Reaction-network reasoning with frontier models for experimentally confirmed catalyst-selectivity hypotheses

Technical Summary: Reaction-Network Reasoning with Frontier Models for Catalyst Discovery

Problem Statement
The discovery of novel catalyst architectures for sustainable chemical manufacturing, particularly in complex systems like electrochemical carbon dioxide (CO₂) reduction, remains a bottleneck dominated by trial-and-error experimentation. In these systems, product selectivity is governed by dynamic interfacial factors and kinetic competition among divergent pathways that often share common intermediates. Conventional computational approaches, including descriptor-based machine learning and first-principles screening, struggle to resolve these critical mechanistic branch points. These methods typically rely on static ground-state descriptors or bulk structural correlations, failing to capture the localized, dynamic environmental variables (e.g., interfacial acidity, electrolyte composition) that dictate whether a reaction proceeds toward a valuable oxygenate (acetate) or a hydrocarbon (ethylene). Furthermore, existing machine learning potentials often suffer from out-of-distribution errors at complex electrochemical interfaces and are trained on ground states rather than the transition-state kinetics required to resolve pathway competition.

Methodology: The CoThinker Framework
To address these limitations, the authors developed CoThinker, a human–AI co-thinking framework that enforces network invariance. Rather than allowing a frontier large language model (LLM) to generate unconstrained text or rely on statistical correlations, the framework constrains the model to reason exclusively over an explicit reaction network.

The workflow operates through four sequential stages:

1. Problem Formulation: A human expert defines the catalytic objective, which the system translates into a mechanistic optimization problem.
2. Reaction-Network Construction: The platform constructs a directed graph of species and elementary steps derived from literature knowledge. For the CO₂ reduction case study, this network was iteratively expanded to include 32 intermediates, 41 elementary steps, and nine competing pathways leading to acetate, ethylene, ethanol, methane, and formate.
3. Iterative Mechanistic Reasoning: The LLM navigates this network to identify "mechanistic hotspots" (branch points) and quantify step leverage. Crucially, the system employs a structured reflect–verify–retry loop. Every proposed hypothesis is subjected to explicit falsification against the complete network to ensure end-to-end pathway coherence. Designs that offer local kinetic improvements but fail global network constraints (e.g., violating scaling relations or failing to resolve upstream bottlenecks) are discarded.
4. Mechanistic Verification: The framework outputs a prioritized set of falsifiable control levers (physical descriptors) rather than a single optimized composition.

Notably, this specific study deliberately bypasses explicit ab initio calculations or neural network potentials (NNPs) during the reasoning phase to prevent the compounding of errors from unstable numerical extrapolations. Instead, it relies on the LLM's encoded literature knowledge to perform logical deduction.

Key Contributions

• Network Invariance: The introduction of a design principle where AI reasoning is anchored to a shared topological map of chemical transitions, ensuring coherence and preventing "contextual drift" common in standard scientific agents.
• Mechanistic Hotspot Identification: The ability to extract specific, actionable physical levers (e.g., local alkalinity, proton-donor accessibility) that govern selectivity, moving beyond correlative predictions to causal, mechanism-guided reasoning.
• Forward Hypothesis Generation: Shifting the computational paradigm from retrospective statistical prediction to the generation of testable, per-step reaction mechanisms.

Results: CO₂ Electroreduction to Acetate
Applied to the electrochemical reduction of CO₂ to acetate on copper-iron (Cu–Fe) oxide catalysts, the CoThinker framework generated specific mechanistic hypotheses which were then prospectively validated:

1. Mechanistic Insights: The framework identified five key mechanistic hotspots. The most critical was hydroxide-assisted ketene desorption and capture, identified as the decisive step locking flux into the acetate pathway. It also identified buffer-assisted protonation of a shared intermediate as the primary competitor diverting flux to ethylene.
2. Design Principles: The reasoning engine distilled the optimal catalyst architecture into specific control levers:
   • Local pH: High local alkalinity is required to promote ketene capture.
   • Proton Accessibility: Interfacial proton donors must be minimized to prevent diversion to ethylene.
   • Fe Incorporation: Iron should act as an electronic modifier upstream of C–C coupling to stabilize intermediates, but contiguous Cu ensembles must be preserved for the coupling step.
3. Experimental Validation: A series of Cu–Fe oxide catalysts (ratios 4:1 to 1:1) were synthesized and tested.
   • Selectivity: The 1:1 Cu–Fe catalyst demonstrated a threefold increase in acetate selectivity compared to Cu-rich baselines.
   • pH Dependence: Increasing electrolyte pH from 6.8 to 7.5 systematically enhanced acetate Faradaic efficiency, confirming the hydroxide-capture hypothesis.
   • Proton Control: Switching to phosphate buffers (high proton availability) suppressed acetate yields, validating the prediction regarding proton-donor accessibility.
   • Structural Confirmation: High-resolution TEM revealed that Fe oxide segregates to peripheral regions while Cu maintains contiguous domains, matching the predicted structural requirement.
   • Computational Support: Universal Machine-learned Potentials (UMAs) confirmed that Fe substitution lowers ketene adsorption energy, facilitating desorption to the solution phase.

Significance
The paper claims that this work provides a broadly applicable blueprint for mechanism-guided materials discovery. By forcing the AI to reason over explicit reaction networks, the framework successfully shifts the objective from predicting a singular "winning" material to deriving falsifiable, step-by-step reaction mechanisms. The successful prospective synthesis of the Cu–Fe catalyst serves as physical proof that this "co-thinking" architecture can navigate complex chemical spaces more rigorously than traditional descriptor-based methods. The authors posit that this approach extends the role of AI from merely accelerating experimentation to defining the precise scientific questions experiments must answer, offering a scalable route for discovering catalysts that rely on minimal, earth-abundant architectures rather than complex surface functionalization. Future iterations of the framework aim to integrate neural network potentials via active learning to manage distribution shifts and model transition states more directly.