This Deep Learning Algorithm From An Austrian Research Proposes Better Numerical Solutions To The Schrödinger Equation

Artificial Intelligence has made its way into almost everyone’s lives with its latest developments. With the growing research now comes a new algorithm that might play an important role in the field of quantum mechanics. One such deep learning algorithm has been developed by researchers in Austria that possess the ability to come up with a solution to Schrodinger’s equation, which has been a great topic of discussion and one of the major challenges in computational Chemistry.

The Schrödinger equation is an essential equation in quantum mechanics that explains the mechanism of an atom or a molecule. It defines how the wave-like state of a physical system changes over time. The equation, named after physicist Erwin Schrödinger and proposed in the year 1926, is very necessary as the development of any new chemical compound depends on its solution. Furthermore, the Schrödinger equation has been extensively used to calculate the properties of quantum systems, such as the energy levels and wave functions of atoms, molecules, and solid-state materials. It highlights the foundation for the study of quantum mechanics and has also been applied in various other domains, such as quantum chemistry, condensed matter physics, and quantum field theory. With this new algorithm, comparatively more accurate numerical solutions for several different molecules can be generated.

The researchers have combined two major neural network architectures to develop this new deep learning algorithm. The first architecture, PauliNet, which the researchers of Berlin have developed, maximizes physical prior knowledge by using the output of the traditional quantum Chemistry method, CASSCF (Complete Active Space Self Consistent Field), as the envelope function. It uses a network with relatively fewer weights to come up with approximate solutions and thus works more quickly. The other architecture, FermiNet, developed by Google’s DeepMind, uses a basic exponential function as the envelope function and a neural network with large weights to produce more precise solutions or energies. 

With the contribution of the two architectures and modifications in the embedding and input features, the deep learning algorithm can produce accurate numerical solutions for the electronic Schrödinger equation. The team has combined this architecture with VMC, i.e., the Variational Monte Carlo approach, to estimate precise ground state energies for various molecules and atoms. This algorithm reduces energy errors by 40 to 70% compared to other deep learning approaches.

The researchers found that by increasing the physical prior knowledge, such as by using CASSCF or some other pre-training, the accuracy didn’t necessarily increase but instead instilled some biases in the architecture. While searching for a suitable starting point in the ground state, excessive pre-training led the model to miss the ground state. Thus the team shows the impact of the growing prior knowledge. Considering the effect and after other additions, the algorithm generates a solution at 6x lower computational cost than classic approaches.

This new development sounds promising as the Schrödinger equation is one of the cornerstones of quantum mechanics as it unfolds how systems evolve. Finding precise solutions to it has always been a tough nut to crack. Thus, this new algorithm is unquestionably great progress and, as compared to previous methods, produces the best solutions as of now.

Check out the Paper and Reference Article. All Credit For This Research Goes To the Researchers on This Project. Also, don’t forget to join our Reddit Page, Discord Channel, and Email Newsletter, where we share the latest AI research news, cool AI projects, and more.