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

TL;DR

ChemReasoner tackles catalyst discovery by fusing language-driven hypothesis generation with quantum-chemical feedback from 3D atomistic representations to search an uncertain catalyst space. The framework uses an LLM-based planner to generate and navigate prompts and combines structure-relaxation and reaction-energy estimates from a graph neural network surrogate as rewards, formalized by $r(c) = -\min_{p\in Paths} \left( \max_{ads_t\in p} (E_{ads_t} - E_{ads_{t-1}}) \right)$. It demonstrates that planning-guided exploration can outperform descriptor-only baselines and that LLM-driven hypotheses can be effectively grounded in physics-based feedback. A CO2-to-methanol case study shows alignment with commercial catalysts in some variants and highlights valuable alloying strategies, with datasets and code made publicly available.

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.

Figures (9)

• Figure 1: ChemReasoner successively "thinks" in terms of different constraints and factors, which are based on scientific principles and narrow down the set of possible candidates. It accomplishes that by prompting a language model with different combinations of chemical descriptors, yielding a tree-structured space of queries and potential candidates, and returns the optimal answer via efficient exploration of the search-space. ChemReasoner uses automated planning, based on previous reasoning, to initiate the exploration and guides it via a reward obtained from the exploration process to prune unpromising actions. We currently use “adsorption energy”, a key measure of reactivity as the reward function.
• Figure 2: Illustration of ChemReasoner search process (best viewed in color): The initial question generates base candidates, which are iteratively refined by adding an optimal set of constraints to the query and producing a new set of actions (or prompts) to explore the LLMs internal knowledge space. The optimal action set is chosen by 1) sampling from expert specified action space, or 2) automated generated by a planner component as illustrated in Figure \ref{['fig:planner']}. We describe the resultant structure shown as the "search tree" and each node in the tree represents a set of 3-tuple of (question, answer, reward). We refer to the initial query as a "root node."
• Figure 3: Planner-guided search action generation (best viewed in color): Given a query state defined by a question (shown in top-left) and the set of corresponding answers, the LLM is used as an optimizer to generate a "plan" for the next query. The LLM performs internal reasoning as shown in orange boxes. It accounts for the complete context from root query up to the current query node and generates a "query plan" with the attributes "catalyst type", "inclusion criteria", "exclusion criteria" and "relationship to current candidate list".
• Figure 4: Illustration of planner guided heuristic search (best viewed in color) described in section 3.2 below. Note the systematic expansion of the query plan in the orange boxes (left column). The middle column shows illustration of 3D atomistic structures generated from chemical symbols. Each 3D structure is processed by a reward function that involves geometry relaxation and potentially deriving approximations of energy barriers in reaction pathways (right column). Visualizations of materials acquired from the Materials project structure finder matproject.
• Figure 5: An illustration of reaction pathways corresponding to the conversion of CO2 to methanol for two different catalysts (red and blue bars shows the energy of different intermediates). The energy barrier (difference between the lower energy state and the higher energy state, described as hill climbing in the text) associated with these catalysts is shown by the red and blue arrows. For ease of comparison between the catalysts, the energies have been shifted such that both CO2 states have an adsorption energy of 0 eV.