Accurate prediction of the liquidus temperature (T-L) of alloys remains a challenge despite numerous theoretical models. Here, we explore and analyze the degree to which machine learning, ML, strategies can be used to predict T-L. We use established literature data on liquidus temperatures of 85,523 binary alloys to train ML models using various feature vectors to represent the alloys. While our results are comparable to previous studies, the persistent similar to 8% error underscores the limitations of current ML models for practical usage. The suboptimal accuracy leads us to question how well-defined the problem is and to what degree fundamental limitations prevent us from attaining more accurate predictions. We identify two major challenges in predicting the liquidus temperature of binary alloys through supervised ML algorithms. One challenge is representing the relevant characteristics of an alloy that determines liquidus temperature through appropriate features. The other fundamental challenge is the discreteness of atomic properties. The difference between two elements and thereby alloy systems is significant, which makes it difficult to learn from one alloy system to predict properties of another. We argue that these problems can be reduced to some extent, however these challenges are common in complex materials science problems and constitute a fundamental challenge in applying supervised ML strategies in this context.
