Day 1: Physics-Informed & Geometric Deep Learning
Integrating scientific domain knowledge into neural architectures to ensure physical consistency and data efficiency.
- The Inductive Bias Shift: Moving from black-box models to architectures that respect conservation laws such as mass, energy, and momentum.
- Physics-Informed Neural Networks: Implementing Partial Differential Equations directly into the loss function.
- Geometric Deep Learning: Exploring equivariant and invariant networks for 3D spatial data, molecular structures, and fluid manifolds.
Advanced Hands-on: Google Colab
- Project: Developing a PINN to simulate fluid flow or heat transfer.
- Key Skill: Utilizing automatic differentiation to enforce physical constraints without needing massive labeled datasets.
Day 2: Generative Models & Autonomous R&D Pipelines
Accelerating the research cycle through generative design and closed-loop experimental optimization.
- Generative AI in Science: Applications of Diffusion Models and Variational Autoencoders in crystal structure prediction and protein design.
- Bayesian Optimization & Active Learning: Building self-driving labs that intelligently select the next experiment to perform.
- Scientific LLM Agents: Leveraging multi-modal Large Language Models to automate literature synthesis and hypothesis generation.
Advanced Hands-on: Google Colab
- Project: Targeted Molecular Design.
- Key Skill: Training a latent diffusion model to generate novel chemical structures with optimized properties such as specific conductivity or binding affinity using RDKit.
Day 3: Explainability, Uncertainty & Symbolic Discovery
Ensuring reliability and extracting human-readable insights from high-dimensional AI models.
- Uncertainty Quantification: Utilizing Conformal Prediction and Bayesian inference to measure model confidence for high-stakes industrial deployment.
- Symbolic Regression: Transforming neural outputs into human-readable mathematical formulas.
- Model Interpretability: Exploring advanced XAI techniques tailored for scientific data, including attribution maps and structural causal models.
Advanced Hands-on: Google Colab
- Project: AI-to-Formula Discovery.
- Key Skill: Implementing PySR to rediscover physical laws from noisy experimental sensor data, bridging the gap between deep learning and classical theory.
Technical Implementation & Environment
- Platform: All modules are optimized for Google Colab and Jupyter Notebooks, utilizing GPU-accelerated libraries.
- Toolkit: Participants will work with industry-standard frameworks including PyTorch, DeepXDE, RDKit, and PySR.
- Outcome: Participants will leave with a portfolio of advanced notebooks that can be adapted for real-world R&D challenges in chemistry, physics, engineering, and pharmacology.
