Chemometrics and Generative AI: New Possibilities for Analysis Today

Rasmus Bro spoke to LCGC International about the evolving relationship between classical chemometrics and generative AI. He argues that the real impact of generative approaches lies not in incremental performance gains, but in enabling entirely new ways of integrating analytical data with vast textual knowledge. From automating the curation of complex chromatographic outputs to supporting interpretation and contextualisation of results, generative AI has the potential to transform analytical chemistry workflows. Bro also discusses trust, interpretability, and the enduring importance of domain expertise, suggesting that AI will augment—rather than replace—the chemist’s scientific judgment.

Q. Where do you see genuine, practical overlap between classical chemometrics and modern generative AI tools, and are there specific problems in analytical chemistry where large language models or generative approaches actually add something that chemometrics cannot?¹

A: I think that there are very, very many places where genAI can add to our tools. In fact, I think that is the benefit. So far, a lot of modern chemometric AI has been about how these new tools can do what we used to do but better. But that is not very interesting. The revolutionary aspect of genAI is that it now gives us completely new possibilities. Things we couldn’t dream of before. Essentially, we have comprehensive information now. We have all existing information available and in a manner that can be made operational. We already know how to identify and quantify a compound. But now we also know everything there is to know about that compound, the boiling point, where it is seen, how it mixes, how it often appears, what functional properties it can have. The list goes on. I think that we have only seen the tip of the iceberg.

Q. You mention that integrating text resources with analytical data will fundamentally change how you do food research, could you give a concrete example of what becomes possible when you combine those two data types that simply wasn't possible before?

A: We develop very advanced (if you talk to a chromatographer)/fairly classical (if you talk to a chemometrician) tools for automating the analysis of complex chromatographic data. What experienced chromatographers spend many days on per experiment, we can do in minutes—and we do not even need a skilled person to do it.

This is all well and good, but afterwards, when you have your peak table of compound names and concentration, you will have to curate that list, that is, remove items that are just column material; handle compound names that clearly make no sense. This also requires the significant time of a person skilled in not only chromatography but also in the domain knowledge the samples are coming from. It is likely that 98% of that work can be done by genAI.

And then we come to the actual thing; analysis and interpretation of the data and especially putting those interpretations in relation to what is already known. Again, genAI can play a very significant role here. It is not quite there yet, but it is only a matter of time.

Q. As models become more complex and less interpretable, how do we maintain trust in the outputs of data-driven analytical methods, and what does a trustworthy result actually look like?

A: You know, I don’t think models will be less interpretable. We are only at the beginning of a long journey and the “childish” things we are doing now are going to look naive in years to come. I think that the models will have to develop in a direction where logical reasoning and traceability will be built in when needed, as it is not always needed. So, when you ask what is trustworthy, it really depends on several aspects; some of which are not always clear even to the person asking. But as we gain experience, as we develop, and as we build new use cases, the need for accountability—whatever that will mean—will arise and will need to be developed. Right now, we are in the wild west period where it is more about doing crazy stuff and finding the boundaries. That will change.

Q. How much chemistry do you need to know to do good chemometrics, and conversely how much data science does a chemist need—where does the balance lie, and is it shifting?

A: A lot! That is our savior! It really isn’t enough to be a great data analyst. You see it repeatedly. You need to know your data science, but it is seldom the crucial part, which is sad for a person like me who develops data scientific tools. But what makes the difference is always domain knowledge. That is what makes me so confident when people think they can do better than me! But here is the plot change. AI potentially knows all known chemistry. I don’t. Working with AI I can become a much better chemometrician.

Q. How should we be training the next generation of analytical chemists to be both scientifically rigorous and data literate, and are current university curricula keeping pace with what the field actually needs?

A: As I said, we are in the middle of a hype. Let us wait until things have settled down. But basically, nothing has changed in my mind. We just got a new fancy pocket calculator and now we need to adjust everything accordingly. Certain things will be easier and other things obsolete, but the skills that a chemist will need remain the same. It’s just a slight change of focus on what the chemist has to spend the mental resources on. Creativity based on insight will most definitely still be needed.

Reference

1. Bro, R. Beyond the Hype: What Chemometrics Can Teach Generative AI. Presented at Pittcon 2026, in San Antonio, Texas. Available at: https://app.swapcard.com/event/pittcon-2026/planning/UGxhbm5pbmdfNDMwOTkxMQ== (accessed 2026-04-13).