New research utilizing neural networks, which are inspired by the workings of the brain, presents a potential solution to the difficult task of modeling molecular states.
New research utilizing neural networks, which are inspired by the workings of the brain, presents a potential solution to the difficult task of modeling molecular states.
This study demonstrates how this technique can assist in resolving essential equations in intricate molecular systems.
This advancement may have practical applications in the future, enabling scientists to design new materials and chemical processes through computer simulations prior to their actual laboratory production.
The research, spearheaded by scientists from Imperial College London and Google DeepMind, is featured today in Science.
Excited molecules
The team explored the challenge of understanding how molecules shift between “excited states.” When molecules or materials are energized by significant external energy sources like light or high temperatures, their electrons can temporarily shift to a new arrangement called an excited state.
The specific amounts of energy that are absorbed and emitted during these state transitions create a distinctive fingerprint for each molecule and material. This fingerprint influences the effectiveness of various technologies, including solar panels, LEDs, semiconductors, and photocatalysts. Additionally, it plays a crucial role in biological light-related processes such as photosynthesis and vision.
Nonetheless, modeling this fingerprint is exceptionally challenging because excited electrons are governed by quantum mechanics, which implies that their exact locations within the molecules are uncertain and can only be expressed in terms of probabilities.
Dr. David Pfau, the lead researcher from Google DeepMind and the Department of Physics at Imperial, stated: “Representing the state of a quantum system is exceptionally difficult. Each possible electron position configuration must have a corresponding probability assigned.”
“The configuration space is vast—if you were to create a grid with 100 points along each dimension, the number of potential electron arrangements for a silicon atom would outnumber all the atoms in the universe. This is precisely where we believed deep neural networks could provide assistance.”
Neural networks
The researchers devised a novel mathematical strategy and applied it using a neural network named FermiNet (Fermionic Neural Network). It marked the first instance where deep learning was effectively employed to calculate the energy of atoms and molecules based on fundamental principles with practical accuracy.
The team assessed their method with various examples, yielding promising outcomes. In the case of a small yet complex molecule known as the carbon dimer, they attained a mean absolute error (MAE) of 4 meV (millielectronvolt—a very small energy unit), which is significantly closer to experimental data compared to previous gold standard techniques that had an MAE of 20 meV.
Dr. Pfau remarked: “Our method was tested on some of the most formidable systems in computational chemistry, particularly where two electrons become excited at the same time, and we found ourselves within roughly 0.1 eV of the most intricate calculations achieved to date.”
“Today, we are releasing our most recent findings as open-source, hoping that the research community will use our methods to investigate the surprising ways in which matter interacts with light.”
