Human-in-the-Loop LLM-Assisted Catalyst Design for CO₂-to-DME : From Hypothesis Generation to Experimental Validation
Year: 2026
Type: other
Abstract: Large language models (LLMs) are increasingly used in catalysis research, yet their integration has so far clustered around autonomous robotic laboratories, while human-mediated closed-loop workflows accessible to most laboratories remain underreported. To address this gap, this work uses the direct conversion of CO₂ to dimethyl ether (DME) as a case study to document, implement, and critically evaluate a human-mediated, LLM-assisted catalyst design cycle. The LLM's (Claude Sonnet) domain knowledge and mechanistic reasoning were assessed before any design task was delegated. The workflow then followed the conventional catalyst research cycle under continuous human oversight. The model was briefed on the project context, then directed to retrieve and analyse the literature, formulate hypotheses, and propose candidates. An expert panel selected and refined a single formulation (In₂O₃/ZrO₂–H-Ferrierite) for testing, and the results were fed back to the model to inform reformulation. Every synthesis change concerned the In₂O₃/ZrO₂ hydrogenation component, with the H-Ferrierite dehydration component held constant throughout. The co-precipitation route proposed in the model produced tetragonal rather than monoclinic ZrO₂, resulting in poor activity, low DME selectivity, and high CO selectivity (56–69%). Mechanistic analysis traced the fault to the synthesis method and led to a reformulated route using impregnation onto preformed monoclinic zirconia, thereby raising the DME space-time yield by roughly 11-fold, from 0.017 to 0.19 mmol g⁻¹ h⁻¹. However, CO selectivity remained high (50–56%), and deeper analysis traced this to insufficient In–O–Zr interfacial contact, intrinsic to the impregnation method. Three independent model sessions then converged on a sol-gel route predicted to restore this contact through more intimate indium–zirconia mixing. The work's primary contribution is the human-mediated feedback loop itself, a cycle not previously documented in the catalysis literature. It also documents how well-known LLM failure modes manifest in catalyst design, showing researchers what to expect and how each was managed. This was demonstrated on a single system; thus, the work's value lies in its transparency rather than in its generality.
Source: Doria (University of Helsinki)
Author: Korir, Wayne Kipngeno
Institution: University of Akron
Language: English
Subfield: Catalysis
Field: Chemical Engineering
Domain: Physical Sciences
Open Access status: green
Funder: European Commission
OpenAlex ID: W7203847897