Data-Driven Surrogate-Assisted Evolutionary Fluid Dynamic Optimisation

计算流体动力学（CFD）是现代工程设计的基础，从飞机、汽车到家用电器。它允许对流体的行为进行计算模拟，并对新设计进行评估。找到最好的设计是非常具有挑战性的，因为有大量的设计需要探索。计算优化是现代科学、商业和工业的一项关键技术。它允许自动调整计算模型的参数，以最大化某些利益，并可以揭示真正创新的解决方案。例如，可以优化飞机的形状以最大化计算出的升力/阻力比。解决优化问题的一套非常成功的方法被称为进化算法，之所以这么叫，是因为它们的灵感来自于自然界进化机制优化生物体适应性的方式。这些算法的工作原理是基于先前评估的解决方案的重组和/或变化，迭代地提出新的解决方案（飞机的形状）进行评估，并通过保留好的解决方案和丢弃表现不佳的解决方案，进化出一群优化的解决方案。如果每个新解决方案的评估都需要很长时间，可能需要数小时或数天，就像复杂的CFD模拟一样，那么在具有许多参数的非常复杂的问题上使用进化算法就会出现障碍。在进化优化过程中必须评估的大量解决方案（通常是数千个）使整个优化变得不可行的。本研究旨在通过替代计算更简单、动态生成的“替代”模型来代替完整的CFD评估，从而加速优化过程。挑战在于从相对较少的精心选择的完整评估中自动学习适当的替代。我们的工作旨在弥合代理模型之间的差距，代理模型在只有少数设计参数需要优化时工作良好，但在大型工业规模的问题上失败。我们的方法有几个相互关联的方面。一个有吸引力但具有挑战性的途径是加快计算模型的速度。这里的关键是这些模型中的许多都是迭代的，一遍又一遍地重复相同的过程，直到获得准确的结果。我们将研究在早期迭代中利用部分信息来预测准确的结果，以及在进化搜索中使用粗略的早期结果来代替准确的结果。本研究的另一个主要目的是使用先进的机器学习方法从完整的评估中学习设计参数与被评估目标的关系。在这里，我们将通过研究新的机器学习方法来解决与许多设计参数相关的计算困难，以发现哪些参数在优化的任何阶段都是相关的。与此相关的是“主动学习”方法的发展，在这种方法中，代理模型自己选择最具信息量的解决方案来进行全面评估。一种整合部分信息、先进机器学习和主动学习的协同方法将被创造出来，以解决大规模优化问题。这项工作的一个重要组成部分是我们与从事实际CFD的合作伙伴的密切合作。我们将与英国航空航天技术研究所和QinetiQ合作进行复杂的空气动力学优化，与Hydro International合作进行旋风分离，与Ricardo合作进行柴油颗粒跟踪。这种多样化的合作将确保研究是由现实的工业问题驱动的，并建立在现有的工业经验基础上。这项工作的成功结果将是新的代理辅助进化算法，该算法被证明可以加速全尺寸工业CFD问题的优化。

Computational fluid dynamics (CFD) is fundamental to modern engineering design, from aircraft and cars to household appliances. It allows the behaviour of fluids to be computationally simulated and new designs to be evaluated. Finding the best design is nonetheless very challenging because of the vast number of designs that might be explored. Computational optimisation is a crucial technique for modern science, commerce and industry. It allows the parameters of a computational model to be automatically adjusted to maximise some benefit and can reveal truly innovative solutions. For example, the shape of an aircraft might be optimised to maximise the computed lift/drag ratio.A very successful suite of methods to tackle optimisation problems are known as evolutionary algorithms, so-called because they are inspired by the way evolutionary mechanisms in nature optimise the fitness of organisms. These algorithms work by iteratively proposing new solutions (shapes of the aircraft) for evaluation based upon recombinations and/or variations of previously evaluated solutions and, by retaining good solutions and discarding poorly performing solutions, a population of optimised solutions is evolved.An obstacle to the use of evolutionary algorithms on very complex problems with many parameters arises if each evaluation of a new solution takes a long time, possibly hours or days as is often the case with complex CFD simulations. The great number of solutions (typically several thousands) that must be evaluated in the course of an evolutionary optimisation renders the whole optimisation infeasible. This research aims to accelerate the optimisation process by substituting computationally simpler, dynamically generated "surrogate" models in place of full CFD evaluation. The challenge is to automatically learn appropriate surrogates from a relatively few well-chosen full evaluations. Our work aims to bridge the gap between the surrogate models that work well when there are only a few design parameters to be optimised, but which fail for large industry-sized problems.Our approach has several inter-related aspects. An attractive, but challenging, avenue is to speed up the computational model. The key here is that many of these models are iterative, repeating the same process over and over again until an accurate result is obtained. We will investigate exploiting partial information in the early iterations to predict the accurate result and also the use of rough early results in place of the accurate one for the evolutionary search. The other main thrust of this research is to use advanced machine learning methods to learn from the full evaluations how the design parameters relate to the objectives being evaluated. Here we will tackle the computational difficulties associated with many design parameters by investigating new machine learning methods to discover which of the many parameters are the relevant at any stage of the optimisation. Related to this is the development of "active learning" methods in which the surrogate model itself chooses which are the most informative solutions for full evaluation. A synergistic approach to integrate the use of partial information, advanced machine learning and active learning will be created to tackle large-scale optimisations.An important component of the work is our close collaboration with partners engaged in real-world CFD. We will work with the UK Aerospace Technology Institute and QinetiQ on complex aerodynamic optimisation, with Hydro International on cyclone separation and with Ricardo on diesel particle tracking. This diverse range of collaborations will ensure research is driven by realistic industrial problems and builds on existing industrial experience. The successful outcome of this work will be new surrogate-assisted evolutionary algorithms which are proven to speed up the optimisation of full-scale industrial CFD problems.