Among the major challenges of the 21st century, whether it is studying how to produce clean electricity or developing high-temperature superconductors, we need to be able to design new materials with specific properties. If this is to be done on a computer, electrons need to be simulated, because these subatomic particles dominate how atoms bind components, but also responsible for the flow of current in solids.
It can be said that understanding the position of electrons within a molecule is crucial to explaining the structure, properties, and reactivity of molecules. Despite decades of effort and some significant advances, scientists have made some significant advances, but to date, accurately simulating the quantum mechanical behavior of electrons remains a daunting challenge.
Now, in a paper newly published in the journal Science, The DeepMind research team is introducing a new neural network, DM21, that can understand the characteristics of molecules by predicting the distribution of electrons within a molecule. It can calculate the properties of some molecules more accurately than existing techniques.
Theoretically, the structure of materials and molecules is determined entirely by quantum mechanics. Nearly a century ago, physicist Erwin Schr dinger proposed the famous Schrödinger equation, which describes the behavior of particles in quantum mechanics. However, it is very difficult to apply this equation to electrons in molecules. Because all electrons interact, this makes it extremely difficult to calculate molecular structures or molecular orbitals based on Schrödinger's equations, which seem to require us to be able to track the position of each electron. Such a job, even in the case of a small number of electrons, is a nightmarishly complex task.
A major development occurred in the 1960s. At that time, theoretical physicists Pierre Hohenberg and Walter Kohn realized that there was no need to track the behavior of each electron individually, and that knowing the probability of any electron at each location, i.e. electron density, was enough to accurately calculate all interactions.
After proving this, Cohen developed density functional theory (DFT), which helps us to accurately approximate the distribution of electrons in molecules. While DFT involves some degree of approximation in nature, it is the only practical method that can be used to study how and why matter behaves in a certain way at the microscopic level. As a result, since the 1960s, DFT has been one of the most widely used technologies in all fields of science. Cohen also won the 1998 Nobel Prize in Chemistry.
However, this technique has its obvious limitations. While it demonstrates a mapping between electron density and interaction energy, the exact nature of this mapping is still unknown for more than 50 years and must be approximated. For decades, researchers have proposed approximations of different precisions for density functionals, but these approximation methods have their own systematic errors, which sometimes lead to DFT failure. For example, they give incorrect results to certain types of molecules, and in one famous example, the DFT method incorrectly predicted sodium chloride, so that even if a chlorine atom is infinitely far away from the sodium atom, the chlorine atom retains an electronic part of the sodium atom.
Now, DeepMind's researchers have successfully implemented learning these functions without system errors by representing this function as a neural network.
The design goal of the DESY neural network is to obtain electron density and output interaction energy. This neural network is capable of making more accurate predictions than traditional DFT methods. | Image credit: DeepMind
During the study, they trained the neural network with data from 1161 precise solutions obtained from Schrödinger's equations. To improve accuracy, they also incorporated some known laws of physics into the network. They then tested the trained system with a set of molecules and came up with stunning results. The new findings suggest that this new AI model is capable of making more accurate predictions and better describing a range of chemical reactions than traditional DFT methods.
Traditional DFT methods, such as B3LYP, are less predictors of how molecules share electron density (blue), while DM21's predictions are closer to reality. | Image credit: DeepMind
In particular, in the results of this study, the researchers solved two problems in the long-standing and traditional functionals.
The first problem is called de-local error: in DFT calculations, the functional determines the charge density of a molecule by finding an electron configuration that minimizes energy. Therefore, errors in the functional cause errors in the calculated electron density. Most existing approximations to density functionals do not allow electron density to be precisely confined to a molecule or atom, but tend to scatter electron density around several atoms or molecules in an impractical form.
Another problem is called spin symmetry breaking: when describing the destruction of chemical bonds, existing functionals tend to unrealistically favor a configuration in which the fundamental symmetry is broken, which is spin symmetry. Since symmetry plays a vital role in our understanding of physics and chemistry, this artificial destruction of symmetry becomes a major problem for existing functionals.
Theoretically, any chemical-physical process involving charge motion is prone to exit error; any process involving bond breaking is prone to spin symmetry breaking. The motion of charges and the breaking of bonds are at the heart of many important technical applications, but it can also lead to serious problems with functionals in describing some of the simplest molecules, as in the case of sodium chloride we mentioned earlier.
DFT is such an important technology that it's not hard to understand that scientists must ensure the correctness of simple chemistry before expecting these functionals to explain more complex molecular interactions. By using neural networks to represent functionals and train datasets to capture the fractional electronic behavior of the desired precise functional, the new study solves both of these problems. Their functionals have shown a high degree of accuracy in extensive, large-scale benchmarks.
Computer simulation plays a central role in modern engineering, making it possible to provide answers to a variety of practical application questions.
We've used them to simulate whether bridges can be supported, whether rockets can lift off, and as technology moves more and more to the quantum scale, hopefully they'll also help us explore questions about materials, drugs, and catalysts, as well as many that we've never seen or even imagined.
Now, deep learning has shown the ability to accurately simulate matter at the level of quantum mechanics. Letting artificial intelligence calculate electron density is perhaps the most ambitious attempt DeepMind team has ever made, and it's the ultimate result of DFT calculations.
#创作团队:
Compilation: Light rain
#参考来源:
https://deepmind.com/blog/article/Simulating-matter-on-the-quantum-scale-with-AI
https://cen.acs.org/physical-chemistry/computational-chemistry/Machine-learning-solves-long-standing/99/web/2021/12
#图片来源:
Cover image: DeepMind