Using Deep Learning to Solve Fundamental Problems in Computational Quantum Chemistry
Introduction
Computational quantum chemistry has come a long way in simulating complex molecular interactions and predicting their properties. The advent of deep learning techniques has further enabled researchers to tackle long-standing challenges in the field. In this blog post, we explore the application of deep learning in solving fundamental problems in computational quantum chemistry, focusing on the development of FermiNet, a neural network architecture for computing the energy of atoms and molecules from first principles.
Research
Using deep learning, we demonstrated how to solve the fundamental equations of quantum mechanics for real-world systems. Not only is this an important fundamental scientific question, but it also has the potential to lead to practical uses in the future, enabling researchers to prototype new materials and chemical syntheses using computer simulation before attempting to create them in the lab.
Our neural network architecture, FermiNet, is well-suited to modeling the quantum state of large collections of electrons, the fundamental building blocks of chemical bonds. We released the code from this study so that computational physics and chemistry communities can build upon our work and apply it to a wide range of problems.
FermiNet and the Challenge of Computing Excited States
While FermiNet focused on the ground states of molecules, understanding how molecules transition to and from excited states when stimulated is a fundamental challenge in computational quantum chemistry. Excited states are critical for understanding how matter interacts with light and have applications in a wide range of fields, from solar panels and LEDs to semiconductors, photocatalysts, and biological processes involving light.
Accurately computing the energy of excited states is significantly more challenging than computing ground state energies. Even gold standard methods for ground state chemistry, such as coupled cluster, have shown errors on excited states that are dozens of times too large. Our research proposes a solution to this challenge, a novel approach to computing excited states that is more robust and general than prior methods.
Results and Conclusion
Our approach, validated on a wide range of benchmarks, showed highly-promising results. On a small but complex molecule called the carbon dimer, we achieved a mean absolute error (MAE) of 4 meV, five times closer to experimental results than prior gold standard methods reaching 20 meV. We also tested our method on some of the most challenging systems in computational chemistry, where two electrons are excited simultaneously, and found we were within around 0.1 eV of the most demanding, complex calculations done to date.
Frequently Asked Questions
Frequently Asked Questions
Question 1: What is FermiNet?
FermiNet is a neural network architecture developed by researchers to solve the fundamental equations of quantum mechanics for real-world systems.
Question 2: What is the significance of computing excited states in computational quantum chemistry?
Computing excited states is critical for understanding how matter interacts with light and has applications in a wide range of fields.
Question 3: What are some of the challenges in computing excited states?
Computing excited states is significantly more challenging than computing ground state energies, and even gold standard methods have shown errors on excited states.
Question 4: How did you validate your approach to computing excited states?
We validated our approach on a wide range of benchmarks, achieving highly-promising results on small and complex molecules, and on some of the most challenging systems in computational chemistry.
Question 5: What are the potential applications of this research?
The potential applications of this research are vast, ranging from understanding the behavior of materials in biological processes involving light to developing more efficient solar panels and LEDs.
