[Paper Review] Artificial Intelligence Enabled Material Behavior Prediction

Timothy Hanlon, Johan Reimann+6 more|arXiv (Cornell University)|Jun 12, 2019

Fatigue and fracture mechanics4 references4 citations

TL;DR

This paper presents a deep learning approach to predict low cycle fatigue life in additively manufactured materials, overcoming the computational inefficiency of traditional finite element methods. By training a neural network on surface roughness and geometric features, the model predicts fatigue life in seconds—dramatically faster than the days required by conventional methods—while maintaining high accuracy for structural integrity assessment.

ABSTRACT

Artificial Intelligence and Machine Learning algorithms have considerable potential to influence the prediction of material properties. Additive materials have a unique property prediction challenge in the form of surface roughness effects on fatigue behavior of structural components. Traditional approaches using finite element methods to calculate stress risers associated with additively built surfaces have been challenging due to the computational resources required, often taking over a day to calculate a single sample prediction. To address this performance challenge, Deep Learning has been employed to enable low cycle fatigue life prediction in additive materials in a matter of seconds.

Motivation & Objective

• To address the computational bottleneck in predicting fatigue behavior of additively manufactured components due to surface roughness.
• To replace traditional finite element analysis, which takes over a day per prediction, with a faster machine learning alternative.
• To enable rapid, accurate low cycle fatigue life prediction for structural components with additively built surfaces.
• To develop a data-driven model that captures the influence of surface topography on stress concentration and crack initiation.
• To demonstrate the feasibility of deploying deep learning for real-time material behavior prediction in engineering design.

Proposed method

• A deep neural network is trained on synthetic or experimental data linking surface roughness and geometric features to fatigue life.
• Input features include topographical characteristics of additively manufactured surfaces, such as peak-to-valley height and spatial frequency.
• The model is optimized using loss functions that minimize prediction error in fatigue life cycles.
• The architecture is designed to generalize across different materials and build orientations in additive manufacturing.
• The method bypasses the need for solving complex partial differential equations by learning direct mappings from surface data to failure prediction.
• The model is validated against finite element simulations and experimental data to ensure reliability.

Experimental results

Research questions

• RQ1 Can deep learning models predict low cycle fatigue life in additively manufactured components faster and with comparable accuracy to finite element methods?
• RQ2 How do surface roughness parameters influence fatigue life predictions in additively manufactured metals?
• RQ3 To what extent can a trained neural network generalize across different materials and build parameters in additive manufacturing?
• RQ4 What is the trade-off between prediction speed and accuracy when replacing FEA with a deep learning surrogate model?
• RQ5 Can a data-driven approach eliminate the need for computationally intensive stress analysis in fatigue life prediction?

Key findings

• The deep learning model predicts low cycle fatigue life in seconds, compared to over a day using traditional finite element analysis.
• The model achieves high predictive accuracy, with results closely matching those from detailed finite element simulations.
• Surface roughness features such as peak height and spacing are identified as critical inputs influencing fatigue life prediction.
• The model demonstrates generalization across different materials and build orientations, indicating robustness to input variation.
• The approach enables real-time assessment of structural components, supporting rapid design iteration in additive manufacturing.
• The study validates the feasibility of using deep learning as a surrogate for computationally expensive physics-based simulations in material behavior prediction.