Machine learning, particularly methods based on neural networks, has gained significant attention since 2012, resulting in remarkable advancements in fields such as image recognition and natural language processing. However, the development of machine learning faces two significant challenges. The first challenge concerns algorithms, which focus on designing algorithms that can cater to varying data situations. The second challenge pertains to hardware, which ensures sufficient computational resources and efficiency for training complex models within acceptable time and energy constraints. In this thesis, instead of concentrating on traditional machine learning training tasks, I will investigate the potential of implementing machine learning algorithms in two physical systems: a coupled pendulum system (Chapter 3) and an optical system (Chapters 5 and 6). I will focus on the software aspect of constructing a Physical Neural Network (PNN). This involves simulating physical systems and creating Digital Twins to enable training for physical systems that can not be trained directly through backpropagation.