Artificial Intelligence for Materials Discovery: Towards Neuromorphic Computing, Quantum Computing and Catalysis
Title: Artificial Intelligence for Materials Discovery: Towards Neuromorphic Computing, Quantum Computing and Catalysis
SNIC Project: SNIC 2021/22-596
Project Type: SNIC Small Compute
Principal Investigator: Jose Luis Lima de Jesus Silva <jose.silva@liu.se>
Affiliation: Linköpings universitet
Duration: 2021-08-19 – 2022-09-01
Classification: 10299
Homepage: https://jseluis.com/research
Keywords:

Abstract

Memristive devices based on 2D materials have been highlighted as potential candidates to solve the bottleneck between memory and processing units on von Neumann architectures. They are usually composed of top/bottom metal electrodes with an intermediate active insulating layer. In addition, these devices have attributes such as non-volatility, CMOS compatibility, high endurance, density integration, and high-speed switching, making them reliable for building the next generation of neuromorphic computing architectures. Low-dimensional nanomaterials have shown functionalities that can mirror biological neurons and synaptic devices since layered 2D materials allow gate voltage tuning due to electron confinement along the semiconductor channel. Materials such as transition metal dichalcogenides and Mxenes can be produced as single-layer or stacked multi-layers through the exfoliation of bulk crystals due to inter-layer weak van der Waals forces. On the other hand, memristive devices relying on large-scale metal-oxide-based (typically TiO2-based insulators) crossbar arrays provide high energy consumption, poor reliability, and low integration density. Additionally, confinement effects of bulk channel materials reduced to sizes below 5 nm is a bottleneck for gate voltage control. As a result, materials that exhibit nano-scale thickness are a promising alternative for the era post-Moore. 2D-based materials have also shown significant advantages in Quantum Computing. For example, topological low dimensional 2D materials can represent quantum bits (qubits), and probing exotic topological quantum states can shed light on potential encoded Fault-tolerant qubits. Therefore, the discovery of superconductors may guide new experiments and possible advances for quantum computing architectures. 2D-based materials have also been used to build micro-devices as reactors for catalytic applications for energy materials for hydrogen production and storage, CO2 reduction, artificial leaves, fuel cells, and batteries. The primary objective of this research is to develop novel methods for the computer-aided discovery of materials using artificial intelligence aiming at potential candidates for neuromorphic computing, quantum computing, and catalysis. Therefore, this research project is based on: (i) Identifying, designing, and benchmark Deep Learning-based architectures (focusing on Graph Neural Networks - i.e., OGCNN, CGCNN AGNN) for materials discovery using open-source databases (Materials Project Database and 2D Materials Database). Apply transfer learning techniques to reuse pre-trained models with data from DFT calculations and Molecular Dynamics. (ii) Apply DFT-based calculations and Deep Learning-based Molecular Dynamics to simulate switching mechanisms of metal-2D materials-metal. The goal is to track mechanisms such as bond breaking and formation, unipolar switching, bipolar switching, and stability of 2D materials. (iii) Simulate multiple nonvolatile resistive-memory devices built as metal-2D Materials-metal structures and use the data for predicting new memristors based on MXenes. In addition, we will use a very similar data-driven machinery based on Graph Neural Networks to find efficient catalysts for diverse applications and hybrid interfaces with applications in quantum computing devices. The success of this data-driven materials discovery project will significantly contribute to the transition for efficient appliances in computing and catalysis.