We investigate the transferability of Earth weather foundation models to planetary atmospheres by adapting the GraphCast graph neural weather forecasting model to Mars. While GraphCast achieves state-of-the-art performance for terrestrial forecasting, its applicability to non-Earth environments remains unexplored. Using the Mars Climate Database (MCD), which provides global atmospheric fields across vertical altitude levels (similar to Earth pressure levels), we evaluate zero-shot and fine-tuned GraphCast predictions of Martian temperature and wind fields. Zero-shot forecasts produce a surprisingly accurate depiction of current conditions but fail to reproduce diurnal variability and rapidly decay toward climatological mean states. To address this limitation, we fine-tune GraphCast using MCD variables and top-of-atmosphere solar radiation forcing while holding humidity constant. Fine-tuning enables rapid learning of Martian thermal variability. Within as few as 10 training epochs, the model begins to capture the diurnal cycle and forecasts up to 10 days reproduce seasonal and vertical temperature structure. Prediction quality improves with training sample size and exhibits sensitivity to seasonal initialization. These results demonstrate that Earth-trained AI weather models can be adapted to simulate Martian atmospheric dynamics, providing a pathway toward rapid planetary weather prediction to support mission operations, dust storm risk mitigation, and future human exploration.
M. Carroll, J. Li, S. Guzewich et al.· 0 citations
To address the low accuracy of multi-component-multi-target interaction prediction in Bupleurum chinense caused by complex molecular structures and limited target associations, this paper employs advanced graph models combined with biological network topology information. The same graph-based representation strategy is relevant to materials informatics, including the screening of molecular or polymer structures for electromagnetic sensing, dielectric response, and functional coating design. Specifically, RDKit is used to construct chemically defined molecular graphs, with atoms as nodes and covalent bonds as edges, and to encode multidimensional atomic features. These graphs are directly input into Graphormer. Graphormer then incorporates a fully connected attention mechanism that integrates topological distances and spatial geometric information from RDKit graphs to extract higher-order molecular representations. Node2Vec is used to embed the protein-protein interaction network in a low-dimensional space and identify potential target pathways. Finally, compound and protein embeddings are concatenated and fed into a two-channel neural network to predict interaction probabilities. With negative sampling for training, the method achieves stable average AUROC values of 0.91-0.93 and AUPRC values of 0.88-0.90. Even when only 10% of targets are visible, the F1-score reaches 0.683±0.021 and Coverage Rate@50 reaches 76.4%, indicating strong inductive reasoning under sparse knowledge conditions.
Reliable transformer fault diagnosis under limited fault samples remains a significant challenge in intelligent power systems. To address the difficulties associated with weak fault signatures, severe environmental interference, and insufficient training samples, this study investigates transformer fault location technology based on acoustic feature recognition and field perception data fusion. The generation mechanism and propagation characteristics of transformer acoustic signals are first analyzed, and an improved time–frequency feature extraction method is developed to enhance feature representation under small-sample conditions. A multi-physics data fusion framework integrating acoustic, vibration, and electrical sensing information is then established, and a dedicated attention mechanism is designed to achieve deep feature fusion across heterogeneous data sources. Finally, an enhanced deep neural network model is employed for accurate fault localization and condition identification. Experimental results demonstrate that the proposed framework effectively improves fault recognition performance and location accuracy under small-sample constraints. The study provides technical support for intelligent power equipment monitoring and offers methodological references for signal propagation analysis, sensor fusion, and electromagnetic condition monitoring systems.
Y. G. Li, L. J. Feng, R. R. Li et al.· Advanced Electromagnetics· 0 citations