High-Performance Liquid Chromatography (HPLC) method development is a critical step in analytical chemistry. SDLabs optimizes the multi-dimensional parameter space to find conditions that maximize peak resolution while minimizing total runtime.

The optimization landscape features a resolution vs. speed trade-off with hidden column-gradient interactions. The most obvious path — increasing runtime — gives decent resolution but misses the true optimum, which requires discovering that C18 + Step gradient is a uniquely powerful combination. Flow rate follows van Deemter behavior, and three distinct optima compete for the optimizer's attention. See the full landscape description below.

HPLC method development traditionally relies on trial-and-error or systematic grid screens. With 6 parameters and complex interactions — particularly column-gradient combinations — the search space is large. Maximizing peak resolution directly conflicts with minimizing runtime, making this a natural multi-objective problem.

The Chimera hierarchy first ensures good resolution, then minimizes runtime. Van Deemter effects are modeled in the flow rate response. Column-gradient interactions play a key role — the C18 + Step gradient combination is the optimal pairing. The global optimum requires a specific categorical combination that a grid search would take many more experiments to find.

In 35 experiments (7 iterations of 5), the optimizer reaches a peak resolution of 10.7 — 86% of the theoretical maximum (12.5) — while keeping runtime at ~45 min. Early iterations explore diverse column-gradient combinations (resolution 0.5–6.4). By iteration 2, the Gaussian Process identifies C18 as the best column and resolution jumps to 8.6. The critical breakthrough comes at iteration 6 when the model discovers the C18 + Step interaction, pushing resolution above 10.7. The progress chart shows a clear two-stage pattern: broad exploration (iterations 1–5), then focused exploitation of the optimal region (iterations 6–7). The model correctly learns that HILIC + Step is a poor combination (−15% penalty) without wasting many experiments on it.

Objective convergence — Peak Resolution improves steadily while Total Runtime is optimized within the resolution constraint:

Prediction Explanation:

Pareto front — Visualizing the resolution vs runtime trade-off across all evaluated conditions:

The core challenge is resolution vs. speed. The most obvious way to improve peak separation is to increase the gradient runtime — but this directly increases total analysis time. An analyst doing trial-and-error would quickly find a decent resolution at 50+ minutes and stop there, missing the much better solution hiding at a specific column-gradient combination.

Column-gradient interactions are the key. Not all column-gradient pairings are equal. C18 with a Step gradient gives dramatically better resolution than any other combination — but this interaction is invisible unless you test it. Meanwhile, HILIC with a Step gradient performs poorly. The optimizer must test enough combinations to discover these hidden interactions.

Flow rate follows van Deemter behavior. Very low flow (0.3–0.5 mL/min) gives the best peak efficiency, but pushing flow below 0.3 offers little additional benefit while substantially increasing runtime. Higher flow rates (>1.5 mL/min) degrade resolution and can cause co-elution when combined with high organic eluent content.

Three distinct optima exist in the landscape: the global optimum at C18 + Step with ~40 min runtime and ~55% organic (resolution ~12.5), a local optimum reachable by simply running long Linear gradients (resolution ~9.5), and a secondary local at Phenyl-Hexyl + Concave (resolution ~8.0). The optimizer must avoid the "just increase runtime" trap to reach the true best conditions.