MatterGen is "a generative AI tool that tackles materials discovery from a different angle. Instead of screening the candidates, it directly generates novel materials given prompts of the design requirements for an application. It can generate materials with desired chemistry, mechanical, electronic, or magnetic properties, as well as combinations of different constraints. MatterGen enables a new paradigm of generative AI-assisted materials design that allows for efficient exploration of materials, going beyond the limited set of known ones."

They go on to say people have tried to do this with generative models, evolutionary algorithms, and reinforcement learning.

"Generative models are promising since they can efficiently explore new structures and be flexibly adapted to different downstream tasks. However, current generative models often fall short of producing stable materials according to density functional theory (DFT) calculations, are constrained by a narrow subset of elements, and/or can only optimize a very limited set of properties, mainly formation energy."

The key thing to understand about MatterGen is that it is a diffusion-based generative model. That means it works in a manner similar to image generating models, not language models. Language models take a series of tokens as input (for example representing the words in your prompt) and output a series of tokens (which can be turned back into words). A diffusion model works in a different way. They use this counterintuitive process of removing noise from an image. The way they are trained is by starting with an image and adding tiny bits of Gaussian noise to it, bit by bit turning it from a clear image to pure noise, which looks like that multicolored snow. The model is challenged at each step to learn the reverse step -- how to reverse the noise. This is coupled with a "contrastive" learning system that links the images to text descriptions. This is what enables the diffusion model to remove the noise from an image that starts out as pure random noise in the direction of your text prompt. It is weird that this works, but it does.

Here, though, instead of the diffusion model operating on pixels on a screen, it is operating on a representation of atoms in a material. Instead of removing noise in the direction of your text prompt, it removes noise in the direction of your desired chemical properties, like chemical composition, symmetry, magnetic density, electronic properties, or mechanical properties.

"Compared to previous state- of-the-art generative models for materials, MatterGen more than doubles the percentage of generated stable, unique, and novel materials, and generates structures that are more than 10 times closer to their ground-truth structures at the DFT local energy minimum."

The main paper discusses the experiments done with the model but the details of the model are pushed off to the supplementary materials, so you'll have to get that if you want to know the details of how it works. Basically it has a list of atoms, it has 3D coordinates for where those are positioned inside a 3D lattice cell, and it has additional numbers describing the way the lattice repeats in 3D space, so the system is not limited to cubic lattices but can handle other patterns. The diffusion model performs the "reverse noise" operation on these sets of numbers.

"MatterGen generates stable materials by reversing a corruption process through iteratively denoising a random structure. The forward diffusion process independently corrupts atom types A, coordinates X, and the lattice L towards a physically motivated distribution of random materials. An equivariant score network is pre-trained on a large dataset of stable material structures to jointly denoise atom types, coordinates, and the lattice. The score network is then fine-tuned with a labeled dataset through an adapter module that adapts the model using the encoded property. The fine-tuned model generates materials with desired chemistry, symmetry, or scalar property constraints."

MatterGen: A new paradigm of materials design with generative AI

#solidstatelife #ai #genai #diffusionmodels #chemistry #inorganicchemistry #materialsscience