#SciML
These are public posts tagged with #SciML. You can interact with them if you have an account anywhere in the fediverse.
@chrisrackauckas@fosstodon.orgIf you work in controls, you know: write C code for real-time embedded hardware. You can't use #python or #rstats etc. for that, right? With #julialang v1.12, we demonstrate it's possible to ahead of time compile to small binaries for use in controls applications. #sciml
Since time immemorial an old adage has always seemed…
arXiv.org
@chrisrackauckas@fosstodon.orgUsing higher order automatic differentiation to improve stiff ODE solvers? Using a third order Newton-like method (Halley's) inside the #sciml #julialang ODE solvers with Taylor-mode AD, ~25% faster. This shows a path for non-standard automatic differentiation to become standard within numerical algorithms and is an example of symbolic-numeric programming outperforming standard numerical algorithms. See the manuscript for details!
arxiv.org/abs/2501.16895
@chrisrackauckas@fosstodon.orgThe problem of building neural surrogates #sciml for real-world industrial problems is not a problem of choosing neural network architectures, it's a problem of gathering the right training data from the model you're seeking to emulate. We demonstrate this on a turbofan jet engine, achieving 0.1% relative error through an active learning process. This is one of the demonstrations from #scitech showcasing the advancements of industrialization of #SciML
Details: arxiv.org/abs/2501.07701
@chrisrackauckas@fosstodon.orgNew fully adaptive Radau IIA method, achieves state-of-the-art performance for high accuracy on highly stiff ODEs. It has a fully automated order construction with adaptive order, and thus if you use higher precision numbers it can automatically construct 17th, 21st, etc. order versions of the method on the fly. Outperforms the classic Hairer Fortran implementation of radau by about 2x across the board!
For more details see: https://arxiv.org/abs/2412.14362 #julialang #sciml
@chrisrackauckas@fosstodon.orgNew version of a very good ODE solver today! IRKGaussLegendre released a SIMD and multithreaded mode. 16th order Implicit Runge-Kutta integrator IRKGL16 for non-stiff symplectic equations which require high accuracy.
For more benchmarks, see https://github.com/SciML/IRKGaussLegendre.jl/blob/master/Benchmarks/NLS-WorkPrecision.ipynb
#politics news is getting on top of me today. To calm down I watched some videos explaining how to implement #stochasticdifferentialequations in #julialang.
Thanks, #SciML -- I needed a break.
It looks like Optim.jl and IPNewton are probably not really recommended and I should switch to something like Ipopt.jl and OptimizationMOI.jl which I'm trying... #sciml #julialang
@chrisrackauckas@fosstodon.orgSolving f(x)=0 is just Newton's method, right? Well the #julialang nonlinear solvers have had lots of innovations in this very common numerical problems. Our nonlinear solvers demonstrate robustness where #SciPy fails to converge, high performance automated sparsity via integrated compiler tricks, and many more tidbits. Together, this makes NonlinearSolve.jl simple to use yet hit the highest level of performance from naive usage. #sciml
Efficiently solving nonlinear equations underpins numerous…
arxiv.org @chrisrackauckas@fosstodon.orgRequirements of scientific machine learning (#SciML) for surrogates in industry does not match how that academics think! New preprint that describes how SciML metrics need to be reconsidered in the context of industrial requirements. #julialang
@DeutscherWetterdienst Der Titel des Artikels ist schon reichlich absurd, denn er impliziert, dass zur Zeit nicht "Maschinen" sondern etwa fleißige Menschen mit Stift und Papier die Berechnungen zur Wettervorhersage durchführen.
Und dann diese eigenartige, fast schon feindselige Gegenüberstellung von Numerik und machine learning. Würde mich ja sehr wundern, wenn nicht auch schon in den aktuellen Modellen Parameter anhand historischer Daten geschätzt wurden.
Ganz zu schweigen von #sciml ...
@chrisrackauckas@fosstodon.orgIs your software stack #quantum ready? The #julialang #sciml differential equation solvers are able to to not only target CPUs, GPUs, and IPUs with good performance, but quantum computers as well through the QuDiffEq.jl backend without changing your code. Check out this work where a group of researchers tested its accuracy for modeling power systems dynamics, showing its correctness and readiness for real-world DAEs!
Power system dynamics are generally modeled by high…
arxiv.org @chrisrackauckas@fosstodon.org#julialang #sciml for perfume engineering. Mixing physical models with machine learned quantities in order to predict and classify odors.
https://chemrxiv.org/engage/chemrxiv/article-details/656703395bc9fcb5c9a3a476
@chrisrackauckas@fosstodon.org#julialang GPU-based ODE solvers which are 20x-100x faster than those in #jax and #pytorch? Check out the paper on how #sciml DiffEqGPU.jl works. Instead of relying on high level array intrinsics that #machinelearning libraries use, it uses a direct kernel generation approach to greatly reduce the overhead.
https://www.sciencedirect.com/science/article/abs/pii/S0045782523007156
@chrisrackauckas@fosstodon.org#Julialang #SciML solving nonlinear systems of equations w/ automated detection of sparsity patterns, symbolic simplification, and of course faster (GPU) solvers all right out of the box with automatic differentiation support. Faster than #scipy and #matlab!
Quoted in the Nature article "Is AI leading to a reproducibility crisis in science?" by Philip Ball, I may sound a bit harsh, but it's the truth…
#SciML #reproducibility
https://www.nature.com/articles/d41586-023-03817-6
@chrisrackauckas@fosstodon.orgNew #sciml in Nature Machine Intelligence: Physics-enhanced deep surrogates for partial differential equations
Uses a #julialang neural network to give an augmented course-grained geometry that corrects the biases to better predicts the high-fidelity behavior.
@chrisrackauckas@fosstodon.orgNew talk on the kinds of numerical issues that can occur when doing scientific machine learning (#sciml) and physics-informed machine learning (#piml), along with some techniques for avoiding potential numerical instabilities! Derives cases where automatic differentiation fails, and explains how #juilalang tools work around these issues.
[Chris Rackauckas](https://youtu.be/vG6ZLhe9Hns?t=4870) discusses symbolic numerics in scientific computing. To paraphrase, " "It's all transformations of code, whether procedural (as in numerics) or symbolic (as in math.)" #juliacon #julia #sciml #climate
I suppose this year's #juliacon conference is really three conferences in one. Bit of logo overload, perhaps, I wonder if something like this ties it together ... 
#julia #julialang #juliacon2023 #SciML #jumpdev
I learnt what a lemniscate was today..
