Impact of Artificial Intelligence and Machine Learning on Applications and Simulation Tools of Mechanical Engineering: A Comprehensive Review
TL;DR
This review synthesises the state of the art across five interconnected domains: structural simulation and finite element analysis, computational fluid dynamics, design automation and topology optimisation, manufacturing process simulation with particular attention to additive manufacturing, and prognostics together with robotic control.
Abstract
The integration of artificial intelligence and machine learning into mechanical engineering has moved from a peripheral research interest to a central force reshaping design, simulation, manufacturing, and maintenance practice. This review synthesises the state of the art across five interconnected domains: structural simulation and finite element analysis, computational fluid dynamics, design automation and topology optimisation, manufacturing process simulation with particular attention to additive manufacturing, and prognostics together with robotic control. Surrogate and physics-informed learning architectures reduce the computational burden of high-fidelity simulation by orders of magnitude, but they also introduce new questions around generalisability, interpretability, and validation against first-principles physics. Deep generative models are reframing topology optimisation and conceptual design as a learned mapping rather than an iterative search, and hybrid digital twins increasingly couple physics-based solvers with data-driven correction terms to support real-time decision making on the shop floor. Prognostics and health management has benefited substantially from deep sequence models for remaining useful life estimation, while reinforcement learning is maturing as a control paradigm for robotic manipulation and mechatronic systems. Persistent challenges include data scarcity in high-value manufacturing contexts, the limited physical interpretability of black-box predictors, inconsistent verification and validation protocols, and the computational and environmental cost of training large models. The review closes by identifying future research directions centred on physics-constrained architectures, federated and transfer learning across heterogeneous industrial datasets, and standardised benchmarking, before outlining the principal limitations of the present narrative synthesis.