Machine learning for design of chemical engineering unit operations - a microevaporator, leading to a 3D structured multiphase absorber

在过去的几年里，机器学习以令人难以置信的速度发展，由于各种原因，机器学习的应用主要集中在特定的领域，包括自然语言处理、计算机视觉和一些自然科学。我们计划使用ML将通过模拟和实验产生的有关化工装置操作(例如微型蒸发器或多相反应器)的数据结合在一起，不仅预测设备特性，而且还设计和提出设备改进建议，以及解释ML模型提出这些改进的原因。我们进行了一项初步的概念验证研究，其中我们证明了机器学习模型，特别是卷积神经网络(CNN)能够预测流动装置的性能。在我们初步研究的基础上，我们计划利用新型机器学习工具和最先进的模拟能力，开发生成性ML模型来设计微蒸发器和3D结构多相吸收装置。虽然这项任务听起来非常适合应用，但它反过来将使我们能够改进现有的、广泛适用的机器学习方法。我们项目的第一个目标是使用机器学习方法预测新的流道结构，包括使用不确定性感知2D卷积神经网络开发集成的主动学习工作流程，以及基于虚拟高通量筛选、生成模型和用于探索和优化的遗传算法的基于ML的设备设计评估。第二个目标是设计新颖的微流控蒸发器，该蒸发器由流道排列成阵列组成，利用ML模型预测产生高质量的饱和蒸汽。在这里，我们计划开发ML模型来弥合多通道微结构蒸发器的高精度直接数值模拟和成本效益模拟之间的差距。第三个目标是使用基于ML的设计来探索复杂多相流系统的全新3D几何结构。我们打算开发一种方法，用于在来自模拟和实验的异类数据集上训练生成的ML模型，纳入适合打印的启发式方法，最终用于金属3D打印和ML预测的实验评估。

Machine learning (ML) has evolved at an incredible pace over the past years, and for various reasons the applications have been focused most heavily in particular domains, including natural language processing, computer vision, and some of the natural sciences. We plan to use ML to combine data about unit operations in chemical engineering (e.g. micro-evaporators or multiphase reactors) generated using simulations and experiments, to not only predict device characteristics, but to additionally design and suggest device improvements, along with explanations of why the ML model suggested them. We conducted an initial proof-of-concept study, in which we showed that machine learning models, in particular convolutional neural networks (CNNs) are capable of predicting properties of flow devices. Building on top of our preliminary study, we plan to develop generative ML models to design microevaporators and 3D structured multiphase absorbers, leveraging the capabilities of novel machine learning tools and state-of-the-art simulations. While this task sounds highly application specific, it will in turn enable us to improve existing, widely applicable machine learning methods. Focus along these lines will be on uncertainty quantification and active learning, as well as on scientific interpretation and of generative models.The first objective of our project is the predicting novel flow channel structures using machine learning methods, including the development of integrated active learning workflows using uncertainty aware 2D convolutional neural networks, as well as the evaluation of ML based device design based on virtual high throughput screening, generative models and genetic algorithms for exploration and optimization. The second objective is the design novel microfluidic evaporators consisting of arrangements of flow channels into arrays that produce high quality saturated vapour using ML model predictions. Here, we plan to develop ML models to bridge the gap between highly accurate direct numerical simulations and cost effective simulations of multiple-channel microstructured evaporators. The third objective is the exploration of completely novel 3D geometries for complex multiphase flow systems using ML based design. We intend to develop methods for the training of generative ML models on heterogeneous datasets from simulation and experiment, for the incorporation of heuristics for printability and finally for the metal 3D printing and experimental evaluation of ML predictions.