Max Welling: AI, Uncertainty, and Automating Scientific Discovery

AI's Transformative Role in Science

• 💡 AI is driving a paradigm shift in scientific discovery, moving from traditional physical experiments to digital simulation and emulation.
• 🚀 The process involves storing simulation data in databases and transforming it into neural network predictors to accelerate future experiments, creating a "flywheel" effect.
• 🎯 This approach is applicable across diverse fields, from weather forecasting to molecular design.

Advancements in Weather and Molecular Modeling

• ⚡ AI models are now outperforming traditional numerical methods in weather forecasting, leveraging vast meteorological datasets and combining different data resolutions.
• 🔬 For molecular design, AI helps overcome challenges like complex quantum mechanics and vast search spaces, enabling faster understanding and manipulation of molecules.
• 🔑 Equivariant neural networks incorporate physical symmetries (e.g., rotation/translation invariance) to improve model efficiency and understanding of the physical world, especially in graph neural networks.

The Critical Need for Uncertainty Quantification

• ⚠️ Uncertainty quantification (UQ) is vital for AI models, particularly when encountering out-of-distribution data where model predictions might be unreliable.
• 📊 Traditional UQ methods, like ensemble models, are often too computationally expensive and impractical for large-scale applications.
• 🧠 A new Variational Bayesian method using variational dropout offers an efficient way to predict uncertainty in both sequential models (e.g., PDE solvers like BARNN) and non-sequential models (e.g., molecular force fields like BLIP).

Automated Material Design and Societal Impact

• 🌱 The CuspAI platform automates material discovery using generative models, property predictors, and Bayesian optimization, accelerating the search for new materials.
• 🧪 This platform is being applied to develop carbon capture technologies (e.g., porous metal-organic frameworks) and other materials for critical societal challenges.
• ✅ AI-driven material design holds immense potential to address global issues such as climate change, clean energy, environmental protection, technological progress, and health (drug discovery).