Large language models (LLMs) and related foundation-model workflows are emerging as promising tools for advancing foundry intelligence across the casting value chain. This review examines their applications in material design and property prediction, process parameter optimization and intelligent control, and defect detection and quality tracing in casting environments. The surveyed studies indicate that LLM-enabled systems can help integrate unstructured technical knowledge with multimodal industrial data. This integration supports composition design, simulation-assisted process optimization, diagnostic reasoning, and knowledge-grounded decision support. However, current evidence shows that the transition from pilot demonstrations to robust industrial deployment remains constrained by several practical barriers, including heterogeneous data integration, insufficient traceability across process stages, reliability under physical and safety constraints, and the latency and resource limitations of shop-floor environments. We further highlight key research directions for real-world foundry applications, including multimodal cognitive systems, lightweight domain-adapted models, trustworthy retrieval-augmented and physics-aware reasoning, and human-in-the-loop validation frameworks. Overall, the review suggests that the future of foundry intelligence will depend not only on model capability, but also on data governance, deployable system design, and reliable integration with metallurgical knowledge and industrial workflows.

Sodium-ion batteries (SIBs) are regarded as promising next-generation energy storage systems, benefiting from the abundant reserves and low cost of sodium resources. However, safety hazards and performance bottlenecks originating from electrolytes, anodes, and cathodes have not only led to frequent accidents but also severely hindered their industrialization process. Current review works, which mainly focus on isolated issues or single components, lack the holistic guidance required for the development of high-safety and high-performance SIBs. To fill this research gap, this review systematically summarizes the latest advances in the application of machine learning (ML) in SIBs development from the perspective of core components. First, we clarify the root causes of key challenges for each component, and then evaluate the role of ML in accelerating material discovery, optimizing electrochemical performance, and mitigating safety risks. Specifically, ML methodologies such as graph neural networks, multi-objective optimization, and physics-informed models are highlighted for their unique advantages in deciphering the structure-performance relationships of SIBs materials. This work demonstrates that ML can efficiently explore the high-dimensional design spaces of electrodes and electrolytes, thereby establishing a data-driven paradigm for SIBs optimization. Finally, we propose that future research should prioritize the construction of standardized data ecosystems, the development of integrated computational-experimental pipelines, and the establishment of cross-component safety design frameworks, aiming to bridge the gap between computational predictions and practical industrial applications.