Key points
Our CEO recently came back from a conference of European Chemical Industry executives in Germany with a very clear message. Across many conversations, three themes emerged as fundamental current concerns for the sector.
First, unsurprisingly, was how to manage dramatic increases in the cost of energy.
Second was sustainability. How can businesses hit ambitious targets to lower the environmental impact of products and processes?
Finally, how can chemical process organisations innovate more rapidly and effectively – both in response to those other two factors, and to win in a competitive market where margins are always under pressure?
Can Machine Learning help?
At Intellegens, we’re a machine learning software company (originally a University of Cambridge spin-out) that works closely with the sector. So, what interested us most about these challenges was, of course, how machine learning approaches might help.
Machine learning is a branch of Artificial Intelligence (AI) in which computational methods learn from existing data to build mathematical models of a system that can then be applied to understand, predict, and optimise the system’s behaviour.
Although the terminology around AI smacks of science fiction, you can think of machine learning as simply a very smart data analysis technique. Among the smart things about it are that it starts with no assumptions about the data being analysed and just learns from that data.
This means that it often finds relationships that are missed by other analytical methods and are not obvious to humans. It can also model very complicated, subtle (‘non-linear’) relationships.
How is this relevant to those chemical process industry challenges? Responding effectively is mostly about optimising products, often chemical formulations of various types, and their associated processes. Most importantly, this response needs to be rapid as circumstances change.
Often, R&D organisations are not short of data on these formulations and processes. They have experimental results from past projects, production and quality assurance data, simulation and modelling results, and information from suppliers.
But it can be hard to ask new questions of data that was not gathered or structured for the purpose of answering those questions. And homing-in on solutions to formulation and process problems usually requires new experiments, which are usually costly and time-consuming.
Machine learning, with its ability to ingest data and exploit hidden patterns within that data, should be able to help, both by extracting more value from legacy data and by providing guidance for new experiments.
Real-world Machine Learning examples
The good news is that there are plenty of examples where this has indeed happened. Machine learning has proven itself able to find beneficial changes to ingredients or processing parameters while reducing the amount of experiment needed to validate those changes. Let’s consider an example relevant to each of those three big challenges.
On energy, a recent machine learning project was delivered through the ‘Build UK-India’ project, a collaboration driven by UKRI, Godrej, Tata, and Johnson Matthey. The project used machine learning to find routes to reduce waste, and thus energy consumption, in a manufacturing process.
Sustainability has been a factor in work at Domino Printing Sciences to develop new ink formulations, with machine learning speeding up the search for alternatives to regulated chemicals by recommending which experimental routes would be most productive.
Faster innovation was demonstrated by a recent project in which the University of Sheffield Advanced Manufacturing Centre and its industrial partners used machine learning to reduce by over 50% the amount of experimentation required to develop use of a new additive manufacturing material.
There are plenty more examples. Researchers at the University of Cambridge have designed new concrete materials and a team at IFF and Optibrium predicted the sensory properties of chemical compounds, an important factor in developing flavours and fragrances. In both cases, machine learning has shown how it can reduced reliance on experimentation.
Challenges to overcome with Machine Learning
Despite this potential, machine learning is far from ubiquitous in chemical product and process development. Many organisations are still figuring out how to get full value from the technology. This is because there are several challenges that need to be overcome in deploying and applying machine learning in practical experimental or process applications.
Machine learning often works poorly when the input data is sparse (that is, it has gaps in it) or noisy (perhaps due to errors or inherent variability in measurements). Most methods require training data that is complete and clean for that initial stage of the process, when the machine learning builds a model by learning from this data.
This is not realistic in many practical scenarios. For example, if you are trying to learn from legacy data gathered over multiple projects that may have been conducted for different purposes, it is unlikely that every possible property of interest will have been measured in a consistent manner in every project.
Even if a good model can be built from real-world data, then a degree of uncertainty is inevitable – it will not deliver perfect predictions. This should not be a problem. We can make good decisions based on imperfect information if we have a feel for the uncertainty in that information.
We need to know which data points we can trust the most, and which are more like guesswork. So, it is important that the machine learning methods we use deliver a robust quantification of the uncertainty in its predictions.
It can also be true that machine learning is difficult to use. It often requires skills in coding. And results can be difficult to interpret. A machine learning model is not an equation, which can be inspected to understand how the variables and parameters relate.
Good ‘explainable AI’ tools are needed to dig into what the model is telling us about the system being studied. One way to do this is to use the model to compute what happens to system outputs as the inputs change and to then plot the results. This enables users to understand what the model is telling us about which inputs most impact the outputs. The importance chart shown below is an example of such an analysis.
Importance chart shows which inputs to a polymer formulation have the greatest impact on properties.
Realising the potential of Machine Learning
If you are considering using machine learning in formulation or process development, you need to consider these possible roadblocks. Find a method that can handle sparse, noisy data. Make sure it offers robust quantification of uncertainty in its predictions.
And consider the usability of software that delivers the method and the tools that it offers for extracting useful information from the results. We’ve already seen how, when these factors are navigated, machine learning can help to accelerate innovation and meet product performance, energy, cost, and sustainability goals.
It’s also interesting to think about how the technology could reach beyond R&D. For example, some leading chemical businesses are already looking at how to embed machine learning methods into their production processes. An example might be to enable manufacturing teams to respond in real time to changes in process inputs.
For example, if a plastic manufacturing process relies on recycled feedstock that is subject to variations in its quality and composition, the machine learning model could be used to suggest changes in the other ingredients and process parameters to keep the properties of the product within tolerance.
It seems inevitable that AI in general will have a great impact on the chemical process industries. As a pragmatic and useful analytical tool that can be implemented today, machine learning will be at the sharp end of this impact. This article has offered some tips for those looking to get started. It will be interesting to see where we go from here.
