A gradient descent method for inverse learning of a sintering model with specific physical parameter denormalization

Assessment of the physical sintering parameter from dilatometry curves can be an arch and time-consuming task. A hybrid method is proposed, coupling an inverse problem approach applied to the physics of sintering and the machine learning methodology to identify the parameters by convergence. The identification approach is based on the gradient descent optimization to identify simultaneously all the parameters, dataset learning/testing split to mitigate the experimental noise and the dataset normalization for an efficient gradient descent learning process. A challenging physical parameters denormalization has been developed for the sintering case. The resulting approach can be applied to experimental data like Master Sintering Curve (MSC) dilatometry, a very common approach in the literature. This inverse learning approach has similar architecture than artificial neurons. These smart physical neurons are very efficient as only two succeed in modeling complex sintering problems, a promising prospect for machine learning approaches.