ߣߣ��Ƶ

Can we use artificial intelligence (AI) to make chemistry greener? How does that work, and where do we start?

To practising chemists, green and sustainable chemistry is more important than ever. As global resources continue to diminish, we need to consider the life cycle of chemical materials and the impacts they can have on our health, safety and environment. 

As a guide to help chemists move towards more sustainable ways of working, scientists Paul Anastas and John Warner published The 12 Principles of Green Chemistry in 1998. The book outlines the challenges involved in making chemical reactions greener, and how to address them.

• Greener labs don’t need bigger budgets – just better habits
• Go green, AI!
• How can we decarbonise knowledge production?

However, making chemistry greener is not always as easy as swapping out a bad chemical for a good chemical. Small changes to reaction components or procedures can have a substantial effect on the outcome of a reaction, and in the busy life of a researcher, identifying and testing these changes often falls by the wayside. 

Our explores how AI systems can be used to solve this problem. AI can spot non-intuitive trends from large amounts of data, providing insight into how to make reactions greener. 

Isn’t AI bad for the environment?

It’s true that the data centres hosting the most popular large language models (LLMs) consume huge amounts of electricity and water, because of the enormous processing power these models need. However, while every LLM is AI, not all AI is an LLM. There are many examples of models that have much smaller energy requirements, but are just as useful for identifying greener chemical reactions. These models can be trained and run on basic hardware, such as a laptop. 

Specificity over generality

When solving a challenge like green chemistry, it helps to have many tools to accomplish different tasks. Identifying well-defined single problems makes it easier to validate models and convince the scientific community of their worth. It also helps to produce more insightful models, allowing focus on specific nuances in smaller datasets. 

These tools can feed into larger AI ecosystems called agents, where a central decision machine can deploy many individual tools to solve complex problems without sacrificing specificity. 

So, what problems can we solve?

Solvents account for more than half the waste produced when making pharmaceutical ingredients, and most common solvents can be harmful to health or the environment. Replacing these solvents can have a big impact upon sustainability, but solvents have significant effects on reactions which are difficult to model and not always well understood. Use AI tools to find sustainable solvent alternatives, and in nearly all cases, there are replacements available that mimic important solvent properties.

Another big win for sustainability is to reduce the number of experiments run in the lab. The fewer experiments, the less energy we use and the fewer natural resources we consume. Again, there are several ways to do this, from statistical modelling with design of experiments or Bayesian optimisation, through to chemical reaction yield prediction and computer-aided retrosynthesis. 

One of the 12 principles of green chemistry, as outlined by Anastas and Warner, is to make molecules “benign by design”. Consider the environmental and health impacts of potential molecules before making them in the lab! 

AI tools can help here by predicting molecular properties such as toxicity or global warming potential. To consider the full life cycle of molecules, conduct life cycle assessment, which is a thorough study of all the environmental impacts of a process. This can be complicated and time-consuming, so leveraging AI tools in this context can help to provide a bigger picture. 

The human side of the problem

More and more AI models are published in the field of chemistry every year. However, the chemist in the lab is yet to feel the benefits of these. Models are rarely published with accessible interfaces for non-experts to use, and this is a real barrier. 

Building simple, intuitive interfaces encourages use of AI tools beyond publication, but even in these cases, it can be hard to reach into the laboratory. How can we encourage chemists to engage with AI tools in their day-to-day work, to help them identify greener choices?

The answer is to build these tools directly into their digital workflows, instead of necessitating another software platform. The most obvious place to get this benefit would be in the Electronic Laboratory Notebook or Laboratory Inventory Management System environments, where chemists plan and design their experiments. Placing these tools directly in the chemist’s path will surely encourage use. 

What next?

The future looks bright for green chemistry and AI. More models are being built and published than ever before and these can augment decision-making in chemistry. This expanding field has, however, highlighted a major flaw in the capture, storage and publication of chemical data. 

The chemical literature is highly biased towards positive results, often omitting important experimental details, which make results difficult to reproduce. These issues highlight the need for clean, robust and balanced chemical data to train AI models. Individual researchers should embrace the good practices that will provide such data. Publishers and funding agencies also have a part to play in driving the required shift in culture. 

To bring chemistry along in the AI revolution, we desperately need to change our approach to publishing chemical reaction data.