Sharp　phase　interfaces　and　accurate　temperature　distributions　are　important　criteria　in　the　simulation　of　solid-liquid　phase　changes.　The　multi-relaxation-time　lattice　Boltzmann　method　(MRT-LBM)　shows　great　numerical　performance　during　simulation;　however,　the　value　method　of　the　relaxation　parameters　needs　to　be　specified.　Therefore,　in　this　study,　a　random　forest　(RF)　model　is　used　to　discriminate　the　importance　of　different　relaxation　parameters　to　the　convergence,　and　a　support　vector　machine　(SVM)　is　used　to　explore　the　decision　boundary　of　the　convergent　samples　in　each　dimensional　model.　The　results　show　that　the　convergence　of　the　samples　is　consistent　with　the　sign　of　the　decision　number,　and　two　types　of　the　numerical　deviations　appear,　i.e.,　the　phase　mushy　zone　and　the　non-physical　heat　transfer.　The　relaxation　parameters　chosen　on　the　decision　boundary　can　further　suppress　the　numerical　bias　and　improve　numerical　accuracy.