Use Case: Hydroformylation of Vinyl Arenes

Rhodium-catalyzed hydroformylation converts alkenes into aldehydes and is a cornerstone of industrial chemistry. In collaboration with dsm-firmenich, SDLabs optimized the hydroformylation of vinyl arenes to maximize both conversion and linear (anti-Markovnikov) selectivity while dramatically reducing precious Rh catalyst loading.

The optimization landscape presents a fundamental conversion vs. selectivity trade-off: higher temperatures boost conversion but erode linear selectivity through isomerization. Ligand bite angle is the dominant lever — wide-bite-angle ligands like XANTPHOS deliver both high conversion and selectivity, but this must be discovered from the data. Pressure and Rh loading offer further room for optimization, with the key finding that catalyst loading can be cut 10–30x without sacrificing performance. See the full landscape description below.

Traditional catalyst screening tests every ligand/temperature combination exhaustively. The original study explored a 7-dimensional search space across 88 experiments. The goal: maximize conversion while maintaining high linear (anti-Markovnikov) selectivity — two objectives that often conflict.

The Chimera hierarchy handles the conversion-selectivity trade-off. Higher temperature boosts conversion but reduces selectivity. XANTPHOS (wide bite angle 108°) gives the best overall performance. BIPHEPHOS gives the best selectivity but lower conversion. The optimizer navigates this landscape automatically, balancing the two objectives according to the specified hierarchy and tolerances.

In 25 experiments (5 iterations of 5), the optimizer reaches 95% conversion and 90% linear selectivity. The Chimera hierarchy drives a clear two-phase behavior: early iterations focus on conversion (the top-priority objective), then once conversion stabilizes within tolerance, the model shifts to improving selectivity. The Gaussian Process learns the ligand bite angle → selectivity relationship after just 10–15 data points, correctly identifying XANTPHOS and BIPHEPHOS as the best ligands without being told their chemical structure. The progress chart shows conversion plateauing by iteration 2, followed by a selectivity jump as the optimizer exploits the learned landscape.

Objective convergence — Conversion stabilizes first, then Linear Selectivity improves as Chimera shifts focus:

Ligand exploration — The optimizer learns which phosphine ligands perform best and focuses on XANTPHOS and BIPHEPHOS:

Parallel coordinates — Visualizing the trade-off between conversion and selectivity across all evaluated conditions:

The core challenge is a conversion vs. selectivity trade-off. Raising the temperature above 95 °C boosts conversion but promotes isomerization side reactions that erode linear selectivity. The optimizer cannot simply crank up every parameter — it must find the sweet spot.

Ligand choice is the dominant lever for selectivity. Wide-bite-angle ligands like XANTPHOS (108°) and BIPHEPHOS deliver the highest linear:branched ratios, while small-bite-angle ligands like DPPE (78°) give poor selectivity regardless of other conditions. Importantly, the best ligand for conversion (XANTPHOS) is also one of the best for selectivity — but this is not obvious upfront, and the optimizer must discover it from the data.

Pressure has a moderate effect on both objectives. Conversion responds sharply between 5 and 8 bar, then plateaus. Selectivity peaks around 12 bar where CO insertion kinetics favor the linear product. The optimal pressure window (10–12 bar) is a compromise the optimizer finds naturally.

Rh loading can be reduced dramatically without losing performance. Above ~0.003 mol%, additional catalyst gives diminishing returns for conversion and slightly hurts selectivity. This is the key resource-efficiency finding: the optimizer confirms that 10–30x less Rh is sufficient.