Guiding Machine Learning Algorithms in Automation for LC-MS Method Development: A Closer Look at Bayesian Optimization Kernels

Artificial intelligence is rapidly reshaping how we design, optimize, and evaluate analytical methods. In liquid chromatography-mass spectrometry (LC-MS), machine learning approaches, particularly Bayesian optimization (BO), offer the potential to revolutionize method development by accelerating decision-making, improving analytical Quality by Design (AQbD) practices, and enabling more systematic exploration of experimental space. Recent advances with the AutoLC platform have shown that fully closed-loop, unsupervised method development is not only feasible, but can operate robustly without human intervention, highlighting the growing maturity of AI-driven workflows.

At the heart of BO lies its kernel: the mathematical structure that encodes assumptions about chromatographic response surfaces and governs how the algorithm learns from data. Despite its importance, kernel choice is often treated generically, even though LC-MS method parameters follow well-defined physicochemical relationships that could be explicitly leveraged. In this talk, we will explore how different kernel families, default, ANOVA-style, stick-breaking, and entropy-search portfolios, shape optimization behaviour in automated LC-MS workflows. We discuss how domain knowledge, chromatographic response functions, and mechanistic constraints may be incorporated into kernel design to better reflect realistic method-development landscapes. The presentation outlines a framework for evaluating kernel suitability and highlights how knowledge-guided kernels may enhance future AI-driven method development systems.