Why does machine learning work really well in many engineering problems?

Abstract Recent studies continue to document successful implementations of machine learning (ML) across engineering disciplines (even when used with default configurations). This observation challenges the notion that rigorous parameter optimization and exhaustive model tuning are, in fact, stable prerequisites for ML models. This phenomenon raises intriguing questions about why ML methods appear to work well in engineering contexts. This paper explores this phenomenon and identifies six notions that support and justify the aforementioned observation. Our findings indicate that engineering data is often well-structured and governed by consistent physical laws, which makes it naturally suitable for ML. Further, the maturity of ML algorithmic libraries/packages, which include already optimized defaults and best practices, enables quick convergence to good solutions when applied to engineering problems. In addition, domain expertise in engineering indirectly enhances ML performance through careful data collection and feature engineering. Finally, we also identify potential research gaps and challenges that ought to be addressed to better understand and tune ML systems as they become more deeply integrated into engineering and decision-making processes.