Updated 11 Jun 2021
In this repository, we implemented and trained the SciNet network described in: Iten, R., Metger, T., Wilming, H., Rio, L., & Renner, R. (2020). Discovering Physical Concepts with Neural Networks. Phys. Rev. Lett., 124, 010508. [arXiv:1807.10300v3]. SciNet attempts to formalize a simplified view of physical modelling thinking process and translate it into a neural network architecture.
Unlike other deep-learning approaches to the description of physical systems, in addition to the input-output mapping SciNet automatically provides information on the number of independent underlying physical parameters given the data at hand.
The network architecture is that of modified variational autoencoder that allows the user to provide data and ask questions about the physical problem at hand. The code provided allows the user to train a deep neural network on time series describing the damped oscillations of a linear oscillator. Training is performed by providing to the network:
Once trained, the network is used as follows:
The repository contains MATLAB® implementation code as well as MATLAB trained models. The demo is implemented as a MATLAB project. The project will manage all paths and shortcuts you need. To Run:
Requires MATLAB release R2019b or newer
The license for Discovering physical concepts with neural networks is available in the LICENSE.TXT file in this GitHub repository.
Copyright 2021 The MathWorks, Inc.
Moubarak Gado (2021). Physical-Concepts-Scinet (https://github.com/matlab-deep-learning/Physical-Concepts-Scinet/releases/tag/v1.0.0), GitHub. Retrieved .
Find the treasures in MATLAB Central and discover how the community can help you!Start Hunting!