Locally-Deployed Chain-of-Thought (CoT) Reasoning Model in Chemical Engineering: Starting from 30 Experimental Data

TL;DR

This work tackles the challenge of predicting molecular solubility and related properties with very limited data in chemical engineering by introducing locally deployed Chain-of-Thought (CoT) reasoning. It combines traditional surrogate models (Gaussian processes) with large-language models (LLMs) in two configurations: LLM-CoT and ML-LLM-CoT (the latter embeds a Gaussian ML model within the CoT loop). Through careful deployment of local models (DeepSeek-R1:14b, Qwen-2:7b) and error-driven prompting, the study shows that ML-LLM-CoT offers superior efficiency and lower high-deviation occurrences compared to LLM-CoT, particularly for dissimilar molecular structures, while maintaining strong performance for similar structures. The results suggest that CoT-based hybrid reasoning can improve rapid property predictions and process optimization in chemical engineering, with potential for privacy-preserving, local deployment and extension to larger models and broader molecular-property tasks.

Abstract

In the field of chemical engineering, traditional data-processing and prediction methods face significant challenges. Machine-learning and large-language models (LLMs) also have their respective limitations. This paper explores the application of the Chain-of-Thought (CoT) reasoning model in chemical engineering, starting from 30 experimental data points. By integrating traditional surrogate models like Gaussian processes and random forests with powerful LLMs such as DeepSeek-R1, a hierarchical architecture is proposed. Two CoT-building methods, Large Language Model-Chain of Thought (LLM-CoT) and Machine Learning-Large Language Model-Chain of Thought (ML-LLM-CoT), are studied. The LLM-CoT combines local models DeepSeek-r1:14b and Qwen2:7b with Ollama. The ML-LLM-CoT integrates a pre-trained Gaussian ML model with the LLM-based CoT framework. Our results show that during construction, ML-LLM-CoT is more efficient. It only has 2 points that require rethink and a total of 4 rethink times, while LLM-CoT has 5 points that need to be re-thought and 34 total rethink times. In predicting the solubility of 20 molecules with dissimilar structures, the number of molecules with a prediction deviation higher than 100\% for the Gaussian model, LLM-CoT, and ML-LLM-CoT is 7, 6, and 4 respectively. These results indicate that ML-LLM-CoT performs better in controlling the number of high-deviation molecules, optimizing the average deviation, and achieving a higher success rate in solubility judgment, providing a more reliable method for chemical engineering and molecular property prediction. This study breaks through the limitations of traditional methods and offers new solutions for rapid property prediction and process optimization in chemical engineering.

Figures (10)

• Figure 1: A flowchart comparing two approaches for predicting data when 30 experimental results are available: traditional machine-learning with hyperparameter optimization and a method involving the DeepSeek agent, highlighting challenges and considerations such as prediction range, data size, and requirements for fine-tuning.
• Figure 2: Flowchart depicting the process of building a Chain-of-Thought (CoT) using a Deepseek agent. It involves leveraging 30 known data points, promoting and condensing data, predicting with recording of the reasoning process, conducting error analysis and summary, and testing on 20 unknown data points categorized as most similar and most dissimilar, with the aim of creating a useful corpus.
• Figure 3: Flowchart of building CoT with LLM themself
• Figure 4: This figure presents the results of the molecular property prediction process during the construction of the Chain-of-Thought (CoT). Sub-figure (a) showcases the single-attempt and multiple-attempt prediction performance of individual molecules, and sub-figure (b) offers in-depth details about the molecules that required multiple predictions, highlighting the importance of error-based iterative prediction and analysis in improving prediction accuracy.
• Figure 5: (a) Deviation of the 20 most dissimilar molecules with the original 30 molecules in the 1128-data-point open-source dataset. Blue bars represent normal deviation values of the molecules. (b) Deviation of molecules with extreme error among the 20 most dissimilar molecules. Red bars indicate extreme deviation values. (c) Consistency of solubility determination for the 20 most dissimilar molecules, where the blue part represents the proportion of molecules for which prediction and reality are consistent, and the red part represents the proportion of inconsistent ones.