Perspectives and Final Remarks

Abstract

In this chapter, we consider perspectives in deep learning for fluid animation. Firstly, the input data generation for neural network training depends on simulating fluid dynamics using traditional methods to produce the training data, which is time-consuming. To address this issue, we describe the basic model for physics-informed neural networks (PINNs) applied to the Navier–Stokes simulation. In this case, the fluid equations and the initial and boundary conditions are included in the model as new terms in the loss function, which can reduce the necessity for large training databases. Secondly, the success of deep architectures happens with the increase in the difficulties of understanding how deep neural networks come to decisions. This fact motivates the development of explainable artificial intelligence (XAI) techniques, like Shapley additive explanations (SHAP), to quantify the importance of hidden layers for model outcomes. From the rendering perspective, machine learning is a new approach that brings many exciting possibilities to improve visual results for the entertainment industry. But at the same time, it still has drawbacks to scientific applications, where visual accuracy is a vital requirement. We end the material by discussing how machine learning techniques improve the rendering algorithms.