ACCELERATING MICROSTRUCTURE AND MELT POOL PREDICTION IN LASER POWDER BED FUSION PROCESSES USING DATA DRIVEN NEURAL NETWORK

Laser powder bed fusion (LPBF) is a pivotal additive manufacturing process characterized by selective melting and solidification of powder layers to form complex parts.The microstructure of the final part plays a crucial role in determining its properties, which are heavily influenced by process parameters such as laser power, scanning speed, powder layer thickness, and material properties.Understanding the influence of process parameters on melt pool geometry is crucial for minimizing defects due to unmelted powder regions and optimizing hatch spacing.Additionally, the microstructure plays a vital role, as process parameters can lead to the formation of columnar grains, promoting anisotropy in the fabricated parts.Therefore, understanding microstructural characteristics is key to ensuring the reliability of manufactured parts.Predicting the microstructure and melt pool geometry in laser powder bed fusion (LPBF) processes via finite element modeling is computationally demanding and time-intensive.To mitigate these challenges, we propose the development of a machine learning model capable of accurately predicting melt pool geometry and microstructure based on process parameter variation, thereby reducing time in the identification of optimal process parameters for further experimental investigation.This study utilizes a physics-based numerical model within ANSYS Additive to predict melt pool geometry and microstructure in the LPBF process.The model's accuracy was validated using single-track metal additive manufacturing (AM) experiments extracted from relevant literature.A comprehensive database was compiled by numerical simulation, including melt pool geometry, microstructure, process parameters, and material IDs.This dataset was then employed to train a data-driven model, with process parameters and material IDs serving as inputs and melt pool width, depth, and microstructure images as outputs.Additionally, the images were resized to 100x200 pixels for computational efficiency.Various neural network architectures, including Multilayer Perceptron (MLP) and Convolutional Neural Networks (CNN), were utilized to enhance the predictive capabilities of the model.These architectures were trained and fine-tuned to achieve precise predictions that closely matched the actual dataset.The study showcased the efficiency of neural networks in accurately predicting melt pool geometry and microstructure, significantly reducing prediction times compared to conventional numerical simulations by several orders of magnitude.Future research direction will focus on incorporating physics constraints into the loss function to accelerate the training process, thereby reducing the number of epochs required.However, the development of a rapid data-driven model with minimal mean absolute error is vital for enhancing part optimization and promoting the widespread adoption of MAM.