Generalized neural-network representation of high-dimensional potential-energy surfaces

Area of Science:

• Computational chemistry
• Materials science
• Machine learning applications

Background:

• Density-functional theory (DFT) is crucial for chemical process accuracy but computationally expensive.
• Simulating large systems with DFT is often unfeasible due to high computational cost.

Purpose of the Study:

• Introduce a novel neural network (NN) representation for DFT potential-energy surfaces.
• Develop a method that is orders of magnitude faster than DFT while maintaining high accuracy.
• Enable efficient simulations of large chemical systems.

Main Methods:

• Developed a new neural network architecture to represent DFT potential-energy surfaces.
• The NN model calculates energy and forces based on atomic positions.
• Validated the method's accuracy and speed against DFT and empirical potentials for bulk silicon.

Main Results:

• The neural network representation is several orders of magnitude faster than DFT.
• Demonstrated high accuracy for bulk silicon, comparable to DFT.
• The method is applicable to both periodic and nonperiodic systems of arbitrary size.

Conclusions:

• The developed neural network method offers a computationally efficient and accurate alternative to DFT for simulating chemical processes.
• This approach significantly advances the feasibility of long simulations for large chemical systems.
• The general applicability across system types makes it a versatile tool in computational science.