Physics-Informed Bias Method for Multiphysics Machine Learning: Reduced Order Amyloid-β Fibril Aggregation

The simulation of multiphysics problems is a popular research area—and one of the potential concerns too—for researchers studying a wide range of quantifiable systems. For one, reduced order mechanical models are often the best way to deduce inferences about physically complex systems. Obtaining numerical solutions to the differential equations underpinning reduced order models can become computationally expensive, especially as complexity grows. Many multiphysical computations can be limited by this burden. Machine learning (ML) offers tools which can avoid the prohibitive computational expense when making predictions on traditional multiphysics models. However, deep ML techniques needed to properly simulate complex systems require large amounts of data, thus revisiting the computational expense of training on numerically derived data and/or observational challenges of training with empirical data. Instead, ML techniques can be trained by reinforcing or integrating the understood physics of the system. Such a physics-informed approach to machine learning leverages the understood physical qualities of the system for the creation of more accurate and efficient ML models. This work presents an example of physics-informed machine learning that alleviates the common costs associated with multiphysical modeling, with dependable accuracy. More specifically, the problem discussed here is an example of automatic training dataset generation for an informed ML approach to modeling Amyloid-β fibril aggregation.