On-Demand LLM-Augmented Bayesian Optimization for Amine Sorbent Discovery in Direct Air Capture

Solid amine sorbents are among the most promising materials for Direct Air Capture (DAC), but their design is bottlenecked by costly synthesis-and-testing cycles. Bayesian Optimization (BO) reduces the number of experiments, yet its Gaussian Process surrogate becomes informative only after a sufficient number of data points; in the DAC setting, where historical data is scarce, the acquisition func- tion selects near-random points. Large Language Models (LLMs) can propose formulations from literature knowledge, but may hallucinate chemically invalid suggestions. We present On-Demand LLM-Augmented Bayesian Optimization (ODBO), which combines BO with an LLM and real-time web search. A GP-based scheduler consults the LLM when the posterior is uncertain and otherwise delegates to BO for local refinement. A retrieval-based prompt compression reduces token consumption by 17% without measurable loss in suggestion quality. On a literature-based DAC benchmark with 18 materials and 30 iterations, ODBO with web search achieves 60% success in rediscovering the held-out optimal sor- bent, the same as without web search, but reaches higher average capacity (3.61 vs. 2.76 mmol/g); this difference is suggestive rather than established as significant. On the same DAC search space, Random Search significantly outperforms pure BO (4.40 vs. 4.24mmol/g, Wilcoxon p < 0.0001, 30 seeds), while ODBO outperforms both (4.66 mmol/g, +10.0% over BO, p < 10−9). In 15 independent closed-loop simulation runs, ODBO achieves a mean best capacity of 3.85 ± 0.64 mmol/g. Together, these results indicate that the primary driver of ODBO’s improvement is the LLM candidate proposal mechanism, not the GP surrogate or web search. We position this work as an empirical study of when LLM priors help in low-data materials discovery, not as a general-purpose optimization framework; all reported gains are conditional on the small discrete search space, the 10-point cold-start budget, and the specific LLM used.