The field of scientific machine learning is moving towards the development of more flexible and powerful models that can handle heterogeneous and multimodal data. Researchers are exploring new architectures and techniques that can capture complex patterns and relationships in scientific data, such as partial differential equations and chaotic systems. A key direction is the development of foundation models that can be pre-trained on diverse datasets and fine-tuned for specific tasks, enabling zero-shot or few-shot generalization across different domains. Noteworthy papers include: MORPH, which introduces a shape-agnostic foundation model for partial differential equations, and ChaosNexus, which proposes a foundation model for universal chaotic system forecasting with multi-scale representations. These models have shown state-of-the-art performance in various benchmarks and have the potential to advance the field of scientific machine learning.