Arvind Mohan
@arvindmohan.bsky.social
Scientific ML for PDEs, Fluid Dynamics & Earth Sciences. Scientist @Los Alamos National Lab. Aerospace Engineer and Mountaineer.
Opinions my own, not LANL/ US DOE
Opinions my own, not LANL/ US DOE
Or more accurately, discussing numerical methods for gradient descent while undergoing (rapid) gradient descent ⛷️
Talking numerical methods with @arvindmohan.bsky.social in our second, better office
December 2, 2024 at 2:18 AM
Or more accurately, discussing numerical methods for gradient descent while undergoing (rapid) gradient descent ⛷️
Reposted by Arvind Mohan
I'm really excited about this paper. Some context for 🔭🧪🔬 folks, as the AI summary may be a bit dry...
A common activity in scientific ML is to train AI models on numerical data, generated by simulation. The simulation data is assumed to be ground truth... 🧵
A common activity in scientific ML is to train AI models on numerical data, generated by simulation. The simulation data is assumed to be ground truth... 🧵
Arvind, me, and Jonah released a new pre-print on some pen and paper analysis of fundamental failure modes and old school stability analysis for neural PDEs typically used in AI for Science application. arxiv.org/abs/2411.15101. 1/n
What You See is Not What You Get: Neural Partial Differential Equations and The Illusion of Learning
Differentiable Programming for scientific machine learning (SciML) has recently seen considerable interest and success, as it directly embeds neural networks inside PDEs, often called as NeuralPDEs, d...
arxiv.org
November 29, 2024 at 5:03 PM
I'm really excited about this paper. Some context for 🔭🧪🔬 folks, as the AI summary may be a bit dry...
A common activity in scientific ML is to train AI models on numerical data, generated by simulation. The simulation data is assumed to be ground truth... 🧵
A common activity in scientific ML is to train AI models on numerical data, generated by simulation. The simulation data is assumed to be ground truth... 🧵
Reposted by Arvind Mohan
Cool article by @marccoru.bsky.social et al. exploring the use of spherical harmonics and very shallow SIREN networks to convert longitude and latitude meaningful geospatial embeddings on the sphere (code is also available) arxiv.org/abs/2310.06743
November 28, 2024 at 3:27 PM
Cool article by @marccoru.bsky.social et al. exploring the use of spherical harmonics and very shallow SIREN networks to convert longitude and latitude meaningful geospatial embeddings on the sphere (code is also available) arxiv.org/abs/2310.06743
This is a fantastic resource to make research more accessible!
One of my favourite data discoveries this year: Google's mind-blowing ARCO-ERA5 dataset: hourly data for ~300 climate variables, available globally from 1940! 🤯
Loadable with a single line of Python code from a single cloud-friendly Zarr file! Below: a month of wind waves + swell: 🌊
Loadable with a single line of Python code from a single cloud-friendly Zarr file! Below: a month of wind waves + swell: 🌊
November 28, 2024 at 1:39 AM
This is a fantastic resource to make research more accessible!
To add, even in cases where extrapolation is seen, its likely because of "lucky" interactions in numerical dissipation between the discretization scheme, the grid and the initial condition. Our paper provides formal tools to a priori estimate error and identify extrapolation limits *before* training
Arvind, me, and Jonah released a new pre-print on some pen and paper analysis of fundamental failure modes and old school stability analysis for neural PDEs typically used in AI for Science application. arxiv.org/abs/2411.15101. 1/n
What You See is Not What You Get: Neural Partial Differential Equations and The Illusion of Learning
Differentiable Programming for scientific machine learning (SciML) has recently seen considerable interest and success, as it directly embeds neural networks inside PDEs, often called as NeuralPDEs, d...
arxiv.org
November 26, 2024 at 5:21 PM
To add, even in cases where extrapolation is seen, its likely because of "lucky" interactions in numerical dissipation between the discretization scheme, the grid and the initial condition. Our paper provides formal tools to a priori estimate error and identify extrapolation limits *before* training
Just joined - Where the sky is still blue, but the bird is long gone.
November 26, 2024 at 5:57 AM
Just joined - Where the sky is still blue, but the bird is long gone.