Controlling the evolution of complex physical systems is a fundamental task across science and engineering. Classical techniques suffer from limited applicability or huge computational costs. On the other hand, recent deep learning and reinforcement learning-based approaches often struggle to optimize long-term control sequences under the constraints of system dynamics. In this work, we introduce Diffusion Physical systems Control (DiffPhyCon), a new class of method to address the physical systems control problem. DiffPhyCon excels by simultaneously minimizing both the learned generative energy function and the predefined control objectives across the entire trajectory and control sequence. Thus, it can explore globally and identify near-optimal control sequences. Moreover, we enhance DiffPhyCon with prior reweighting, enabling the discovery of control sequences that significantly deviate from the training distribution. We test our method in 1D Burgers' equation and 2D jellyfish movement control in a fluid environment. Our method outperforms widely applied classical approaches and state-of-the-art deep learning and reinforcement learning methods. Notably, DiffPhyCon unveils an intriguing fast-close-slow-open pattern observed in the jellyfish, aligning with established findings in the field of fluid dynamics.

Graph neural networks (GNNs), a type of neural network that can learn from graph-structured data and learn the representation of nodes through aggregating neighborhood information, have shown superior performance in various downstream tasks. However, it is known that the performance of GNNs degrades gradually as the number of layers increases. In this paper, we evaluate the expressive power of GNNs from the perspective of subgraph aggregation. We reveal the potential cause of performance degradation for traditional deep GNNs, i.e., aggregated subgraph overlap, and we theoretically illustrate the fact that previous residual-based GNNs exploit the aggregation results of 1 to $k$ hop subgraphs to improve the effectiveness. Further, we find that the utilization of different subgraphs by previous models is often inflexible. Based on this, we propose a sampling-based node-level residual module (SNR) that can achieve a more flexible utilization of different hops of subgraph aggregation by introducing node-level parameters sampled from a learnable distribution. Extensive experiments show that the performance of GNNs with our proposed SNR module outperform a comprehensive set of baselines.

As one of the most popular GNN architectures, the graph attention networks (GAT) is considered the most advanced learning architecture for graph representation and has been widely used in various graph mining tasks with impressive results. However, since GAT was proposed, none of the existing studies have provided systematic insight into the relationship between the performance of GAT and the number of layers, which is a critical issue in guiding model performance improvement. In this paper, we perform a systematic experimental evaluation and based on the experimental results, we find two important facts: (1) the main factor limiting the accuracy of the GAT model as the number of layers increases is the oversquashing phenomenon; (2) among the previous improvements applied to the GNN model, only the residual connection can significantly improve the GAT model performance. We combine these two important findings to provide a theoretical explanation that it is the residual connection that mitigates the loss of original feature information due to oversquashing and thus improves the deep GAT model performance. This provides empirical insights and guidelines for researchers to design the GAT variant model with appropriate depth and well performance. To demonstrate the effectiveness of our proposed guidelines, we propose a GAT variant model-ADGAT that adaptively selects the number of layers based on the sparsity of the graph, and experimentally demonstrate that the effectiveness of our model is significantly improved over the original GAT.