Using Physics-Informed Machine Learning to Improve Predictive Model Accuracy
“[Deep Learning Toolbox provides a] nice cohesive framework where you can do signal analysis, image processing, deep learning, and high-performance computing all in one script.”
Key Outcomes
- Use a cohesive framework comprising signal analysis, image processing, HPC, and deep learning
- Improve accuracy and efficiency by implementing physics principles and synthetic data
- Eliminate the need to know the governing equations to conduct the analysis
Machine learning models process large data sets but neglect physics when minimizing the error between data and prediction. It is computationally intensive to consider the laws of physics, but neglecting to consider them limits the efficacy of extrapolation to desired conditions.
Another challenge is that physics-based partial differential equations (PDEs) are very generalized, but are not specific enough for application without boundary conditions. On the other hand, empirical equations are good at explaining the data from which they were built but are useless elsewhere.
Dr. Sam Raymond of Stanford University’s Department of Bioengineering combines scientific computing and deep learning to improve the accuracy and generalizability of predictive climate models.
Dr. Raymond and his fellow researchers use Deep Learning Toolbox™ to integrate computational fluid dynamics (CFD) and field data with deep learning. This approach, called physics-informed machine learning, brings the benefits of high-performance computing (HPC) to large data sets. Using MATLAB® enables researchers to reach beyond the computational limits of machine learning alone.