Safe Overfitting of Boosted Tree Algorithm in Heat Transfer Modeling

Overfitting generally considered to be an avoidable pitfall in any kind of machine learning setting. Both academic literature and applied machine learning is saturated with tools and techniques preventing overfit and estimating the likelihood of its occurrence, and the potential damages it causes during model generalization. However in this paper it will be outlined, that if certain strict conditions are met, overfitting may still occur, but the generated machine learning model is still effective and applicable, as long as it is based on an actual physical phenomena behind the data, and there is an abundance of learnable events provided. In this paper, a heat transfer example is shown as the empirical proof of this concept.