[Paper Review] ChemReasoner: Heuristic Search over a Large Language Model's Knowledge Space using Quantum-Chemical Feedback

Henry W. Sprueill, Carl N. Edwards+7 more|arXiv (Cornell University)|Feb 15, 2024

Advanced Text Analysis Techniques4 citations

TL;DR

ChemReasoner introduces a hybrid AI framework that combines large language model (LLM)-driven heuristic search with quantum-chemistry feedback via graph neural networks (GNNs) to accelerate catalyst discovery. By iteratively generating hypotheses, evaluating them using 3D atomistic simulations, and refining search via adsorption energy and reaction barrier rewards, the method achieves superior performance—outperforming expert-crafted descriptor approaches in 2 out of 3 benchmarks—while enabling fully automated, trustworthy, and scientifically grounded catalyst screening.

ABSTRACT

The discovery of new catalysts is essential for the design of new and more efficient chemical processes in order to transition to a sustainable future. We introduce an AI-guided computational screening framework unifying linguistic reasoning with quantum-chemistry based feedback from 3D atomistic representations. Our approach formulates catalyst discovery as an uncertain environment where an agent actively searches for highly effective catalysts via the iterative combination of large language model (LLM)-derived hypotheses and atomistic graph neural network (GNN)-derived feedback. Identified catalysts in intermediate search steps undergo structural evaluation based on spatial orientation, reaction pathways, and stability. Scoring functions based on adsorption energies and reaction energy barriers steer the exploration in the LLM's knowledge space toward energetically favorable, high-efficiency catalysts. We introduce planning methods that automatically guide the exploration without human input, providing competitive performance against expert-enumerated chemical descriptor-based implementations. By integrating language-guided reasoning with computational chemistry feedback, our work pioneers AI-accelerated, trustworthy catalyst discovery.

Motivation & Objective

• To address the challenge of discovering optimal catalysts for complex reactions like CO2-to-methanol conversion, where traditional descriptor-based methods are limited by empirical understanding and combinatorial complexity.
• To overcome the limitations of LLMs in reasoning about 3D atomic-scale interactions and multi-step reaction pathways by grounding their hypotheses in quantum-chemical simulations.
• To develop an autonomous, planner-driven framework that guides LLM search through catalyst space without human input, using computational feedback as a reward signal.
• To demonstrate that integrating language-based reasoning with physics-based scoring (e.g., adsorption energy, reaction barriers) leads to more reliable and efficient catalyst discovery than expert-crafted descriptors alone.

Proposed method

• The framework formulates catalyst discovery as a reinforcement learning-like search in an uncertain environment, where an LLM acts as an agent exploring a knowledge space of chemical hypotheses.
• At each step, the LLM generates search prompts based on dynamically selected chemical descriptors (e.g., 'resistance to poisoning', 'ability to dissociate CO2') using automated planning.
• Candidate catalysts are converted into 3D atomistic representations and evaluated via a GNN model trained on quantum-chemistry data to compute adsorption energies and reaction barriers.
• A reward function based on these physical properties—particularly adsorption energy and activation barriers—guides the LLM toward energetically favorable, stable, and selective catalysts.
• The planner uses feedback from prior steps to prune unpromising actions and refine search criteria, including dynamic inclusion/exclusion rules (e.g., 'must adsorb and activate CO2') and relationship logic ('different from' existing candidates).
• The system supports iterative, tree-structured reasoning, enabling multi-step hypothesis generation and refinement grounded in scientific principles and computational feedback.

Experimental results

Research questions

• RQ1 Can an LLM-guided heuristic search, when grounded in quantum-chemical feedback, outperform expert-designed descriptor-based screening in catalyst discovery?
• RQ2 To what extent can automated planning in an LLM agent reduce reliance on human-crafted chemical descriptors while maintaining or improving search efficiency and accuracy?
• RQ3 How effective is the integration of 3D atomistic simulation feedback (via GNNs) in steering LLM reasoning toward physically plausible and energetically favorable catalysts?
• RQ4 Can the framework discover non-noble metal catalysts with high activity for CO2-to-methanol conversion, expanding beyond traditional noble-metal assumptions?
• RQ5 Does the use of multi-property feedback (e.g., adsorption energy, reaction barriers, stability) lead to more robust and selective catalyst predictions than single-property screening?

Key findings

• The LLM-planned approach (ChemReasoner-Planner) outperformed expert-crafted descriptor-based search (ChemReasoner-Expert) in 2 out of 3 evaluation categories, demonstrating the superiority of automated, feedback-driven planning.
• The framework successfully identified high-performing catalysts, including Pd-Au, Pt-Ru, Ru-Au, Rh-Pd, and Pt-Au alloys, with scientifically grounded justifications based on CO2 activation and CO reduction mechanisms.
• By incorporating reaction pathway and stability analysis, the method avoided candidates with high energy barriers or poor structural stability, enhancing prediction reliability.
• The integration of quantum-chemical feedback enabled the discovery of non-noble metal catalysts not previously emphasized in expert lists, such as Pd-Au and Ru-Au alloys, expanding the search space beyond noble metals.
• The use of dynamic inclusion/exclusion criteria—such as 'must adsorb and activate CO2'—allowed the system to refine candidate sets and avoid unproductive search directions autonomously.
• The framework achieved competitive performance with state-of-the-art LLM-based implementations while significantly reducing human intervention, marking a step toward trustworthy, AI-accelerated scientific discovery.