Induration in steel industries is the process of pelletizing iron ore particles. It is an important unit operation which produces raw materials for a subsequent chemical reduction in Blast Furnace. Of the enormous amount of energy consumed by Blast Furnace, a large portion is utilized in processing the raw materials. High-quality raw materials, therefore, ensure less consumption of energy in the Blast Furnace. Thus, optimization of induration process is necessary for conservation of a significant amount of energy in steelmaking industries. To realize this, a highly non-linear, industrially validated, 22 dimensional first principles based model for induration is created and a multi-objective optimization problem is designed. However, the physics-based model being computationally expensive, Multi-layered Perceptron Networks (MLPs) are trained to emulate the induration process. Novelty in this work lies with the optimal architecture design of MLPs through a multi-objective integer non-linear programming (MO-INLP) problem and with simultaneous training size estimation through four different Sobol sampling-based algorithms. Successful emulation of induration process resulted in 10-fold speed increment in optimization through surrogate models. To justify the parsimonious behavior of resultant MLPs, five different tests are performed for checking whether they are over-fitted. Comparison with Kriging adds to other highlights. © 2019, © 2019 Taylor & Francis.