Compiling Chemical Knowledge into Executable Descriptors for Materials Prediction

Chemically Meaningful Textualization Enables Explainable Validation of Metal-Organic Frameworks by Large Language Models

Computation-ready metal-organic framework (MOF) databases are essential for high-throughput screening, yet many reported crystal structures remain chemically unreasonable or disordered, compromising simulation fidelity. Existing validation approaches can identify non-computation-ready structures, but they often rely on heuristic rules, license requirement, or offer limited interpretability. Here, we show that large language models (LLMs) can serve as interpretable validators of MOF structures when crystallographic information is transformed into chemically meaningful text. By benchmarking nine descriptors, we find that successful LLM-based validation depends not on the amount of structural information alone, but on whether local coordination, framework connectivity, and chemical context are organized into a linguistically learnable representation. Fine-tuned LLMs using specialized descriptors (mof2text) achieve performance comparable to graph-based models in identifying unreasonable MOFs. Importantly, these models extend beyond black-box classification by generating diagnostic rationales for likely error sources, including abnormal bonding, connectivity, and charge states, as well as error-category predictions for annotated datasets. This work establishes chemically informed textualization as the key step that transforms LLMs from generic text models into practical and explainable tools for curating MOF databases.

Language models as QSPR predictors: Unleashing potential through in-context learning and instruction tuning

The Role of Knowledge Representation & Reasoning in Deciphering Chemical Complexity

Modern chemistry is pushing the limits of traditional Artificial Intelligence (AI) models, placing unprecedented demands on data availability to address humanity's most pressing challenges. One particular concern is AI's dependence on large, curated data and its tendency to deviate from or misrepresent fundamental chemistry principles. Nonetheless, this concern is often overshadowed by the urgent demand for emergent solutions to real‐world problems. This perspective describes the incorporation of a domain‐specific knowledge representation & reasoning (KR&R) framework with machine learning (ML) for predictive chemistry. KR&R is presented as a framework to represent chemical knowledge, making a formal connection between inductive hypothesis generation and deductive reasoning. By integrating scientific rules into data‐driven processes, upholding a “chemist in the loop” approach, KR&R ensures that ML models are understandable and consistent with existing chemical theory. These concepts are illustrated by case studies where KR&R improves the interpretability of ML predictive models targeting thermodynamic properties (Δ G sol , Δ vap H m °), reaction yields, and catalytic performance. These examples also show KR&R's importance in managing the complexity of modern computational chemistry, establishing it as a key component of explainable AI in the field.

onepot-Bench 0: towards lab-aware in silico chemistry benchmarks

Onepot-Bench 0 is introduced, a proprietary benchmark suite for evaluating language models on synthetic chemistry capabilities relevant to wet-lab execution and probes basic competency, reliability, and deeper knowledge, all skills which are required for reliable performance in the lab.

ChemDIRT: A Diversified Instruction, Representation, and Task Benchmark for Robust Chemistry-LLM Evaluation

This work introduces ChemDIRT (Diversified Instruction, Representation, and Task Benchmark), a comprehensive evaluation framework designed to assess the robustness of chemical reasoning in LLMs and benchmark a diverse set of open- and closed-source LLMs.

Multi-Granular Rationale-Guided Molecular LLM for Property Prediction

This is the first method to expose GNN-derived attributions to an LLM as evidence for property prediction, and achieves the best overall results among generalist models and narrows the gap to specialist models tuned for each task.