Machine-learning Based Microstructure-property Relationships for Ceramics Using High-Throughput, Ultra-fast Laser Sintering
GENG X. 1, TANG J. 1, SARKAR S. 1, SHERIDAN B. 1, LI D. 2, SHI Y. 3, TONG J. 1, XIAO H. 1, PENG F. 1, BORDIA R. 1
1 Clemson University, Clemson, United States; 2 Advanced Manufacturing LLC, East Hartford, United States; 3 Rensselaer Polytechnic Institute, Troy, United States
We report laser sintering of alumina, and a machine learning approach to predict the microstructure and hardness of laser-sintered alumina. Laser sintering allows ultra-fast sintering close to full density within a few tens of seconds. The microstructure and density-grain-size trajectories of laser–sintered alumina is different from those of the furnace–sintered alumina. Therefore, we developed a machine learning (ML) algorithm to predict the microstructure under arbitrary laser power. This algorithm realistically regenerates the SEM micrographs under the trained laser powers. Further, it also accurately predicts the alumina’s microstructure under unexplored laser power. To generate the high-throughput material’s data, we fabricated an alumina sample array that contains hundreds of individual sample units, in one laser scan. Due to laser power distribution and the sample location, the individual units in this sample array have different but controllable microstructure. A microstructure-sensitive property, hardness, of the units in the large sample array was measured using micro-indentation. The microstructure of selected units was characterized. Using the results of microstructure and hardness we developed an ML algorithm to predict the expected microstructure of alumina of arbitrary hardness. An independent neural network was developed to measure the difference between the predicted microstructure and real ones. The neural network showed the predicted microstructure had less than 10% error for hardness from the experimental ones.