Huge search spaces, hidden patterns
A lot of scientific problems share the same basic shape: there's an enormous number of possible answers, far too many to check one by one, but the space isn't random. It has structure. Which arrangement of atoms makes a stable, useful material, which molecule will bind tightly to a disease-related protein, which three-dimensional shape will a chain of amino acids fold into. These questions all involve searching through a combinatorial space so large that brute-force enumeration is hopeless, yet the underlying physics and chemistry mean that good answers aren't scattered randomly. They cluster in ways a model can learn to recognize.
This is exactly the kind of problem machine learning is good at. If you covered machine learning fundamentals or neural networks before, you already know the core idea: a model trained on enough examples learns to recognize patterns and generalize to new cases it hasn't seen. Applied to science, that means a trained model can look at a new candidate molecule or material and make an informed guess about its properties, without exhaustively simulating or synthesizing it first.