A team from Imperial and BASF has won the 2023 Computers & Chemical Engineering Best Paper Award for their AI technique that has the potential to advance chemical research and development.
The prestigious journal of process systems engineering rated the paper as the best paper out of more than 280 published that year.
The trial and error process in chemical research and development is costly and can take weeks for some experiments, so chemists need to find the optimal manufacturing setup with as few experiments as possible.
The paper, by doctoral student JosƩ Pablo Forti and colleagues from BASF and Imperial, applies a set of classical Bayesian statistical methods, used to obtain the most useful information possible from a finite number of experiments, to a specific research and development (R&D) methodology used by a chemical company.
“The editorial team appreciated that the authors' work straddles theory and practice by developing a novel Bayesian optimization approach and applying it to an industrially relevant application,” said Professor Stratos Pistikopoulos, Editor-in-Chief of Computer & Chemical Engineering at Texas A&M University.
“It is fantastic to see research that is both academically pioneering and has important practical value for industry being recognised by a flagship journal for the process systems engineering community,” said co-author Ruth Misener, professor in the School of Computing and the EPSRC Statistics and Machine Learning Doctoral Training Centre.
R&D Challenges
Experimentation is key in R&D in the chemical industry. Before setting up a new production line or facility, industrial chemists test various manufacturing parameters, such as temperature settings and raw materials, to maximize product purity and minimize economic and environmental costs.
“It is fantastic to see such pioneering and important research for industry recognized by a leading journal in process systems engineering.” Professor Ruth Misener Compute
This trial-and-error process is itself costly ā some experiments can take weeks or even months ā and requires significant resources to run, which is why chemists need to run as few experimental iterations as possible to find near-optimal manufacturing settings.
Imperial and BASF team behind the win Computers and Chemical Engineering In this paper, we created a new algorithm based on Bayesian optimization, a statistical method that helps to obtain optimal manufacturing parameters with a small number of experiments.
Bayesian optimization uses a small amount of initial experimental data to create a curve that models the relationship between a particular manufacturing parameter (e.g., temperature) and performance (e.g., purity), and assigns different levels of certainty to different parts of the curve based on the available data.
It tells the experimenter which parameter values āāto test next by finding a compromise between test values āāthat are already expected to give good performance and high-risk experiments where the outcome is very uncertain but may give even better performance. It updates the curves and confidence intervals for each data point after each iteration and then recommends the next iteration.
Chemistry Advances
The algorithm the researchers devised improves on traditional Bayesian optimization by better adapting to the specific experimental techniques used in chemical research and development.
“Machine learning is growing rapidly and in this paper we have shown how to apply the state-of-the-art of machine learning to chemistry. The challenge is to make sure that the mathematics actually describes a real-world problem. Thanks to our collaboration with BASF chemists and data scientists we have been able to achieve this,” said Folch.
“There are some things in chemistry that Bayesian optimization doesn't work well with,” Misener added. “We address two of them in this paper. One is 'multi-fidelity,' the fact that some data sources return more reliable data than others. The other is 'asynchronous batching,' the fact that experiments take different amounts of time and multiple experiments may be running at the same time.”
A widely used approach in chemical research and development is to conduct experiments using relatively quick and cheap approximations of the manufacturing process they are developing. For example, researchers developing manufacturing techniques for electric vehicle batteries can supplement their time-consuming and costly experiments on the pouch batteries used in the vehicles with an approximation using easy-to-manufacture coin-cell batteries.
The new algorithm is designed to recommend cheaper but less precise experiments to test parameter values āāwhere the predicted outcomes have very high uncertainty, and to recommend more expensive but more precise experiments where the outcomes have low uncertainty.
It is also designed to accommodate the fact that faster experiments return results sooner than slower but more precise experiments, avoiding idle resources by basing recommended experiments on currently available data and unknown but expected future results.
Through empirical testing using battery data provided by BASF, the researchers showed that with a limited number of experiments, the algorithm was able to identify more optimal production settings than traditional Bayesian optimization.
Real-world innovation
The research is part of a larger collaboration between Imperial and BASF, the world's largest chemical company, which aims to develop and supply advanced chemical technologies to create a more efficient and sustainable chemical sector.
The partnership recently led to the creation of SOLVE, a spin-out company founded by Dr. Linden Schrecker and with Folch as Chief Scientific Officer, which is actively leveraging novel experimental and AI techniques, such as those outlined in the award-winning CACE paper, within the chemical and pharmaceutical industries to benefit the economy and the environment.
Industry Partnership Opportunities
Companies interested in learning about collaborating and commercialising university research can find opportunities by contacting Imperial Enterprise.