diff --git a/Manifest.toml b/Manifest.toml index f82c82d2..023011bf 100644 --- a/Manifest.toml +++ b/Manifest.toml @@ -553,8 +553,6 @@ version = "0.8.4" [[KitBase]] deps = ["CUDA", "Conda", "Dates", "Distributions", "FFTW", "FastGaussQuadrature", "FileIO", "JLD2", "LinearAlgebra", "OffsetArrays", "Plots", "ProgressMeter", "PyCall", "SpecialFunctions", "Test"] git-tree-sha1 = "038edf1373543cc02f98eccf48f7acdae29f9f3c" -repo-rev = "main" -repo-url = "https://github.com/vavrines/KitBase.jl.git" uuid = "86eed249-3a28-466f-8d3a-596821e1af9a" version = "0.3.4" diff --git a/paper/paper.bib b/paper/paper.bib index e69de29b..f9726039 100644 --- a/paper/paper.bib +++ b/paper/paper.bib @@ -0,0 +1,55 @@ +@article{bezanson2017, + title={Julia: A fresh approach to numerical computing}, + author={Bezanson, Jeff and Edelman, Alan and Karpinski, Stefan and Shah, Viral B}, + journal={SIAM review}, + volume={59}, + number={1}, + pages={65--98}, + year={2017}, + publisher={SIAM} +} +@inproceedings{jasak2007, + title={OpenFOAM: A C++ library for complex physics simulations}, + author={Jasak, Hrvoje and Jemcov, Aleksandar and Tukovic, Zeljko and others}, + booktitle={International workshop on coupled methods in numerical dynamics}, + volume={1000}, + pages={1--20}, + year={2007}, + organization={IUC Dubrovnik Croatia} +} +@misc{clawpack2020, + title={Clawpack software}, + author={{Clawpack Development Team}}, + url={http://www.clawpack.org}, + note={Version 5.7.1}, + doi={https://doi.org/10.5281/zenodo.4025432}, + year={2020}} +@article{Flux2018, + author = {Michael Innes and + Elliot Saba and + Keno Fischer and + Dhairya Gandhi and + Marco Concetto Rudilosso and + Neethu Mariya Joy and + Tejan Karmali and + Avik Pal and + Viral Shah}, + title = {Fashionable Modelling with Flux}, + journal = {CoRR}, + volume = {abs/1811.01457}, + year = {2018}, + url = {https://arxiv.org/abs/1811.01457}, + archivePrefix = {arXiv}, + eprint = {1811.01457}, + timestamp = {Thu, 22 Nov 2018 17:58:30 +0100}, + biburl = {https://dblp.org/rec/bib/journals/corr/abs-1811-01457}, + bibsource = {dblp computer science bibliography, https://dblp.org} +} +@misc{xiao2020, + title={Using neural networks to accelerate the solution of the Boltzmann equation}, + author={Tianbai Xiao and Martin Frank}, + year={2020}, + eprint={2010.13649}, + archivePrefix={arXiv}, + primaryClass={physics.comp-ph} +} \ No newline at end of file diff --git a/paper/paper.md b/paper/paper.md index 2a2454d5..082873b7 100644 --- a/paper/paper.md +++ b/paper/paper.md @@ -18,17 +18,21 @@ bibliography: paper.bib # Summary -``Kinetic.jl`` is a lightweight Julia toolbox for computational fluid dynamics and scientific machine learning. It focus on the theoretical and numerical studies of many-particle systems of gases, photons, plasmas, neutrons, etc. The finite volume method (FVM) is implemented to conduct 1-3 dimensional numerical simulations for any advection-diffusion type equations. +``Kinetic.jl`` is a lightweight toolbox for computational fluid dynamics and scientific machine learning. It focus on the theoretical and numerical studies of many-particle systems of gases, photons, plasmas, neutrons, etc. Under the hood, the finite volume method (FVM) is implemented with various flux functions and update algorithms. Therefore, 1-3 dimensional numerical simulations can be conducted for any advection-diffusion type equations. Special attentions have been paid on Hilbert's sixth problem, i.e. to build the numerical passage between kinetic theory of gases, e.g. the Boltzmann equation -The main module consists of ``KitBase.jl`` with basic physics and ``KitML.jl`` with neural dynamics. The high-performance Fortran library ``KitFort.jl`` is not included by default, but can be manually imported when the executing efficiency becomes the priority. A Python wrapper ``kineticpy`` has been built as well to locate the structs and methods through ``pyjulia``. The ecosystem is designed to balance the efficiency for scientific research and the simplicity for educational usage. + +and the Navier-Stokes equations. + +``Kinetic.jl`` is an open-source project hosted on GitHub and