Use Case: CO2-to-Methanol Catalyst Formulation

In collaboration with ETH Zurich's SwissCAT+ hub, SDLabs guided a closed-loop, AI-driven campaign to find the optimal heterogeneous catalyst for converting carbon dioxide into methanol — compressing a century of catalyst discovery into six weeks of automated experimentation.

The challenge is navigating a space of more than 20 million catalyst combinations, where the most obvious metric — high CO2 conversion — is actively misleading: iron-rich catalysts convert CO2 readily but waste almost all of it to methane. The real optimum is a low-loading copper-zinc-cerium formulation that is simultaneously the most selective, one of the cheapest, and invisible to any single-metric screen. See the full landscape description below.

Heterogeneous catalyst development for CO2 hydrogenation is notoriously slow because the search space is large, multi-dimensional, and filled with deceptive local optima. Eight metals (Cu, Zn, Ce, In, Fe, Mn, Ga, Re) can each be loaded onto a ZrO2 support at any fraction, yielding over 20 million distinct formulations. Conventional one-at-a-time screening takes one week per catalyst; covering even a small fraction of the space is impractical. Furthermore, optimizing conversion alone is counterproductive — highly active iron catalysts produce methane, not methanol. A genuine solution requires simultaneously maximizing selectivity, maximizing conversion, minimizing methane, and minimizing catalyst cost.

All eight parameters represent weight-percent loadings of a metal on a ZrO2 support. ZrO2 fills the remainder, so the total metal loading is constrained to ≤ 27 wt%.

Chimera's satisficing hierarchy is essential here: optimizing conversion alone steers toward iron-rich formulations that produce methane. By placing methanol selectivity at the top of the hierarchy, the optimizer is told — explicitly — that productive conversion is worthless if the wrong product is made. Once selectivity is satisficed, the model then pursues CO2 conversion, then suppresses methane, and finally trims cost. This mirrors how an experienced catalysis chemist would rank the objectives, but is applied automatically and consistently across all 24 candidates per batch.

The Gaussian Process model learns across all eight metal dimensions simultaneously, identifying synergies (Cu-Zn, Cu-Ce) and antagonisms (Cu-In, Fe-MeOH selectivity) from just a few batches of data. This cross-dimensional learning is what makes BO far more efficient than any factorial design in this space.

The optimizer starts with no prior knowledge and explores chemical space broadly in the first one or two batches, testing diverse metal combinations including pure-support baselines and single-metal formulations. By iteration 2 it has learned that copper-containing catalysts are far more promising than iron-rich ones, and it begins concentrating recommendations in the Cu-Zn region. Iteration 3 introduces the first Ce-doped candidates as the GP learns the value of rare-earth promotion at trace loadings. By iteration 5, the best observed catalyst is near the global optimum, matching the century-old industrial discovery.

Chimera's two-phase behavior is clearly visible in the progress chart: methanol selectivity climbs steeply in the first three batches (the primary objective dominating), then levels off while CO2 conversion continues to improve as the secondary objective takes over once selectivity is satisficed. Metal cost drops steadily as the model learns to avoid indium and rhenium. See the progress chart for the actual convergence curves.

Objective convergence — Methanol selectivity climbs first, then CO2 conversion improves while methane and cost fall:

[Screenshot: Multi-objective convergence chart showing methanol selectivity rising steeply in early iterations followed by CO2 conversion improvement]

Parameter exploration — The optimizer narrows from broad metal space down to the Cu-Zn-Ce region over six generations:

[Screenshot: Parameter recommendation chart showing metal loadings converging toward Cu-Zn-Ce formulations over iterations]

Parallel coordinates — All 144 catalysts plotted across all four objectives, revealing the trade-off between conversion and selectivity:

[Screenshot: Parallel coordinates chart of 144 catalyst experiments showing Pareto front between methanol selectivity and CO2 conversion]

The most important feature of this landscape is that the single best metric — CO2 conversion — is actively misleading. Iron is the most active metal in the set, converting up to 30% of CO2 at just 4 wt% loading. But over 90% of that conversion produces methane, not methanol, making Fe-rich formulations worthless for the actual goal. Any single-metric screen on CO2 conversion alone would lock onto iron and never escape.

Copper is the true primary active metal for methanol synthesis. Its activity follows a clear bell curve centered around 1.85 wt%: below 1 wt% there is too little active surface, above 3 wt% over-hydrogenation sets in and methane formation rises. This narrow optimum is invisible to coarse screening grids.

Zinc is the classical co-promoter for copper. Adding 0.5–1 wt% Zn stabilises copper active sites in the reduced state and suppresses the methanation pathway, lifting methanol selectivity by roughly 10–15 percentage points relative to Cu alone. The Cu-Zn synergy is the foundation of the 1960s ICI industrial catalyst.

Cerium is the critical trace dopant. Its effect is concentrated at very low loading (≈0.05 wt%) where it provides surface oxygen mobility that keeps the active site clean and steers intermediates toward methanol rather than CO or methane. Above 0.5 wt%, cerium starts to dilute the active copper surface. This sharp, low-loading optimum is exactly the type of feature that conventional screening misses and that Bayesian optimization is designed to find.

Indium represents a second, independent pathway to high methanol selectivity — the In2O3/ZrO2 family identified in the 2010s. Around 3 wt% In, this formulation achieves selectivity competitive with Cu-Zn-Ce. However, indium is roughly 25× more expensive per gram than copper, so the cost objective keeps the optimizer from settling here once it has found the Cu-based solution. Combining copper and indium is worse than either alone: the two active-site mechanisms compete rather than cooperate.

Rhenium (the 1980s Re/ZrO2 system) is modestly active per gram but four hundred times more expensive than copper per gram, making it economically untenable unless performance is dramatically superior — which it is not at the loadings achievable.

Manganese and gallium are mild promoters that offer incremental gains only in the presence of copper. Neither is worth loading at the cost they add, and the optimizer learns to leave them near zero.