---
title: "Revisiting MAE datasets and model building"
url: "https://ouro.foundation/posts/mmoderwell/revisiting-mae-datasets-and-model-building"
id: "019980da-4819-7820-8baa-671bb0b1bce8"
type: "post"
---

# Revisiting MAE datasets and model building

Post by @mmoderwell in #permanent-magnets (all) · 2025-09-25

- Asset id: `019980da-4819-7820-8baa-671bb0b1bce8`
- Page: https://ouro.foundation/posts/mmoderwell/revisiting-mae-datasets-and-model-building

# Revisiting MAE datasets and model building

Once again we're at a stopping point because of our inability to effectively predict MAE. Our AI discovery agents have discovered materials that have all the properties we can currently predict. This is great and we should keep these going, however we aren't able to distinguish between a hard and soft magnet in the search/evolution process.

Today I will be working on coming up with a solution to this.

Two approaches I'm thinking about:

- Figure out some AWS service that will let me set up containers that can scale to zero and use the specialized hardware that our HamGNN + TB2J approach takes.
- Or, find a way to make this implementation not dependent on the specialized hardware (recompile scripts).
- Use the datasets we've collected and train a GNN model to predict MAE.

Check out a lot of the prior work here:

> Embedded post: [Building a magnetocrystalline anisotropy energy model, attempt 1](https://ouro.foundation/posts/mmoderwell/building-a-magnetocrystalline-anisotropy-energy-model-attempt-1)

> Embedded post: [Building a magnetocrystalline anisotropy energy model, attempt 2](https://ouro.foundation/posts/mmoderwell/building-a-magnetocrystalline-anisotropy-energy-model-attempt-2)

> Embedded post: [Diving into the Novamag database](https://ouro.foundation/posts/mmoderwell/diving-into-the-novamag-database)

I'm leaning towards the figure out what we have and get that to work. That model is able to predict the Hamiltonian which is useful for understanding many other properties of a material. I'd love to be able to explore more API endpoints around that, so getting it running in a container and behind an API is a prerequisite.

Building a model is also very enticing, especially with how much I've learned since the last time I attempted it.

This time around I'd try to train a model from scratch such that the features the neural network builds/learns are specific to predicting MAE. That was a problem last time. Data availability and scale is still a problem though.

## Assets in this post

- [Building a magnetocrystalline anisotropy energy model, attempt 1](https://ouro.foundation/posts/mmoderwell/building-a-magnetocrystalline-anisotropy-energy-model-attempt-1) (post, id `01970ccd-5c46-7ede-bcdc-318a0c231bfe`)
- [Building a magnetocrystalline anisotropy energy model, attempt 2](https://ouro.foundation/posts/mmoderwell/building-a-magnetocrystalline-anisotropy-energy-model-attempt-2) (post, id `01980bc3-87a3-75f8-9517-ec39be1ffc38`)
- [Diving into the Novamag database](https://ouro.foundation/posts/mmoderwell/diving-into-the-novamag-database) (post, id `019707b9-0c0f-7f34-b78c-691176e49bd0`)
