Mateo Sánchez

CS student | AI for Science, not just software | PINNs, CFD, Mars thermal physics

[Mateo Sánchez @tembluso]

First physics project: 12 weeks learning CFD + Physics-Informed Neural Networks

Starting from zero. Going to document what I learn and build as I figure out how to apply AI to fluid dynamics equations.

Repo: https://t.co/SzwKtyuARL

Weekly Thread below

[Mateo Sánchez @tembluso]

@kslote1 Never worked with KANs before. How do they compare with PINNs?

[Mateo Sánchez @tembluso]

@JFajardoValera That is actually very smart. Do you use PINNs only when the GPS degrades or rely on them fully?

[Mateo Sánchez @tembluso]

@Ofirlin From what you found how much accuracy do you trade off for the inference speedup?

[Mateo Sánchez @tembluso]

@Sri26762339 Looks very impressive. I've only used physics informed models on numerical data. How do you bake the physics loss into an image based pipeline like this?

[Mateo Sánchez @tembluso]

@CollinsKimotho Right now I'm working on PINNs and CFD.

[Mateo Sánchez @tembluso]

@Behdad_Jamshidi Looks fascinating but how exactly would this work?

[Mateo Sánchez @tembluso]

@said_gokmen I feel the same. Everything feels very superficial without real depth. AI has much more potential than wrappers and standard SaaS.

[Mateo Sánchez @tembluso]

@elon_lit Damn this is good. The 2x speedup on PINN training is interesting. What problem types did you test it on?

[Mateo Sánchez @tembluso]

@AllanatrixQ very cool. Do you think the model can get to reliable top-1 identification or is that not realistic with current MS/MS data?

[Mateo Sánchez @tembluso]

Week 13/12 recap

https://t.co/Qw0oHsA0TU

[Mateo Sánchez @tembluso]

Week 13 recap

Paper is taking shape.

Underestimated the amount of work it is, but just a few touches left.

Final stretch.

[Mateo Sánchez @tembluso]

First physics project: 12 weeks learning CFD + Physics-Informed Neural Networks

Starting from zero. Going to document what I learn and build as I figure out how to apply AI to fluid dynamics equations.

Repo: https://t.co/SzwKtyuARL

Weekly Thread below

[Mateo Sánchez @tembluso]

Week 12/12 recap

https://t.co/xTC3WAorOc

[Mateo Sánchez @tembluso]

Week 12 recap

Started writing my first scientific paper.

Realized most of my results weren't polished enough. A couple of my convergence claims haven't been rigorously tested.

Found even better solutions.

Paper is 50%done. Extending to week 13 to finish properly.

[Mateo Sánchez @tembluso]

Week 13 recap

Paper is taking shape.

Underestimated the amount of work it is, but just a few touches left.

Final stretch.

[Mateo Sánchez @tembluso]

Week 12 recap

Started writing my first scientific paper.

Realized most of my results weren't polished enough. A couple of my convergence claims haven't been rigorously tested.

Found even better solutions.

Paper is 50%done. Extending to week 13 to finish properly.

[Mateo Sánchez @tembluso]

Week 11 recap

Finished all experiments.

Tried 6 different approaches to model a real-world problem. Most were dead ends but two produced results.

Now polishing results and writing everything up as a proper research paper.

[Mateo Sánchez @tembluso]

I built a 2D axisymmetric PINN solving heat diffusion around the physical sensor geometry.

Forward problem works, inverse problem hit a structural identifiability wall.

Diagnosed exactly why and documented it.

[Mateo Sánchez @tembluso]

Week 10 recap

Implemented the spatial representation of the sensor in the PINN

RMSE dropped significantly but the recovered value was 70% larger than the published benchmark

This is because the sensor itself acts as a thermal fin perturbing the temperature of the soil around.

[Mateo Sánchez @tembluso]

However there is still an issue. The sensor averages over a surface not a point.

Therefore treating it as a point measurement is physically incorrect.

Next I will try to impelment the correct spatial representation.

[Mateo Sánchez @tembluso]

Week 9 recap

Ran inverse problem experiments with the PINN.

It inferred a physical parameter within 7% of the published benchmark value, using a completely different method than traditional approaches.