Researchers use AI to find new magnetic materials | Hacker News Reader

Researchers use AI to find new magnetic materials

phys.org

72 pointsrbanffy19 comments

There is a company called Niron Magnetics based out of Minnesota that looks promising. They are making an Iron Nitride magnet that supposedly is better than neodymium. Since it's made out of iron and nitrogen it should also be a lot cheaper to produce.

Iron nitride is in many respects not better than neodymium, but certainly very promising nonetheless.

Diamonds are made of just carbon.

And your point is?

Diamonds are a cartel backed, price gouged, selectively distributed, commodity whose only production cost is digging them out of the ground.

On top of that they are pretty much the most common gemstone.

The cartel I mentioned earlier has also had a heavy handed approach towards the producers of artificial diamond.

All of this is a long winded way of saying “what is your point?” Diamonds and Common earth element man made magnets have absolutely nothing in common.

It’s just a brief counterexample aiming to challenge the parent’s argument that the relative cost of a material is determined by the relative abundance of its constituent elements.

Industrial diamonds are not that expensive.

And are pretty cheap to produce relative to mine at this point

Since diamonds are made of just carbon, they are cheaply produced using chemical vapor deposition.

https://fortune.com/2023/09/03/diamond-demand-falling-lab-gr...

Yep. And the DeBeers cartel manufactured their association with marriage despite being more common and more expensive than ruby.

in plain language: they've made a statistical model that predicts curie temperature based on a few data points; the researchers justly called it ML, the hype hunrgy copywriters called it AI (it helps to lie a little to get more clicks and show more ads). Edit: it would be AI if that model, after a day long analysis of the raw data, came up with a new principle explaining this data, and wrote down a formula with some explanations.

Linear and logistic regression are also AI. Language models do not come up with new principles and ideas but they're still called AI. The AI these days is just a marketing term for "we took some data and fit a curve to it".

In this case I think the application of optimization to predict material properties is a very good use case for AI. AI after all is basically regression scaled up with lots of variables and large data sets. Anything with well defined inputs and outputs should be using "AI" to reduce human effort in uncovering statistical patterns that can accelerate research and development. Material science and drug chemistry are obvious use cases. Compiler optimization is another one but I haven't yet seen any real effort for optimizing software development with AI other than copilots for code completion.

I had a similar reaction, a better headline (to me) is researchers use statistics to predict superconducting magnetic metals.

So sick of AI headlines now.

Does the distinction between AI and ML really matter outside of AI/ML research?

ML is a discipline, like engineering, and AI is a product with certain characteristics, like a spaceship. Our engineering isn't good enough yet to build a spaceship.

And another group called it statistics, and another one cybernetics. Oh the joy.

They left out any mention of looking for high temperature super conductors ;-)

There are no obstructions to doing that. It is indeed possible to predict properties of superconductors with AI. It's just another regression problem with well defined inputs and outputs. Inputs are the materials and electromagnetic properties of the materials and the output is prediction of superconductivity.

Surprising there is no YC company already working on this. Instead of yet another email company you'd think there would be at least one trying to accelerate R&D for material science and chemistry with AI.

Afaik the fundamental mechanisms governing superconductivity are not known, which means any machine learning guess can only tell you which materials would be statistically likely candidates.

At least the physicists I heard talking about that problem don't see any AI inventing a superconductor directly. Maybe it can help discovering the fundamental mechanisms tho.

Have a look at Predicting Transition Temperature of Superconductors with Graph Neural Networks posted in August.

https://arxiv.org/abs/2308.11160

Reply on news.ycombinator.com