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

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.

Combined Dynamic Virtual Spatiotemporal Graph Mapping for Traffic Prediction

The continuous expansion of the urban construction scale has recently contributed to the demand for the dynamics of traffic intersections that are managed, making adaptive modellings become a hot topic. Existing deep learning methods are powerful to fit complex heterogeneous graphs. However, they still have drawbacks, which can be roughly classified into two categories, 1) spatiotemporal async-modelling approaches separately consider temporal and spatial dependencies, resulting in weak generalization and large instability while aggregating; 2) spatiotemporal sync-modelling is hard to capture long-term temporal dependencies because of the local receptive field. In order to overcome above challenges, a \textbf{C}ombined \textbf{D}ynamic \textbf{V}irtual spatiotemporal \textbf{G}raph \textbf{M}apping \textbf{(CDVGM)} is proposed in this work. The contributions are the following: 1) a dynamic virtual graph Laplacian ($DVGL$) is designed, which considers both the spatial signal passing and the temporal features simultaneously; 2) the Long-term Temporal Strengthen model ($LT^2S$) for improving the stability of time series forecasting; Extensive experiments demonstrate that CDVGM has excellent performances of fast convergence speed and low resource consumption and achieves the current SOTA effect in terms of both accuracy and generalization. The code is available at \hyperlink{https://github.com/Dandelionym/CDVGM.}{https://github.com/Dandelionym/CDVGM.}