Deep Learning in Physics
Deep learning in Physics refers to the use of deep learning models with big data in context of physical functions or systems. Please refer to our article on describing a physical system for more background details. An interesting area is the data-driven inference of temporal evolutions of physical functions with deep learning models. The core idea is that 3D including time of physical functions can be predicted in the latent space of a deep learning network. This creates a deep learning-based simulation algorithm with significant speed-ups compared to traditional algorithms.
The temporal evolution of physical functions that arise in the context of fluid flow problems are one concrete example where deep learning has been succesfully used. This was achieved by using a Long Short-Term Memory (LSTM) model to predict the changes of the pressure field over time. Here the central challenge is a high dimensionality of the Eulerian space and time data sets. But a trained deep neural networks was two orders of magnitudes faster than a traditional pressure solver. We refer to our article on Euler Equation for more details about Eulerian space and time data.
The approach includes a data preparation phase using first a convolutional neural network. One benefit of using a CNN is that it provides a natural way to compute accurate and very efficient non-linear representations. The setup for computing the reduced representation is very important, because it strongly influences how well the prediction network can infer changes over time.
After the data preparation a sequence to sequence learning with neural networks is trained. It learns to predict future representations of the physical space but uses this deep learning model. More Details on using deep learning in physics in this particular context can be found here. In summary the Approach uses a hbyrid LSTM-CNN architecture. This architecture predicts large-scale evolutions of physical 3D functions in learned latent spaces. The evaluation of the approach with many architectural choices and parameters is important. The approach is evaluated with a series of complex liquid simulations. One remarkable fluid flow result was that an initial anvil shape was not part of the training data but the deep learning model successfully generalized to unseen shapes such as the anvil.
Deep Learning in Physics Details
For more pieces of Information about this subject please refer to the following video:
#DeepLearning in #physics enables the use of #bigdata – driven inference of temporal evolutions of physical functions: https://t.co/c4qmLNCmNO pic.twitter.com/OqVIj8i2in
— Big Data Tips (@BigDataTips) March 3, 2019
