Alloy Design and Process Condition Optimization using Machine Learning with Computational Simulation

Abstract

The present doctoral dissertation underscores the potential of machine learning as an instrumental tool in materials engineering. The study involves a dual-pronged approach: devising a neural network model for the design of high entropy alloys (HEAs) and optimizing the Laser Powder Bed Fusion (L-PBF) process. In the first approach, we develop a specially-structured neural network model, designed to navigate the expansive compositional space of HEAs. The model's unique architecture allows it to capture the characteristics of HEAs' constituent elements. By integrating thermodynamics descriptors into the model's input, we enhance its predictive ability concerning the mechanical properties of HEAs. We then employ a conditional random search, an efficient strategy for discovering local optima, as our inverse predictor, resulting in the design of two innovative HEAs. Our experimental validation reveals that these HEAs exhibit an optimal blend of strength and ductility, thereby affirming the effectiveness of our model and alloy design method. We further elucidate the strengthening mechanism of the designed HEAs, focusing on their microstructure and lattice distortion effects. This novel alloy design technique holds promise for researchers, enabling them to create a myriad of novel alloys with intriguing properties. Our second approach introduces a method capable of predicting L-PBF process conditions that yield a relative density ≥ 98% across a diverse range of powder materials. This approach involves the development of an XGBoost model, which relies on a dataset comprised of powder material properties, process conditions, and the consequential relative density. To heighten accuracy, we apply a sigmoid function to the model's output, relative density. The relationships between input features and the target value were thoroughly examined using Shapley additive explanations. Our experimental validation, conducted with STS 316L, AlSi10Mg, and Fe60Co15Ni15Cr10 MEA powders, validates the capability of our method to predict optimal processing conditions, ensuring high relative density or minimal porosity, regardless of the powder's composition. This study contributes to L-PBF additive manufacturing by providing a universally applicable strategy to enhance process conditions.