When it comes to applying machine learning to physical system modeling, it is more and more common to see practitioners moving away from a pure data-driven strategy, and starting to embrace a hybrid mindset, where rich prior physical knowledge (e.g., governing differential equations) is used together with the data to augment the model training.
Under this background, physics-informed neural networks (PINNs) have emerged as a versatile concept and led to many success stories in effectively solving real-world challenges.
As a practitioner who is eager to adopt PINNs, I am keen on learning both the latest developments in training algorithms, as well as the novel use cases of PINNs for real-world applications. However, a pain point I often see is that, although there are abundant research papers/blogs summarizing effective PINN algorithms, overviews of novel use cases of PINNs can rarely be found. One obvious reason is that, unlike the training algorithms which are domain-agnostic, reports of PINN use cases are scattered in various engineering domains and not readily accessible for a practitioner who is usually an expert in one specific domain. As a consequence, I often found myself reinventing the wheel as my ways of using PINNs have already been well addressed by practitioners in another field.
It is exactly my journey and experiences that have sparked the idea of writing this blog: here, I strive to break the information barrier across different engineering domains and distill the recurring functional usage patterns of PINNs. I hope that this review will inform practitioners from different domains about what’s possible with PINNs and inspire new ideas for interdisciplinary innovation.
Toward that end, I have extensively reviewed PINN research papers in the past three years and came up with the following 5 main usage categories:
- Predictive modeling and simulations
- Data-driven insights
- Data-driven enhancement
- Monitoring, diagnostic, and health assessment