Microsoft has released MatterGen, an innovative large model dedicated to inorganic material design.

MatterGen's basic architecture is based on a diffusion model, which can gradually optimize atomic types, coordinates and periodic lattices to quickly generate different types of new inorganic materials. For example, in the energy field, MatterGen can generate a new cathode material for lithium-ion batteries.

By adjusting the atomic types, introducing some transition metal elements with special electronic structures, and accurately determining their position coordinates in the lattice, a crystal lattice with a unique microstructure can be developed to continue the battery life. battery life and lifespan.

Compared with traditional discovery methods, MatterGen can increase the rate of generating stable, unique and novel materials by more than 2 times, and bring the generated structures up to 10 times closer to their DFT local energy minimum. Therefore, MatterGen is of great help to high-tech fields such as electric vehicles, aerospace, and electronic chips.

Many friends may be a little confused about this new field, so let’s give a simple and easy-to-understand explanation. If you want to build a house, the traditional method is to choose from existing house plans, which may not meet our special needs.

Now when you use MatterGen to build a house, you can directly say that you want a five-bedroom, one-living room, a gym, an e-sports room, two small bedrooms, and a master bedroom. It is best to build it outside the house. A small garden. The overall house structure adopts Chinese style, and it is best to add some dragons, phoenixes, etc. on the walls to decorate it.

This means that MatterGen discovers complex inorganic materials through the diffusion process and conducts more detailed decomposition and generation. Based on the specific requirements we have given, we can gradually explore and construct the most appropriate material combination and structural layout.

Start with the type of atoms, just like selecting building materials with different textures and characteristics; then carefully determine the coordinate positions of these atoms in space, and accurately place each brick; and finally build a perfect periodic crystal. grid to build a stable and unique house frame.

Actually, many netizens are confused after reading this. Please treat me like a 5-year-old child and explain this technology~

I know that AI is changing everything, and it will eventually happen. But I didn't expect it to come so quickly.

Look at it discovering some amazing superconductors, improving its computing performance, which in turn enhances its ability to discover more superconducting materials, which further improves its computing power, and so on... you get the idea. Imagine that AI is optimizing everything. A critical critical mass has been achieved.

There could be a revolution in battery cell additives, which have been discussed and in demand in the field in recent years. Based on the images provided by Microsoft, it looks like it could also help produce cathode activators.Model of sexual material.

Thanks this model has implemented AGI.

It’s time for AI to solve the problem of global warming.

Is this equivalent to the AlphaFold model in the material world?

In the MatterGen model, the diffusion process is The core mechanism that generates crystal structures. This process is inspired by the phenomenon of diffusion in physics, in which particles move from areas of high concentration to areas of low concentration until a uniform distribution is achieved. In the context of materials design, the diffusion process is cleverly adapted to generate an ordered, stable crystal structure from a completely random initial state.

The diffusion process starts with a random initial structure, which has no physical meaning and is just a random distribution of atoms in space. Then, MatterGen gradually reduces the "noise" in this initial structure through a series of iterative steps, making it gradually closer to a real crystal structure. This process is not a simple random change, but is strictly guided by the laws of physics and principles of materials science.

In each iteration, MatterGen will fine-tune the atom type, coordinates and lattice parameters. These fine-tunings are based on a predefined physically motivated distribution, which means that the model takes into account the actual physical properties of the crystal material, such as bond lengths and angles between atoms, and the symmetry of the crystal lattice, when adjusting atomic positions and types. sex.

For example, coordinate diffusion respects the periodic boundaries of the crystal, adjusting the positions of the atoms through a wrapping normal distribution, ensuring that the atoms do not leave the periodic structure of the crystal. Lattice diffusion adopts a symmetrical form, the mean value of its distribution is a cubic lattice, and the average atomic density is derived from the training data, which ensures that the generated lattice structure is both stable and physically meaningful.

The equivariant fractional network is another key component in the MatterGen model, responsible for learning how to recover the original crystal structure from the diffusion process. The design of this network is based on an important physical principle - equivariance.

Equivariance refers to the characteristic that a system maintains certain properties unchanged under certain transformations. In crystalline materials, this means that the properties of the material remain unchanged under rotation, translation, etc.

The equivariant fraction network is able to output equivariant fractions of atomic types, coordinates, and lattice by learning patterns in the data. These scores represent the "misfit" of each atomic and lattice parameter in the current structure, that is, their deviation from the ideal crystal structure.

By calculating these scores, the network guides the model on how to adjust atomic and lattice parameters to reduce noise in the structure and make it closer to a stable crystal structure. This is also important for MatterGen to improve accuracy and ideality rates. One of the reasons.

In order to increaseTo increase the flexibility of the model, MatterGen has added an adapter module that can be fine-tuned for different downstream tasks and can change the output of the model according to given property labels. (This is the tailor-made function in our case)

The adapter introduces an additional parameter set in each layer of the model, and these parameters can be adjusted according to task-specific property labels. During the fine-tuning process, these parameters are optimized so that the structure generated by the model better meets the requirements of the specific task. This design not only improves the adaptability of the model, but also reduces the amount of labeled data required for fine-tuning, because the model does not need to learn the characteristics of each task from scratch, but can be adjusted based on pre-training.

For example, when designing a new type of battery material, the model may need to focus on the conductivity and ion diffusivity of the material; while when designing a new type of catalyst, the model may need to focus on the surface of the material. activity and selectivity. Adapter modules enable the model to adapt its strategy for generating structures according to these different requirements.

Currently, Microsoft has published the research in "Nature" and has been recognized by many technology giants. It can be compared with Google’s AlphaFold series of protein prediction models that won the Nobel Prize in Chemistry last year.