Introduction to ML for Chemistry

Recording

Learning Objectives

• Describe the ML workflow (data → features → model → evaluation) and map it to a chemistry problem
• Identify the main families of ML algorithms and know which problem types each suits
• Understand what this bootcamp covers and how the two weeks are structured
• Set up a Python environment with the key libraries (scikit-learn, RDKit, PyTorch)

Key Takeaways

• Takeaway 1. ML models learn statistical patterns from data — they do not encode physical laws, so the quality of training data is the single biggest lever on model quality.

• Takeaway 2. The ML workflow is a loop: data curation → featurisation → training → evaluation → iteration. Skipping evaluation or treating it as a formality is the most common source of over-optimistic results.

• Takeaway 3. Chemistry imposes domain constraints (molecular validity, physical plausibility) that purely data-driven models can violate — always sanity-check predictions against chemical intuition.

• Takeaway 4. No model family dominates all tasks. Random forests, neural networks, and Gaussian processes each have regimes where they excel; choose based on dataset size, interpretability requirements, and uncertainty needs.

Further Reading & Resources