Leveraging large language models for nano synthesis mechanism explanation: solid foundations or mere conjectures?

TL;DR

This work addresses whether large language models truly understand the physicochemical mechanisms underlying gold nanoparticle synthesis. It introduces a mechanism-focused benchmark of 775 expert-level multiple-choice items and a logits-based confidence metric, the $c$-score, to quantify true mechanistic comprehension beyond mere recall. Across open-source and commercial LLMs, results show that top models (e.g., GPT-4, Claude) achieve high accuracy, while $c$-scores reveal nuanced confidence in correct mechanistic reasoning and highlight differences not captured by accuracy alone. The study provides a rigorous framework for evaluating scientific reasoning in materials science, supporting the development of more reliable AI tools for mechanistic discovery, and includes data and code for reproducibility. $c$-score = $\frac{1}{N}\sum_{i=1}^{N}\frac{e^{L_G^i}}{e^{L_A^i}+e^{L_B^i}+e^{L_C^i}+e^{L_D^i}}$ quantifies the model’s probabilistic commitment to the correct answer across questions, enabling interpretable assessment of mechanistic understanding.

Abstract

With the rapid development of artificial intelligence (AI), large language models (LLMs) such as GPT-4 have garnered significant attention in the scientific community, demonstrating great potential in advancing scientific discovery. This progress raises a critical question: are these LLMs well-aligned with real-world physicochemical principles? Current evaluation strategies largely emphasize fact-based knowledge, such as material property prediction or name recognition, but they often lack an understanding of fundamental physicochemical mechanisms that require logical reasoning. To bridge this gap, our study developed a benchmark consisting of 775 multiple-choice questions focusing on the mechanisms of gold nanoparticle synthesis. By reflecting on existing evaluation metrics, we question whether a direct true-or-false assessment merely suggests conjecture. Hence, we propose a novel evaluation metric, the confidence-based score (c-score), which probes the output logits to derive the precise probability for the correct answer. Based on extensive experiments, our results show that in the context of gold nanoparticle synthesis, LLMs understand the underlying physicochemical mechanisms rather than relying on conjecture. This study underscores the potential of LLMs to grasp intrinsic scientific mechanisms and sets the stage for developing more reliable and effective AI tools across various scientific domains.

Figures (6)

• Figure 1: Semantic illustration of our proposed framework for large language model evaluation in nanomaterial synthesis prediction, highlighting concepts and workflow. a) Nanosynthesis study loop: begins with basic conditions, leading to the discovery of novel synthesis rules through experiments involving variable adjustments. b) exemplifies the synthesis mechanism, dissected into causality and correlations, with an emphasis on correlations described through condition-observation pairs. c) outlines the process from sourcing relevant literature (using key area keywords) for benchmark construction and model evaluation.
• Figure 2: Evaluation data set illustration. a) shows the distribution of collected evaluation sets containing 775 questions categorized by synthesis methods and structures, respectively. b) displays a jittered scatter plot of manually curated research papers with the counts of mechanism, conditions and observations, with mechanism relevance from low to high, indicated by varying colors to represent the frequency of observations and varying sizes to represent the biasing towards mechanism. c) showcases the multiple selection question considered in the evaluation. The model is instructed to give the correct option. d) illustration of the probing test in our evaluation study based on the proposed c-score.
• Figure 3: Evaluation accuracy of baselines and the confidence-based scores of top-5 open-sourced models compared to the original accuracy in multiple selection questions. a) x-axis represents different models, while y-axis is the accuracy. The figure delineates the range in accuracy achieved by each model under different temperature settings (from 0.0 to 1.0), where the circles represent the accuracy at each temperature setting, and the diamonds denote their average values. b) The comparison between accuracy and condifence-based scores among 5 top-performance models, showing the performence increasing (green line) and decreasing (red line).
• Figure 4: Evaluation results of temperature effects on baselines. Each subplot illustrates the accuracy (y-axis) of the corresponding model under various temperature settings (x-axis), organized in descending order based on the average of the model performance across multiple-choice questions at different temperatures. Each point denotes the accuracy at a fixed temperature.
• Figure 5: Illustration of the knowledge probing method. Given the input with both question-options and instructions, the model should give answer with predefined vocabulary, which includes A, B, C and D. The probability of each option is drawn based one the ouput logits with fixed setting of the temperature before Softmax. By observing the distribution changes of all options, model behaviors can be revealed upon different test cases. We use this design to test models ability regarding the knowledge of AuNPs synthesis.