GPT-4-powered analysis and prediction of experiments through an effective chain-of-thought prompting strategy: a case study of selective catalytic reduction of NOx with NH3 by metal-oxide composites

The chemistry laboratories of the future are poised to undergo a significant transformation, largely driven by the advent and integration of large language models (LLMs) into everyday tasks and workflows. As we stand at the threshold of this exciting fusion of chemistry research with artificial intelligence (AI), a pressing question arises: "In what ways can LLMs enhance and contribute to the field of chemistry?" This study investigates a potential application of LLMs in interpreting and predicting experimental outcomes based on various experimental variables, by leveraging the human-like reasoning and inference capabilities of LLMs to process and analyse the presented information. This study centres on the selective catalytic reduction of NOx with NH3 using metal-oxide composites, presenting a chemistry challenge for evaluating GPT-4's ability to discern correlations. We specifically assess how GPT-4 relates experimental variables—including the catalyst's composition, synthesis conditions, and operational parameters—to the corresponding outcome, namely the catalytic performance. We implement the Chain of Thought (CoT) concept to create reasoning paths that establish relationships, which are then reinput into GPT-4 to generate inferences. This involves formulating a series of logical steps, akin to human problem-solving, to uncover connections within the data. The results of these reasoning paths serve as a feedback loop, thereby enriching the depth and accuracy of GPT-4's subsequent inferences. Here, we introduce a specialized CoT prompting strategy, termed "Ordered-and-Structured" CoT (OSCoT), wherein each individual feature or aspect of the problem is systematically and sequentially examined. This contrasts with the baseline "one-pot" CoT (OPCoT) approach, where all information is processed in a simultaneous manner. The OSCoT strategy's sequential analysis of each problem aspect significantly enhances the quality and precision of the inferences generated by GPT-4. Our findings show that GPT-4, powered by OSCoT, can accurately predict catalytic performances for both binary and ternary metal-oxide composites.