complex geometries like the Tesla valve. The Tesla valve is a duct that shows a diodicity behavior: it allows a fluid to flow in one direction with lower losses than in the other one. Unfortunately the design of the Tesla valve is two dimensional and it relies on the designer experience to obtain good performance. The method presented here allows the automatic generation of valves similar to the Tesla one, obtained automatically by a topology optimization algorithm. It is the first time that a three dimensional method is presented, the available algorithms in the open literature works in two dimensions. A fluid sedimentation process enables the creation of a new geometry optimized to meet a prescribed set of performance, such as pressure losses. The steepest descent method is used to approximate the integrals met during the calculation process. The optimizer is used to obtain three dimensional geometries for different multi-objective functions. The geometry is compared to an existing similar solution proposed in the open literature and validated. The results are compared to a Tesla valve to show the performance of the optimized geometries. The advantage of the proposed solution is the possibility to apply the design method with any spatial constraints and for a wide range of mass flow.

This work presents a machine learning based method for bi-fidelity modelling. The method, a Knowledge Based Neural Network (KBaNN), performs a local, additive correction to the outputs of a coarse computational model and can be used to emulate either experimental data or the output of a more accurate, but expensive, computational model. An advantage of the method is that it can scale easily with the number of input and output features. This allows bi-fidelity modelling approaches to be applied to a wide variety of problems, for instance in the bi-fidelity modelling of fields. We demonstrate this aspect in this work through an application to Computational Fluid Dynamics, in which local corrections to a velocity field are performed by the KBaNN to account for mesh effects. KBaNNs were trained to make corrections to the free-stream velocity field and the boundary layer. They were trained on a limited data-set consisting of simple two-dimensional flows. The KBaNNs were then tested on a flow over a more complex geometry, a NACA 2412 airfoil. It was demonstrated that the KBaNNs were still able to provide a local correction to the velocity field which improved its accuracy. The ability of the KBaNNs to generalise to flows around new geometries that share similar physics is encouraging. Through knowledge based neural networks it may be possible to develop a system for bi-fidelity, computer based design which uses data from past simulations to inform its predictions.