The Future of AI/ML in Liquid Chromatography

An “AI winter” refers to a period of decreased funding and interest. It is more or less a period of disillusionment, typically as a result of overhyped promises that could not be kept, causing funding to dry up.

In the 1980s, veteran AI researchers like Marvin Minsky (co-founder the Massachusetts Institute of Technology’s AI laboratory) cautioned that public and corporate excitement around AI was growing unsustainably. They foresaw a chain reaction where endless expectations from the public, along with academic pessimism, would trigger skepticism in the media, leading to a sharp drop in funding and ultimately a collapse in progress. Their prediction proved accurate as just a few years later, the once-thriving AI sector faced a dramatic downturn, with research momentum stalling as investments dried up.

Going back to our own chromatographic and analytical science communities, I think we can all agree that there is a hype surrounding AI right now. The above history lesson is therefore very relevant. Most of us see large-language models such as OpenAI and Copilot and start imagining endless possibilities. In this light, I am not too worried about our activated imagination, because we need inspiration to develop the next big thing. However, I am worried about the pressure on us as developers, instrument manufacturers, lab scientists and researchers to introduce new methods based on AI.

The fear of missing out (FOMO) drives people to find a problem that ML – a probabilistic, data-driven AI – can solve, in contrast to arriving at ML as a solution to an existing problem.

The result of the former is that typical feeling of disappointment, when seeing a presentation highlighting the great promise of AI that does not meet the expectations.

If I now sound negative about AI, then please let me correct myself: I am very optimistic and a firm believer of the promise of AI. But I do think the past teaches us that we must advance this sustainably.

There is also an irony, because many powerful ML methods have been introduced into chromatographic data analysis strategies by the chemometrics community long before the emergence of the large-language models, yet are often not associated with AI. In other words, if you have been missing out, you have been doing so for decades.

Recent advances have had a profound, but, so far, uneven impact on LC data processing. While the promise is clear for improved peak detection, more accurate integration, retention and selectivity modeling, retention-time alignment and even automated method development, the actual implementation in daily practice remains limited.

Two of the key hurdles are the quality and scarcity of input data: ML models are only as good as the data they are trained on. In that sense, AI has helped expose the fragility of traditional data processing workflows. We're now at a point where AI is not just a tool for automation, but also a lens to reassess the robustness and reproducibility of our chromatographic data pipelines.

I am generally most excited by the explorative studies that investigate the applicability of various ML techniques to the problem of method development. A case in point is the assessment of reinforcement learning to improve method development in LC by the group of Deirdre Cabooter.

A central limitation is the lack of ground-truth data, which makes it difficult to evaluate the performance of peak detection or retention-time alignment algorithms. Without objective benchmarks, it's hard to know if improvements are real or case-specific.

Another challenge is data complexity: two-dimensional liquid chromatography (2D-LC) and LC×MS generate vast, high-dimensional datasets that often include artifacts, missing values and overlapping peaks. Many algorithms break down under these conditions. Finally, there’s the issue of generalizability; an algorithm that works well for one dataset might fail completely on another. That lack of robustness is a critical bottleneck for automation and machine learning adoption.

In the context of the previous answer, another development I find exciting is the emergence of realistic data simulators tailored for chromatography. If there is one strength our community has developed, then it is that we chromatographers have a well-developed and deep understanding of retention in various LC modes. The simulation tools allow us to leverage this knowledge to generate synthetic yet plausible chromatograms, complete with noise, baseline drift and realistic peak shapes, which can then be used to benchmark signal processing algorithms in a controlled, reproducible way. This is transformative because it finally enables objective and meaningful comparisons between methods, while also helping us to improve the efficiency of automation using ML, because these simulators can alleviate the huge data consumption that ML models need.

Closely related is our own development of multi-fidelity Bayesian optimization and AI-enhanced retention modeling, which combine mechanistic knowledge with data-driven techniques. On the data analysis side, probabilistic algorithms, those that incorporate uncertainty in peak identity or position, offer a path toward more robust peak tracking and alignment. 

AI and ML will increasingly serve as integrated decision-support tools, not as black-box replacements for analysts. Their role will be to assist with complex tasks, like peak deconvolution, outlier detection and experimental design, where traditional rules-based approaches fall short.

We’ll see a shift from expert-driven method development to hybrid models, where human intuition is complemented by data-driven optimization. Eventually, I expect ML to be embedded directly in instruments, enabling adaptive chromatographic methods that adjust in real time. But this requires trust – and that trust can only be built on transparent, validated models.