Periodic Reporting for period 1 - CINEMA (Chemistry informed machine learning in emulsion polymerization processes and products)
Reporting period: 2023-01-01 to 2024-12-31
Machine learning (ML) systems continue to revolutionize many aspects of daily life, but despite their immense potential have yet to impact significantly in polymer science. One major issue that is hindering the more widespread use of machine learning in polymer science, and many other physical sciences, lies in the challenges in amassing sufficient data to efficiently train machine learning models. This in itself is not necessarily a problem, and is an issue frequently encountered in the machine learning field, but can only be resolved by a thorough understanding of the science behind the problem of interest. CINEMA aims to providing a training platform that will allow the next generation of polymer scientists to take polymer science into the 21st century through incorporating the fundamental knowledge gained over many years of research into the training of machine learning systems. Such a knowledge-driven machine learning approach puts the scientific issues of CINEMA at the forefront of the use of machine learning in fundamental scientific problems, and also provides the perfect training platform for the next generation of scientists, for whom the use of AI will be an invaluable tool.
The scientific work of CINEMA has been divided into a number of work packages that are designed to achieve the diverse objectives of the CINEMA program.
A first major focus is to develop machine learning as a way to control polymerization reactors. Two distinct approaches to achieve this are being developed. The first is to use machine learning models as surrogates for computationally expensive first-principles mathematical models, such that they can be used in real-time for online control and prediction of future trajectories. This has proven successful with up to 100 million fold improvement in computation times without any loss of prediction accuracy. The second approach is to develop a machine learning approach that learns a model autonomously from experimental data for later use in control systems.
A second major focus is to develop machine learning as a way to predict polymer properties. In this area, efforts have been made to integrate previously available knowledge of polymer materials alongside limited experimental datasets. By incorporating relevant knowledge across multiple lengthscales, this approach is being successfully used for predicting application properties of polymer materials such as tensile modulus and adhesion, as well as physical properties such as glass transition temperature and polymer solubility.
A first major focus is to develop machine learning as a way to control polymerization reactors. Two distinct approaches to achieve this are being developed. The first is to use machine learning models as surrogates for computationally expensive first-principles mathematical models, such that they can be used in real-time for online control and prediction of future trajectories. This has proven successful with up to 100 million fold improvement in computation times without any loss of prediction accuracy. The second approach is to develop a machine learning approach that learns a model autonomously from experimental data for later use in control systems.
A second major focus is to develop machine learning as a way to predict polymer properties. In this area, efforts have been made to integrate previously available knowledge of polymer materials alongside limited experimental datasets. By incorporating relevant knowledge across multiple lengthscales, this approach is being successfully used for predicting application properties of polymer materials such as tensile modulus and adhesion, as well as physical properties such as glass transition temperature and polymer solubility.
The current scientific impact of the CINEMA project is still relatively limited, as much of the scientific progress has only been shared to a limited extent at conferences with formal publications yet to be released. However, the impact is expected to grow significantly in the coming year as the research begins to be published.
In terms of economic impact, a machine learning model for the prediction of polymer solubility is currently being patented. In addition to this, the secondments with commercial sector companies that are part of the consortium, that are planned for the coming year, will help to transfer knowledge and bridge the gap between academia and industry. This has been reinforced by contact between the academic groups and the companies that are associated members of the consortium.
A major long-term impact will be the doctoral candidates (DCs) themselves, who are benefiting from a unique training program aligned with the growing emphasis on digital technologies and AI, preparing them for future roles in these evolving fields.
In terms of economic impact, a machine learning model for the prediction of polymer solubility is currently being patented. In addition to this, the secondments with commercial sector companies that are part of the consortium, that are planned for the coming year, will help to transfer knowledge and bridge the gap between academia and industry. This has been reinforced by contact between the academic groups and the companies that are associated members of the consortium.
A major long-term impact will be the doctoral candidates (DCs) themselves, who are benefiting from a unique training program aligned with the growing emphasis on digital technologies and AI, preparing them for future roles in these evolving fields.