AI for Science: The 5th Paradigm

Core Concepts

Evolution from Empirical, Theoretical, Computational, and Data-intensive paradigms to AI-driven scientific discovery.

Key Challenges

• The Curse of Dimensionality
• Trade-off between Accuracy and Efficiency
• Physics-Informed Neural Networks (PINNs)

Articles

1. Life Science (Bio) — AI methods for protein structure prediction, molecular design and bioinformatics.
2. Weather — AI weather foundation models (Pangu/GraphCast) and Earth-science applications.
3. CFD — PINNs, neural operators (FNO/DeepONet) and AI-accelerated simulation.
4. Materials — Material property prediction, crystal generation and AI-driven materials discovery.
5. Scientific Foundation Models Frontier — From scientific language models to domain foundation models (protein/material/weather), three grand challenges and the "data → model → validate → apply" exploration path.

Reading Path

• Methods: read the four domain articles (Bio/Weather/CFD/Materials) for the AI4S methodology map.
• Direction: read the foundation-model frontier article to see AI4S evolve from "task-specific models" to "general foundation models."
• Deployment: AI4S engineering landing (surrogates, full-loop cases, scientific agents) lives in 13 · Industry Applications.

Key Takeaways

• The essence of AI4S is using AI to approximate/replace expensive scientific computing: data-driven + physics constraints + HPC verification — all three are required.
• Scientific foundation models are AI4S's next main battleground: domain data + compute + validation loop decide the frontier position.
• AI4S's reliability bottom line: physical consistency must be externally verified — model predictions must be cross-checked against simulation/experiment, not trusted from model confidence.