Evaluation of machine learning based models to predict the bulk density in the flash sintering process

Abstract Flash Sintering (FS) has still unexplained issues of the flash phenomenon that do not allow the construction of dynamic models to predict the ceramic microstructure. This work investigates computational models based on supervised machine learning techniques for modelling the FS process to predict the bulk density (BD) using the holding time (TH) and electric current density (J). Samples of 3YSZ (zirconia stabilised with 3 mol% of yttria) were prepared and shaped in cylindrical geometry (6 mm in diameter and 4 mm in height). The FS parameters were adjusted at a heating rate of 20 °C/min, maximum current density of 100 mA/mm² and an electric field in alternating current mode of 90 V/cm (1 kHz). FS processes were performed for different J values set to 80, 100 and 200 mA/mm² and different TH values set to 60 s, 120 s and 180 s. After preliminary regression analysis, four learning methods were applied to generate the models: Artificial Neural Network, K-Nearest Neighbours, Random Forest, and Support Vector Machine. The results show moderate correlations (>0.5) between the values measured and predicted by the different models in the prediction of BD through TH. The Support Vector Machine based method allowed the best BD prediction models to be built, with r equal to 0.62, MAE equal to 0.95 and RMSE equal to 1.37. These results indicate the potential of applying the methods and allow the definition of new hypotheses to improve the models.