Can Deep Learning Replace Physics-Based Simulations in Future Engineering Design?

Physics-based simulation tools like Computational Fluid Dynamics (CFD) and Finite Element Analysis (FEA) are fundamental in modern engineering design. The calculation methods of these tools are based on different basic physical equations, and reliable physical predictions can be obtained by numerical modeling and physical field simulations. Despite this, physics-based high-fidelity simulations often require considerable computational resources, such as high-performance computing hardware, large memory capacity, and long computation time. Particularly in coupled multiphysics fields, complex operating environments, or system-level simulation scenarios.

Distinct from physics-based simulation, deep learning is a data-driven approach that utilizes large datasets for pattern recognition and prediction. It relies on large amounts of experimental or simulation data to learn complex engineering patterns in order to provide predictions of engineering outcomes. It does not need to solve complex physical scenarios step by step using different equations; instead, it learns from the provided samples, trains models, and rapidly estimates the results. Through different training methods, such as Physics-Informed Neural Networks (PINNs) and surrogate models, it can automatically optimize engineering designs, reduce engineering design time, decrease the number of repeated simulations, and lower computational costs.

So far, it is evident that deep learning cannot eliminate the need for experimental data or physics-based simulation data. The limitation of deep learning lies in the fact that its output is entirely dependent on its input data. Once the quality or scope of the data is poorly controlled, or errors are introduced, deep learning models are unable to distinguish between correct and incorrect results. Consequently, their data analysis lacks physical interpretability. Furthermore, when these data are repeatedly and extensively reused, a generalization problem arises. This means that, without the incorporation of new experimental or simulation data, deep learning models are unable to produce novel outcomes. Likewise, when encountering situations beyond the training data, deep learning models cannot provide correct predictions. Therefore, physics-based simulation and deep learning models are two different applications. Rather than replacing one another, they are better suited to work collaboratively.

Learn more about Nextus Innovations Laboratory, Research Support, Project Consultancy