.png "ACSGCI"){kind=logo}
By Ashley Baker, Scientific Content Manager (Contractor), ACS Green Chemistry Institute
In light of our upcoming Green Chemistry and Engineering Conference theme, “AI-Enabled Green Chemistry,” we spoke with conference co-chairs Jared Piper, Director of Process Chemistry at Pfizer, and Alexei Lapkin, Professor of Sustainable Reaction Engineering at the University of Cambridge, about how computational tools are rapidly accelerating green chemistry and engineering innovations.
Q. Why have a green chemistry conference set around the theme of artificial intelligence (AI), and what’s the relationship between these two fields?
A. Alexei
The entire field of chemistry is transforming very quickly and adopting digital research technology and tools that are increasingly accessible.
Let’s say my company has a chemical manufacturer as a client. They might ask us to design a new molecule that has a specific function, is non-toxic, and is biodegradable. They might also ask us to design a synthetic path to that molecule from cheap reagents with no hazards and no patent protections, and they want it delivered within a few months. In the past, to do this manually would be a program of several years of work. Now, we are doing this work with tools from AI, computer aided synthesis planning, and predictive models for properties. So, this client’s ask is becoming feasible.
In a few years, these tools will be much more ubiquitous and much better performing. When we started to develop the earlier tools in an academic context, let’s say tools for synthesis planning, we didn’t even know what to ask from them. But once we showed them to people, particularly companies, chemists started saying, “Oh, if I have such a tool, I could do this.” Then we went back and started adding more and more functionality to our tools.
One application that I’m particularly keen on is inverse design. We are trying to convince industry at large to shift to a different business model: to sell functions rather than molecules. In this case, you need to do inverse design. You see more and more publications on inverse design in materials, catalysts, and functional molecules, and this is critical because we can set really complicated functional descriptions. It’s not about fitting whatever molecule you currently have into products even if it doesn’t function very well just to sell it. With inverse design, if I want a polymer that performs in a certain way but also has a specific profile at the end of life – either it is a new raw material, or it biodegrades - I can build in circularity, or I can build in how the polymer would eventually be downcycled or upcycled. Then, the sustainability transition becomes much easier because you can plan for how much of your material will be in what use scenarios and what kind of industry will handle it later. But without machine learning (ML) and AI models, inverse design is often completely impossible. This is what AI offers. That, for me, is the huge and important target to go for.
A. Jared
It’s been a while in the making. This is the fifth year that we’ve had an AI session at the GC&E where we demonstrate prediction, machine learning, and other tools that can help accelerate green chemistry. Now it’s blossomed into the whole theme of the meeting which is great to see. Although a few community members have been involved for several years, we now see a broader audience understanding that tools have arrived that allow us to do design for green, and it will become more advanced as we move forward.
People in this field also understand that AI/ML tools won’t solve every problem. It’s not a magic solution. Our vision is that it will accelerate development and remove or minimize unnecessary experimentation. If you have three synthetic routes to make a molecule, one of them has the most potential to be greener than the others based on the kind of bond disconnections, the types of solvents you use, whether you can use an enzyme, and so on. So, you can start by optimizing the route the tools have already predicted to be better, and that prevents a lot of wasted resources and effort.
Q. With Jared’s pharmaceutical experience and Alexei’s academic research, you bring two distinct perspectives. How do you each use or envision using these tools in your respective fields? What areas are impacted the most by their use?
A. Jared
There are many examples of using computational tools to optimize chemical reactions. The number one thing is that we can optimize the selectivity or the yield with fewer experiments and in a shorter timeframe, and we can be confident that we’ve reached its global maximum. Using initial data gathered for a given transformation, these sophisticated algorithms and tools can tell you it’s as good as it’s going to get.
We also have a way of predicting how green a process can be based on statistical models using real data carried out on scale (>1 kg), which is known as the Process Mass Intensity (PMI) Predictor and is available on the ACS GCI website. Pharma has also been using AI to predict biological targets that would potentially lead to viable therapies. The tools run the gamut from discovery through development, all the way to how we implement processes in a manufacturing setting. So, using them, we can predict how a process would work in our plants, how to make changes, how to make the greenest process, and how to get the most benefit for our company. You can optimize reactions based on anything you want, whether it’s cost, PMI, or safety.
This leads me back to Alexei because we’re trying to move into predictive life cycle assessment (LCA) as well. That one is much murkier for us in development; it’s not until later in a product’s life cycle that we understand the full impacts of our chemical process decisions.
A. Alexei
Yes, the LCA prediction is a big target for many research groups. LCA is a pain, it’s highly inaccurate, and it’s really expensive. Computational methodologies would allow LCAs to be much cheaper and more reliable. AI and data science are absolutely essential here.
There are research groups working on predicting life cycle impacts for both new molecules and their manufacturing routes. Through our Cambridge Centre for Advanced Research and Education in Singapore, there is a project that is developing hybrid models to support LCA for small molecules manufacturing. We would not have been able to set these targets without machine learning. You cannot do it with physical models alone, or small data sets; it’s simply not possible.
There are several areas where the community is using machine learning that are specific to green chemistry. Predicting biodegradability, for example, is a hugely important target, and again it’s not possible with solely mechanistic models. They have to be complemented with data-driven models. Predicting toxicity is another big target. So, there are really important targets for green chemistry that can’t be solved without AI and data.
Q. Is there a skill gap in people who can create and implement these computational tools? Are there enough educational resources available to enable chemists to solve problems using them?
A. Alexei
There are certainly not enough people who can do this at the moment. All of our Ph.D. graduates who have had this training are snapped up by companies, start their own businesses, or become professors. Basically, there are no unemployed people who have studied this. Moving forward, this means changing chemistry and chemical engineering education so it covers skills in coding, ML, and AI, while still focusing on core domain knowledge.
