The main goal of this thesis is to use a multiphase approach to study and predict the macrosegregation in steel ingots. Firstly, a two-phase columnar solidification model was employed to study the formation of channel segregation in a laboratory benchmark of Sn-10 wt.% Pb. For the first time 3D lamellar- and rod-structured channel segregates were numerically predicted. This result was later confirmed by H. Combeau and co-workers. A two-step mechanism of the channel formation was proposed, i.e. the initiation and the growth of the channel. The initiation of a channel is caused by flow perturbations near the solidification front in the mushy zone, which can be characterized by a mushy zone Rayleigh (Ra) number. When the maximum Ra number reaches a critical value (0.12 - 0.24), a channel may start to form. After the initial formation, a channel may either further develop (grow) into a stable channel, or disappear depending on the flow-solidification interaction in the two-phase mushy zone. In the second part of this thesis, a three-phase mixed columnar-equiaxed model was employed to study the macrosegregation in a 2.45 ton steel ingot. The general segregation pattern was predicted: cone shape negative segregation in the bottom region, positive segregation in the top region, and some A-segregation bands near the wall. The equiaxed sedimentation was found playing an important role in the formation of global segregation in such big ingot. The formation mechanism of A-segregation was analyzed: flow instability causes the formation of quasi A-segregates, but both the appearance of equiaxed crystals and their interaction with the growing columnar dendrites strengthen the segregated severity significantly. The equiaxed phase was found not a necessary condition for the formation of quasi A-segregates. The predicted macrostructure and macrosegregation results agreed qualitatively with the experimentally reported segregation pattern, although the quantitative discrepancy between the calculations and the experimental results was still significant. The reason for the above discrepancy is due to the ignorance of the dendritic morphology by the original three-phase model. Therefore, in this thesis a further model-development step was made by implementing a simplified dendritic morphology. This consideration has been verified to improve the quantitative accuracy of the numerical prediction significantly. Additionally, a 25-ton steel ingot was also calculated. The global segregation pattern was well predicted, and a reasonable agreement with the experimental result was obtained. As a relative coarse grid was used in this calculation, A-segregation bands cannot be predicted with sufficient resolution. Finally, a four-phase solidification model was established to combine the formation of macrosegregation and shrinkage cavity. An additional phase, i.e. gas phase, was considered. It has predicted the shrinkage cavity in the hot top region, and the predicted shape of the cavity is quite similar to the experiment one. However, the predicted macrosegregation below the shrinkage cavity did not agree with the experimental one. Therefore, further effort is still required to improve the model capability.