Mechanistic machine learning for metamaterial fatigue strength design from first principles in additive manufacturing

Abstract

Digital control in manufacturing processes produces significant amounts of metadata. The production process metadata, such as thermal and optical measurements, enables a higher degree of property grading than uninstrumented manufacturing and feedback for fault detection. This study explores how metadata can design fatigue-resistant structures using physically grounded models such as density functional theory, cyclic plasticity, and fracture mechanics that train machine learning algorithms. Machine learning models work very efficiently in their trained physical space. In comparison, mechanistic models are computationally costly for complex phenomena such as fatigue. We show how fatigue can be administered consistently at all scales by energy-based criteria and how a mechanistic function can be built based on this concept. The energy mechanistic function allows exact quantification of the effect of the existing flaws from manufacturing on fatigue lifetime under certain load boundary conditions. Since the mechanistic function is local and subscale to the prediction scale of the machine learning model, it can be used to build density functions for probabilistic regression of the fatigue property on the scale above. The analysis is applied to the selective laser melting process due to the availability of digital control and metadata generation during deposition.

Keywords

• Density functional theory (DFT)
• Functional grading
• Machine learning (ML)
• Fatigue strength
• Cohesion energy
• Additive manufacturing (AM)