Myung Min is from Ao Fei Temple
Qubits reports | Official account QbitAI
Now, AI can accurately describe matter at the quantum level!
In the latest issue of science, DeepMind's neural network can predict the distribution of electrons within molecules to calculate molecular properties.
It's only been a week since DeepMind appeared on the cover of Nature and solved two major mathematical puzzles.
This breakthrough has important implications for the fields of AI, chemistry, and materials science.
On the one hand, this means that deep learning has great promise in accurately simulating matter at the quantum level; on the other hand, it has important implications for exploring materials, medicine, catalysts and other substances at the nanoscale.
DeepMind also said that they will open source this work for researchers around the world!
No wonder netizens will sigh:
DeepMind——YYDS！
Nature says this will be one of the most valuable technologies in chemistry:
Solve the problem of electron interaction with MLP
This time DeepMind solves the problem related to density functional theory (DFT).
DFT is a method of studying the electronic structure of multi-electron systems by calculating the electron density within molecules, which can describe matter at the quantum level.
By approximation, the DFT first simplifies the complex electron interaction problem to a non-functional problem, and then puts all the errors in a separate term and analyzes the errors separately.
Over the past few decades, it has become one of the most commonly used methods for predicting the properties of various systems in chemistry, biology, and materials.
However, there are still some limitations to this method.
On the one hand, it has delocalization errors.
In DFT calculations, the functional finds the electron configuration at least energy minimization to infer the electron density of the molecule. This creates an electronic error in the function.
Most existing density functions mistakenly distribute electron densities over several atoms or molecules, rather than identifying them around a single molecule or atom.
△ The left picture is the traditional method, and the right picture is the DeepMind proposed method
Another major error comes from the failure of spin symmetry.
If the chemical bonds in the structure are described as breaking, the existing functionals give a configuration in which the spin symmetry is broken.
However, symmetry plays an important role in the study of physical and chemical configurations, so this defect of the current method also causes a large error.
As can be seen in the comparison, the PBE method breaks the spin symmetry.
From this, DeepMind proposed a neural network, DeepMind 2021 (DM21 for short).
This framework uses a multilayer perceptron (MLP) that maps a set of input vectors to a set of output vectors.
After entering precision chemical data such as spin index charge density into a weight-sharing MLP, it can predict the enhancement value of the local charge density and the local energy density.
After integrating these values, add the dispersion correction DFT to the function.
Once trained, the model can be deployed in self-consistent computation.
In the specific data comparison, the error value of DM21 is lower than that of traditional methods.
That is, DM21 can accurately model complex systems such as hydrogen chains, charged DNA base pairs, and transition states of dual radical systems.
Experimental results show that on different benchmarks (GMTKN55\BBB\QM9), the absolute error value of DM21 is lower than that of ordinary methods.
It is not difficult to conclude that DM21 can construct a more accurate description of electron interactions than the DFT method, and deep learning will have great promise for accurately simulating matter at the quantum level.
AI has been used to shock the biological and mathematical communities
One of the results of this research is google DeepMind research scholar James Kirkpatrick.
He said that understanding microscopic phenomena is of great significance for research on clean electricity and microplastic pollution.
This also has profound implications for researchers to explore issues such as new materials, drug development and catalysts at the nanoscale.
And this isn't the first time DeepMind has shocked the scientific community with AI.
This year, they used AlphaFold2 to predict 98.5% of human proteins, shocking the biological community for a while.
Not long ago, they used AI to break through two major mathematical problems and appeared on the cover of Nature, which had a profound impact on knot theory and representation theory.