CHROMATOGRAPHY
Figure 3- Pareto front between analysis time & resolution design workflow using Bayesian optimisation, setting a new benchmark in HPLC method development. The procedure operates as a fully closed-loop set-up with the optimiser autonomously taking its decisions and adjusting the gradient.
The motivation is clear: simplify complexity, save time, and reduce costs. With this closed-loop system, optimisation happens overnight without human intervention, freeing our scientists to focus on innovation rather than repetitive tasks.
BoFire optimisation workflow
At the core of this technology lies BoFire, an open-source Bayesian optimisation engine connected to the HPLC system. After each run, chromatographic data flows back into the optimiser, which then refines gradient profiles automatically( Figure 1).
Multiple parameters, including the gradient programme( solvent composition and timing), are adjusted based on performance metrics. A custom objective function called the sum of capped resolutions guides the process to ensure adequate peak resolution. The optimiser searches for the maximum of this function within the design space initially provided by the operator.
Multi-objective optimisations
What makes this approach truly powerful is its ability to handle multiobjective optimisations. Instead of focusing on a single goal, such as maximising the separation of a sample, the approach can also minimise measurement time while doing so, delivering faster runs with more resolved components.
The objective in the example discussed here was to optimise the separation of a pharmaceutical ingredient from a complex mixture comprising multiple impurities while keeping the analysis time short( Figure 2). Due to multiple targets, the optimiser cannot find one global minimum. None of the results represent the optimum for both resolution and analysis time.
Instead, Pareto ordering needs to be considered. This method compares options based on multiple criteria without collapsing them into a single score. An option is‘ Paretodominant’ if it is at least as good as another option in all criteria and better in at least one. This helps to identify dominant and non-dominant( Pareto optimal) solutions. The result is a Pareto front, where the tradeoff is shown between measurement time and quality of separation, as exemplified in Figure 3.
In the given example, the optimiser performed five runs to predefine the design space and 17 optimisation runs to find the Pareto-optimal conditions with a maximum for the sum of capped resolutions and an acceptable analysis time of 14.6 minutes. The experiment was conducted completely unsupervised and completed overnight without interfering with other tasks that used the set-up during the day shift.
The benefits are tangible: significant time savings through overnight automation and cost-efficiency thanks to reduced solvent use and instrument hours.
Summary & outlook
Chromatographic separations play a crucial role for Evonik in the production environment as well as the analytical space. We have established expertise in large-scale separations and continuous chromatography. Looking ahead, Evonik plans to apply this concept to preparative chromatography, where optimisation could significantly shorten development cycles and increase production efficiency.
Closed-loop automatic gradient design is not just an upgrade; it is a paradigm shift. By combining automation with Bayesian optimisation, Evonik and Agilent have transformed a labour-intensive process into an autonomous workflow that delivers faster, better and more cost-effective separations. ●
* Also contributing to this article were Christian Haas, data scientist, and Johannes Peter Dürholt, senior data scientist, at Evonik
References: 1: J. Boelrijk, B. Pirok, B. Ensing & P. Forré, Journal of Chromatography A 2021, 1659, 462628 2: J. Boelrijk, B. Ensing, P. Forré & B. W. Pirok, Analytica Chimica Acta 2023, 1242, 340789
J j
Sebastian Detlefsen
GROUP HEAD CHROMATOGRAPHY & PURIFICATION
EVONIK sebastian. detlefsen @ evonik. com www. evonik. com
MAR / APR 2026 SPECCHEMONLINE. COM
63