A physically-supported data-driven proxy modeling based on machine learning classification methods: Application to water front movement prediction