distributed under MIT license. The main module consists of ``KitBase.jl`` with basic physics and ``KitML.jl`` with neural dynamics. The high-performance Fortran library ``KitFort.jl`` is not included by default, but can be manually imported when the executing efficiency becomes the priority. A Python wrapper ``kineticpy`` has been built as well to locate the structures and methods through ``pyjulia``. It is easy to perform a subtask with the corresponding portable module. +`Kinetic.jl` leverages the Julia programming language [@bezanson2017] for its implementation. The motivation behind it is to balance the flexibility for scientific research, the efficiency for application, and the simplicity for educational usage. Most of the mature FVM libraries, e.g. the OpenFOAM [@jasak2007], are written in compiled languages (C/C++, Fortran), which enjoy the perfect execution efficiency and sacrifice the speed and flexibility for a second development. For example, it is cumbersome to introduce phase-field evolution in OpenFOAM or integrate it with scientific machine learning (SciML) packages. Regarding this situation, the compromise has been made [@clawpack2020], which split the high-level front-ends and the low-level computational back-ends. Usually the front-end is implemented with interpreted languages, e.g. Python, which is able to call the low-level APIs written by compiled languages. Basically it benefits general users, while researchers need to work on the back-end anyhow if a new feature is in desire. Besides, the two-language problem will introduce additional trade-off in development and execution. -An illustrative example of the shock tube problem +Julia programming language is dynamically typed and designed for high performance computing for broad devices. Based on type inference and multiple dispatch, it is an ideal choice to solve the two-language problem. In `Kinetic.jl`, we provide different hierarchies of abstraction. At the highest level, it is feasible to model and simulate a typical gas dynamic problem within 10 lines of code. At the lowest level, we design the methods for general `Number` and `AbstractArray` , so that it holds the perfect possibility to cooperate with existing packages in Julia ecosystem. As an example, `KitML.jl` depends `Flux.jl` [@Flux2018] to conduct scientific machine learning problems. -configuration file `config.txt` +In the following, we present an illustrative example of the shock tube problem in gas dynamics. With the configuration file `config.txt` set as below, ``` @@ -68,107 +72,34 @@ alphaRef = 1.0 omegaRef = 0.5 ``` -and execute the following codes +we execute the following codes ```julia using Kinetic -ks, ctr, face, t = initialize("config.txt") -t = solve!(ks, ctr, face, t) -plot_line(ks, ctr) +set, ctr, face, t = initialize("config.txt") +t = solve!(set, ctr, face, t) +plot_line(set, ctr) ``` -`solve!` is equivalent as the low-level solution algorithm +The computational setup is stored in `set` and the control volume solutions are stored in `ctr` and `face`. The high-level solver `solve!` is equivalent as the following solution algorithm. ```julia dt = timestep(ks, ctr, t) nt = Int(floor(ks.set.maxTime / dt)) res = zeros(3) for iter = 1:nt - Kinetic.reconstruct!(ks, ctr) - Kinetic.evolve!(ks, ctr, face, dt) - Kinetic.update!(ks, ctr, face, dt, res) + reconstruct!(ks, ctr) + evolve!(ks, ctr, face, dt) + update!(ks, ctr, face, dt, res) end ``` +The result is visualized with built-in function `plot_line`, which presents the profiles of gas density, velocity and temperature inside the tube. + - -``Kinetic.jl`` is interested in theoretical and numerical studies of many-particle systems of gases, photons, plasmas, neutrons, etc. It employs the finite volume method (FVM) to conduct 1-3 dimensional numerical simulations on CPUs and GPUs. Any advection-diffusion-type equation can be solved within the framework. Special attentions have been paid on Hilbert's sixth problem, i.e. to build the numerical passage between kinetic theory of gases, e.g. the Boltzmann equation - - - -e.g. compressible gas dynamics. - - - -Hilbert's sixth problem - - -cubersome -phase field - - - -``Oceananigans.jl`` is designed for high-resolution simulations in idealized -geometries and supports direct numerical simulation, large eddy simulation, -arbitrary numbers of active and passive tracers, and linear and nonlinear -equations of state for seawater. Under the hood, ``Oceananigans.jl`` employs a -finite volume algorithm similar to that used by the Massachusetts Institute of -Technology general circulation model [@Marshall1997]. - -![Fig. 1](free_convection_and_baroclinic_instability.png) -Fig. 1: (Left) Large eddy simulation of small-scale oceanic boundary layer -turbulence forced by a surface cooling in a horizontally periodic domain using -$256^3$ grid points. The upper layer is well-mixed by turbulent convection and -bounded below by a strong buoyancy interface. (Right) Simulation of -instability of a horizontal density gradient in a rotating channel using -$256\times512\times128$ grid points. A similar process called baroclinic -instability acting on basin-scale temperature gradients fills the oceans with -eddies that stir carbon and heat. Plots made with `matplotlib` [@Hunter2007] -and `cmocean` [@Thyng2016]. - -``Oceananigans.jl`` leverages the Julia programming language [@Bezanson2017] to -implement high-level, low-cost abstractions, a friendly user interface, and a -high-performance model in one language and a common code base for execution on -the CPU or GPU with Julia’s native GPU compiler [@Besard2019]. Because Julia is -a high-level language, development is streamlined and users can flexibly specify -model configurations, set up arbitrary diagnostics and output, extend the code -base, and implement new features. Configuring a model with `architecture=CPU()` -or `architecture=GPU()` will execute the model on the CPU or GPU. By pinning a -simulation script against a specific version of Oceananigans, simulation results -are reproducible up to hardware differences. - -Performance benchmarks show significant speedups when running on a GPU. Large -simulations on an Nvidia Tesla V100 GPU require ~1 nanosecond per grid point per -iteration. GPU simulations are therefore roughly 3x more cost-effective -than CPU simulations on cloud computing platforms such as Google Cloud. A GPU -with 32 GB of memory can time-step models with ~150 million grid points assuming -five fields are being evolved; for example, three velocity components and -tracers for temperature and salinity. These performance gains permit the -long-time integration of realistic simulations, such as large eddy simulation of -oceanic boundary layer turbulence over a seasonal cycle or the generation of -training data for turbulence parameterizations in Earth system models. - -``Oceananigans.jl`` is continuously tested on CPUs and GPUs with unit tests, -integration tests, analytic solutions to the incompressible Navier-Stokes -equations, convergence tests, and verification experiments against published -scientific results. Future development plans include support for distributed -parallelism with CUDA-aware MPI as well as topography. - -Ocean models that are similar to ``Oceananigans.jl`` include MITgcm -[@Marshall1997] and MOM6 [@Adcroft2019], both written in Fortran. However, -``Oceananigans.jl`` features a more efficient non-hydrostatic pressure solver -than MITgcm (and MOM6 is strictly hydrostatic). PALM [@Maronga2020] is Fortran -software for large eddy simulation of atmospheric and oceanic boundary layers -with complex boundaries on parallel CPU and GPU architectures. ``Oceananigans.jl`` -is distinguished by its use of Julia which allows for a script-based interface as -opposed to a configuration-file-based interface used by MITgcm, MOM6, and PALM. -Dedalus [@Burns2020] is Python software with an intuitive script-based interface -that solves general partial differential equations, including the incompressible -Navier-Stokes equations, with spectral methods. - -``Kinetic.jl`` is an open-source project hosted on GitHub and distributed under MIT license. +The examples of scientific machine learning are introduced systematically in the paper [@xiao2020] . # Acknowledgements