Automatic Design of Experiments in Chemical Micro-reactors

实验设计领域涉及通过选择不同的实验输入来最大化（或最小化）某个目标函数。贝叶斯优化已被证明可以为此类问题提供解决方案。查询数据用于计算代理模型的后验，我们可以使用它来选择下一个实验（通过优化称为获取函数的函数）。然而，在经典设置中，贝叶斯优化假设我们选择一个单一的实验并立即获得观察结果。微反应器正在改变实验室化学，因为它们允许我们在微观尺度上进行许多实验，因此，它们需要自动实验设计。微滴穿过反应器，每个微滴都可以看作是一个单独的实验。然而，这个问题带来了许多复杂性。这个问题的一个非常重要的部分与时间延迟有关。在得到以前实验的结果之前，我们必须选择许多新的实验。我们还将收到来自多个来源的观测结果，有些将快速但不准确，而另一些将准确但缓慢且昂贵。当我们试图使化学反应保持在稳态时，我们会限制输入的变化。其他重要的挑战包括多目标优化、输入延迟和安全约束。许多这些并发症都是在孤立的环境中进行研究的。特别是，在多保真度和异步贝叶斯优化方面有大量的文献。该项目的目标是提出可以同时考虑尽可能多的复杂性的方法。为了提供这样一种方法，我们必须以新颖的方式结合并从根本上改变以前提出的方法。有一个有效的方法来设计这样的实验将使我们更有效地生产化学品，并减少浪费。该项目的影响将不限于化学品生产。实验设计的设置在许多领域都很重要，从食品制造到机器学习超参数的优化。该研究项目是ESPRC的人工智能和机器人、工程和数学科学主题的交集。这项研究是与化学品制造商巴斯夫合作进行的，巴斯夫为该项目提供资金。这次合作有望使我们能够在现实生活中测试任何已开发的方法，包括微反应器和其他化学实验。它还应该帮助我们弥合这个项目的数学性质和它试图解决的化学应用之间的差距。

The area of experimental design concerns trying to maximise (or minimise) a certain objective function by selecting different experimental inputs. Bayesian Optimisation has proven to provide solutions to such problems. Query data is used to calculate a posterior from a surrogate model, which we can use to select the next experiment (by optimising a function called the acquisition function). However, in the classical setting, Bayesian Optimisation assumes we select a single experiment and immediately obtain an observation.Micro-reactors are changing laboratory chemistry as they allow us to carry out many experiments on the micro-scale and as such, they require automatic experimental design. Micro-droplets travel through the reactor and each can be considered to be a single experiment. However, the problem brings many complications. A very important part of the problem concerns time-delay. We will have to choose many new experiments before we receive the results from previous ones. We will also be receiving observations from multiple sources, some will be quick but inaccurate, while others will be accurate but slowand expensive. We will have restrictions on how much we want to vary our inputs, as we try to maintain the chemical reaction in steady-state. Other important challenges include multi-objective optimisation, input delay, and safety constraints.Many of these complications have been studied in isolated environments. In particular, there is extensive literature on multi-fidelity and asynchronous Bayesian Optimisation. The objective of the project is to propose methods that can take into account as many complications as possible, at the same time. To provide such a method we will have to combine and fundamentally change previouslyproposed methods, in novel ways. Having an effective way of designing such experiments would allow us to have more efficient production of chemicals, and reduce waste. The impact of the project would not be restricted tochemical production. The setting of experimental design is important in many areas, from food manufacturing to the optimisation of machine learning hyper-parameters. The research project falls in an intersection of ESPRC's themes of Artificial Intelligence and Robotics, Engineering, and Mathematical Sciences.The research is being done in collaboration with chemical manufacturer BASF, whom are providing funding for the project. The collaboration will hopefully allow us to test any developed methods in real life settings, involving micro-reactors and other chemical experiments. It should also help us bridge the gap between the mathematical nature of the project, and the chemical applications it is trying to tackle.