How to go from zero to hero in Machine Learning for materials modelling

I am interested in exploring the use of machine learning tools to accelerate molecular dynamics simulation probably by training machine learning potentials. While I have found some papers that apply these concepts, I am having trouble finding good resources that can guide me in a clear and easy-to-understand way on how to proceed with developing:

1. representations of chemical configurations,
2. subsampling of databases,
3. training a machine learning potential,
4. use model to predict properties of materials

I am wondering if someone can direct me to some good resources that can help me learn these skills on my own.