Machine Learning for Explainable Roundtrip Polymer Reaction Engineering

点击翻译按钮获取中文摘要

Polymers are important materials in everybody’s daily life and in numerous technical applications. They have a wide range of properties that can be tailored by the type and the conditions of the production process, ideally by modeling of the polymerization process. The project ML-PRE aims to bridge the gap between state-of-the-art machine learning (ML) methods and their application in modeling and optimization in polymer reaction engineering (PRE). This will enable a novel approach referred to as roundtrip PRE, covering integrated polymerization process modeling and reverse engineering of the polymerization process. The reverse engineering aspect is particularly novel to the field. The overall approach aims at designing new sustainable production processes and developing polymers with better or even new properties.The baseline of the project is the modeling of polymerization processes using kinetic Monte Carlo (KMC) methods. An open-source KMC simulator will be used to generate data sets for training and testing the ML methods. The overall goal is broken down into the following scientific objectives of ML-PRE: (1) To create a coherent and validated suite of scalable ML-based models for polymerization modeling, facilitating fast and efficient simulation-supported learning of ML models, extending the KMC simulator for novel types of problems (e.g. acrylate polymerizations at high temperature, diffusion-controlled termination). (2) To create ML-based approaches for reverse engineering of polymerization processes with modeling and optimization capabilities. (3) To create ML-based models for learning controllers of semi-batch polymerization processes. (4) To create a general and transferable methodology for increasing the transparency of ML created in the project by means of suitable and validated explainability techniques. From the ML perspective, the main innovations are provided by the first and the fourth objective: Objective 1 is about creating a coherent suite of validated ML models with interfaces designed to support the flexibility to support the complex bi-directional workflows in roundtrip PRE. The second main innovation is that through Objective 4 above, we aim at a general and transferable methodology to bring about and maintain transparency and explainability of the ML methods created and validated for roundtrip PRE.Referring to the collaboration matrix of the SPP call we mainly address target area  #1 (optimal decision making); through the work on ML-supported simulation we also cover some aspects of target area #2 (introducing/enforcing physical laws in machine learning models). W.r.t. the collaboration matrix we expect and work towards that results for mechanistic models in the 1st column (Phenomena / Micro-scale) will be transferable to the areas  mechanistic models, experiments (real process), and optimization of the 3rd column (Flowsheet / Process) and vice versa.