Rational Design of Multi-component Materials via Ab-initio Calculations and Machine Learning

Abstract

Multi-component materials (MCMs) with disordered elemental arrangements attract extensive attention because of their excellent mechanical, physical, and chemical properties. Many MCMs display good strength, enhanced hardness, high melting points, irradiation resistance, and corrosion resistance, rendering them with great potential in structural, energy conversion, and energy storage applications. Specifically, the rugged energy landscape provided by MCM may stabilize defects, depressing the diffusivity of interstitials and vacancies. MCM can also form percolation networks that provide special migration channels to boost Li-ion diffusion, which is crucial to developing high-performance battery materials. Moreover, the complex local environments of MCM are promising in breaking the linear scaling relations in catalysis, showing their encouraging applications in energy conversion and environmental protection.

Despite the multiple merits, MCMs are largely reported with equal atomic ratios, and the whole compositional space of MCMs has not been fully explored yet. The reason behind this dilemma is due to expensive traditional trial-and-error approaches, which make experimental exploitation nearly impossible. Besides, it is challenging to reveal the roles played by local environments and parent compositions, and hence, correlations between mechanical and chemical properties are largely unknown.

To alleviate the obstacles in developing high-performance MCMs, I adopt ab-initio calculations powered by density functional theory (DFT) to reveal the shearing and grain-boundary strengthening mechanisms of multi-component intermetallics (MCIs). Moreover, machine-learning models are developed to predict single-phase synthesizability, geometrical distortion of local environments, and mechanical properties of multi-component transition-metal carbides (MTMCs). Additionally, a deep learning model on the basis of atomic graph attention (AGAT) networks is constructed to predict and design excellent high-entropy electrocatalysts (HEECs). Active learning combing AGAT, conditional generative adversarial network (CGAN), and k-nearest neighbor (KNN) clustering is formulated to further reduce the size of the required database and accelerate the rational design loops. Further experiments are conducted to corroborate the recommendations posited by DFT calculations and machine learning predictions.

My results show that DFT can effectively uncover the root cause of peculiar properties of MCMs. Specifically, the local distortion in MCMs plays a key role in the deviations from rule-of-mixtures and classical scaling relations. Additionally, the local distortion, compositional space, and properties of MCMs can be efficiently predicted and exploited by machine-learning algorithms developed in this work. Moreover, the robustness of machine-learning models are validated by experiments.

Date of Award	7 May 2024
Original language	English
Awarding Institution	City University of Hong Kong
Supervisor	Shijun ZHAO (Supervisor)