You can’t do this kind of work by just putting a chemist and a computer scientist in one room. You need a computer scientist to understand chemistry and a chemist to understand a little bit of computer science. You don’t need to be an expert, but you need in-depth knowledge in your core field so you can set the challenges and the problems correctly and explain them. But, you also need to understand what the other side can do. At Cambridge, there is post-graduate training through the “SynTech” Centre for Doctoral Training specifically to ensure that chemists have Ph.D. projects that require them to learn some coding, or to learn how to use automation and high throughput experimentation.
This field has gone through a phase where it was extremely hard to start doing this work, particularly for chemists because they have had to learn how to code. But we needed chemists to go through this initial pain to understand the potential and learn how to speak with computer scientists. Recently, there has been an emergence of a bunch of startup companies offering no-code environments for chemists. In the next few years, the situation will change. Chemists will need to understand what they can ask of the tools. They will still need to stay in touch with the computer scientists, but they won’t have to do it all by themselves. But for now, academia worldwide is struggling to find professors who can teach or do research in this field. There are simply not enough people. Likewise, I think Jared can comment on how “easy” it is to find people who can do this in industry.
A. Jared
That’s perfectly spot on. We have the fortunate occasional chemist who has a strong computer science background and can code or understand the models, but they are not in great supply. More academic programs are adding these requirements and training the emerging graduates of chemistry in this field, and there are some that offer the upskilling of existing chemists in industry.
One great thing has been that this community of people embracing AI is always sharing with and trying to learn from each other. It’s a great community of chemists, but it’s far short of what we need. This is especially true as we’re going forward with more and more of an understanding of where we can apply these tools and the power of the tools that we have. So, we start to see more opportunities, but again, only a small number of people are trained in this field. More are needed. If nothing else, I would say we need more data scientists to get into chemistry.
There are also people out there who can create more user-friendly interface systems so new adopters of the technology do not have as high of a hurdle. You need someone to create that easier interface first, and more chemists will start adopting it.
Q. Do you foresee a challenge in getting people to think about the potential of computational tools rather than the difficulty of letting go of conventional processes?
A. Alexei
Yes, the only way to do it is to show people they can make more money with the new business model than they made before. There are plenty of examples in business literature that the most rapid transition to anything new is to change the business model. If you’re just tinkering with the way you make a particular molecule, it doesn’t matter. However, if you change the business model, the impact is significantly higher, and that’s what we need to do.
There are innovation principles and forecasting methods for technologies, and one of them is that an ideal system that delivers a function does not have a physical embodiment – has no associated costs. That’s the goal, and everything evolves in that direction, such as your communication, your transportation, everything. Chemistry is no different. People will find ways of making molecules in increasingly smaller / local facilities, only when and where you need them.
A. Jared
Also, we all know that pharma and many other companies have net zero goals for the immediate future. I’ll be honest, they are very lofty goals, and it’s not always clear how we’re going to get there. There’s not much time left; some of these goals are for 2040 or 2050. It’s happening quickly. So you absolutely need to have these types of tools to understand the optimal steps to move forward.
Q. What would you say are the pitfalls of these new technologies that you would caution chemists against?
A. Jared
One pitfall is that we’ve had a lot of great chemistry over the years, but it’s been told in a story format, not in a data-driven format. For example, I didn’t get to publish papers unless the yield was over a certain percentage or the reaction worked as advertised. We didn’t broadcast the ones that didn’t work, the ones that were roadblocks. However, that negative data is critical to complete the data sets and understand where the computational models can help. We’re still building an understanding of what data to collect, how to collect it, ensure it’s a complete set, and that it’s high quality.
A. Alexei
This is a very active sub-field of research: understanding the impact of the data sets, the distribution of the data in the data sets, and the impact of unknowns that are not in the data sets. An example that is almost always discussed in this context is the impact of the order of addition of reagents. This is rarely recorded in lab journals, and it’s not reported in papers or patents. But it may impact the outcome of the transformation. This is one of the things that is currently impossible to capture and include in a model, so we’re still figuring out how to handle it. But these are at least known unknowns. There are also unknown unknowns, and we also need to figure out how to deal with these.
Q. To close and bring us back to the upcoming GC&E conference, what are you most excited about, and what new perspectives do you hope attendees walk away with?
A. Alexei
Now, the number of pioneering experts in this field is sufficiently large, the tools are becoming more accessible, and the industry has bought it completely. It’s time to lift up the rest of the community, get everyone on the same page, and show that this is no longer something strange and weird. It is a new research methodology, and it needs to be taught, it needs to be used, and it needs to be used correctly. We need to start posing new research questions that we couldn’t before. There is some recalibration needed as to what we should be setting as fundamental scientific targets for the field.
A. Jared
We’ve been building momentum on this topic as a community for a long time within the green chemistry world. Very early, before the pandemic, I said we needed to have a team focused on this and the Green Chemistry Institute was very supportive. They’ve helped bring us together as a community, and I’m really excited that this is finally the theme of the meeting. But I’m even more excited about how people have responded and how many will attend. The workshops and other hands-on events will do nothing but continue to build the community around in silico tools and bring visibility to opportunities in green chemistry. It’s going to help connect people who know how to solve problems with people who have urgent problems that need to be solved.
________
There’s still time to register for the 28th Annual Green Chemistry and Engineering Conference with the theme, “AI-Enabled Green Chemistry,” taking place June 2-5, 2024 in Atlanta, GA. Don’t miss the re-vamped Student Experience, which includes the Green Chemistry and Engineering AI Hackathon, organized by post-graduate students from the University of Cambridge.