Engineering Data Analysis Beyond Statistics vs. Machine Learning: A Review of Explanation, Prediction, Hybrid Modeling, and Responsible Decision Support

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

This structured narrative review explores how engineers select and evaluate traditional statistical, machine learning, mechanistic, and hybrid methods for explanation, prediction, and responsible decision support. We identified representative literature available through 2025 via iterative searches of Web of Science, Scopus, ScienceDirect, and Google Scholar, and conducted a thematic synthesis. We qualitatively assessed the credibility and relevance of the literature sources. Application areas covered include manufacturing, semiconductor processing, mechanical and materials systems, civil infrastructure, chemical processes, energy and construction, robotics, and greenhouse engineering (excluding healthcare). Comparisons included analytical objectives, data characteristics, hypotheses, interpretability, uncertainty, causality, physical consistency, validation, and deployment. The synthesis yielded three recurring conclusions. First, a method’s applicability depends on information content, dependency structure, operating conditions, and decision consequences, not dataset size or method labels. Second, statistical significance and predictive accuracy do not independently demonstrate causal validity, physical plausibility, or operational effectiveness. Reliable evaluation combines internal validation, feasible external testing, repeatable benchmarking against meaningful benchmarks, and deployment evaluation to address issues such as data leakage, distribution shift, and uneven error costs. Third, hybrid models are appropriate when complementary functionality is required, but the increased computational, validation, and maintenance burden must be justified; simpler, validated models may suffice. Physics-informed machine learning, gray-box models, and digital twins differ in their functional and integration requirements. This paper’s main contribution is a cross-domain decision-making architecture and practical selection guidelines that link engineering goals to modeling choices, validation requirements, decision thresholds, and operational measures. This framework does not rank algorithms or prescribe general workflows; instead, it supports domain-specific judgments about model suitability, uncertainty, security, cost, and human oversight.