Artificial Intelligence–Driven Materials Design for Spintronic Applications

Introduction

The steady increase in energy consumption per capita and the slow transition toward renewable energy sources is becoming a serious global problem, making energy efficiency paramount for new technologies.

Two-Dimensional Materials

Two-dimensional materials offer an encouraging path toward ultra-low-power electronics due to our capability to combine them into Van der Waals heterostructures with tailored quantum properties based on their constituents.

Spin-Orbit Torque Memories

The spin-orbit torque (SOT) memories are technological prospects that consume a fraction of conventional memories' power.

1. They offer superior speed and storage capacity.
2. They were further improved when using 2D materials as building blocks instead of 3D metallic systems.

Theoretical Advances

Recently, theoretical efforts demonstrated the existence of thousands of potentially synthesizable 2D materials, opening an exponentially larger pool to mine for optimized heterostructures which brute-force approaches cannot tackle.

Project Aim

This project aims at developing artificial intelligence that will propose optimized Van der Waals heterostructures for spin-orbit torques.

Methodology

To this end, we will first construct an automatic material assessment (AUTOMATA) tool based on deep neural networks that will perform numerical modeling and quantum transport simulations autonomously to compute the spin-orbit torque efficiencies.

In parallel, we will develop a computer-assisted structure (COMPASS) optimizer that will propose new systems for spin-orbit torques by using an evolutionary strategy.

Integration of Tools

The AUTOMATA tool will rank the candidates generated by the COMPASS optimizer, and we will use those with superior performance to improve the COMPASS optimizer prediction.

The successful combination of these tools will accelerate the development of technologies by automatizing the material selection phase through this quantum mechanical optimization process.

Generalizability

Although we apply it to spin-orbit torques, it is, with little effort, generalizable to any electrical response functions.