Compositionally complex alloys (CCAs), such as metallic glasses (MGs) and high entropy alloys (HEAs) which usually contain more than five constituent elements in equal or near equal molar fractions, have recently attracted great research interest because of their remarkable mechanical and physical properties. Unfortunately, the design of CCAs is facing great challenges because the traditional experimental approaches based on empirical rules are time consuming and ineffective. In this thesis, I aim to use machine learning (ML), a data-driven approach that could solve multi-dimensional problems without specific programing, as a tool for the accelerated design of CCAs.

First, I made a critical appraisal of the existing phase design rules as commonly used in the literature with three ML algorithms. Based on the artificial neural network algorithm, a sensitivity matrix from the ML modeling could be derived, which enabled the quantitative assessment of how to tune a design parameter for the formation of a certain phase, such as solid solution, intermetallic, or amorphous phase. Furthermore, I explored the use of an extended set of new design parameters, which had not been considered before, for phase design in CCAs with the ML modeling. To verify our ML-guided design rule, I performed various experiments and designed a series of alloys out of the Fe-Cr-Ni-Zr-Cu system. The outcomes of our experiments agree well with the predictions of the ML model, which suggests that the ML-based techniques could be a useful tool in the design of CCAs.

Next, I focused on the compositional design of compositionally complex MGs. Based on a database containing ~5000 different compositions of MGs reported since 1960s, I successfully developed a hybrid ML model. Unlike the prior works relying on empirical descriptors for data representation, I designed theory guided data descriptors in line with the recent thermodynamic models on amorphization in CCAs for the development of the hybrid classification-regression ML algorithms. Our hybrid ML modeling was validated both numerically and experimentally. More importantly, it enabled the discovery of MGs (either bulk or ribbon) through the ML-aided deep search of a multitude of quaternary to senary alloy compositions.

Aside from the supervised ML models, I further developed a generative deep learning framework (GDLF) to generate compositionally complex BMGs directly without human inputs, which is more effective for compositionally complex systems. My GDLF is based on the unsupervised Generative Adversarial Network (GAN) algorithm for data generation, and the supervised Boosted Trees algorithm as the surrogate model for result evaluation. I studied systematically the confounding effect of various data descriptors and the literature data on the effectiveness of our framework both numerically and experimentally. Most importantly, aside from alloy compositions, I demonstrated that the GDLF could enable the inverse design of BMGs, because well-trained generator could produce composition-property mappings, based on which one could quickly pinpoint the compositions that can attain the targeted properties. Therefore, this framework could pave the way for the inverse design procedure and greatly accelerate the development of new compositionally complex BMGs.