This repository contains the source code for the paper Yu Yao, Chao Cao, Stephan Haas, Mahak Agarwal, Divyam Khanna, and Marcin, Emulating Quantum Dynamics with Neural Networks via Knowledge Distillation, arXiv 2203.10200 (2022).
The purpose of this code is to illustrate our framework and to replicate the main results of our paper.
We introduce a novel framework for training machine learning-based emulators. It combines ideas of knowledge distillation and curriculum learning. The main goal is to construct a process in which the emulator can extract the basic rules governing the time evolution of a physical system.
The main framework is illustrated below.
First, we simulate the physical system of our interest. Then, we construct individual training examples, and we build a training curriculum. The goal is to represent all the interesting phenomena that we wish to capture. For example, to describe the propagation of quantum packets, we must capture dispersion, scattering, tunneling, and quantum wave-interference. In the next step, we use our curriculum of simple examples to train the emulator. To test the generalization capability of our machine learning model, we measure whether the emulator can predict the evolution of more complex systems.
These instructions are for machines with the Linux Operating System. Installation on Windows or macOS should be similar, just make sure that the gcc compiler for C is installed in the system.
Clone this Repository
git clone https://github.com/yaoyu-33/quantum_dynamics_machine_learning
cd quantum_dynamics_machine_learningCreate Virtual Environment and Install dependencies
python -m virtualenv qwave
source qwave/bin/activate
pip install -r requirements.txtInstall ffmpeg for rendering simulations
sudo apt update
sudo apt install ffmpegTo see a demonstration of our framework, check the Jupyter notebook in the demonstration folder.
Data Preparation
Run the notebook part1_data_preparation.ipynb to generate the training data. The training examples will be saved to a specified folder on your disk.
Model Training
Run the notebook part2_model_training.ipynb to train the emulators.
Forecasting
Run the notebook part3_forecasting.ipynb to forcast the time-evolution of the quantum system. This notebook also includes a series of tests to compare various model architectures.
Our framework differs from a typical supervised learning in a few important aspects. First, the goal is not to learn patterns from data, rather to extract and synthetize knowledge about the studdied system from a physicly-informed simulator. Second, we construct a balanced curriculum of easy examples to promote the generalization of our emulator. Finally, the goal is to apply the acquired knowledge to complex cases that go beyond the scope of the training curriculum. Thus, in our case, the distributions of training and test examples are essentially different. The idea is to train the emulator on easy examples, and then use it to forecast cases that would be hard to emulate in a standard way.
Our framework can be relevant to researchers working in the fields of quantum chemistry, material science, and quantum optics. It can be used to model quantum devices and to simulate quantum information propagation.
@misc{2203.10200,
Author = {Yu Yao and Chao Cao and Stephan Haas and Mahak Agarwal and Divyam Khanna and Marcin Abram},
Title = {Emulating Quantum Dynamics with Neural Networks via Knowledge Distillation},
Year = {2022},
Eprint = {arXiv:2203.10200},
}