Researchers have developed a new machine learning-based method to accurately calculate the formation energies of oxygen vacancies in amorphous silicon dioxide (SiO₂). This breakthrough is crucial because oxygen vacancies are fundamental atomic defects that affect the electrical and optical properties of this ubiquitous material in electronics. The traditional approach, based on density functional theory (DFT), is computationally very expensive for large and complex amorphous systems, limiting the understanding of these defects.

The team trained a machine learning model to predict vacancy formation energies using a database of high-fidelity DFT calculations. This model, termed a "machine learning Hamiltonian," allows for the simulation of much larger systems with greater structural diversity than those accessible with direct DFT. The key lies in its ability to capture complex atomic interactions and local variations in the amorphous structure, which are difficult to model with classical methods.

Results show that the machine learning method not only accurately reproduces DFT-obtained formation energies for known configurations but also allows for the exploration of a much broader configuration space. This has revealed a significantly wider distribution of oxygen vacancy formation energies than previously thought, with direct implications for the stability and functionality of SiO₂-based devices. The computational efficiency of the new method is orders of magnitude superior to DFT, paving the way for large-scale simulations.

This advance is fundamental for materials engineering, as a detailed understanding of defects in SiO₂ is essential for optimizing the fabrication of transistors, memories, and other microelectronic components. The ability to accurately predict how defects affect material properties will enable the design of devices with enhanced performance and reliability. The next steps include applying this method to other types of defects and amorphous materials, as well as exploring its impact on the dynamic properties of these systems.