Research on Density Prediction of Laser Powder Bed Fusion Process Parameters for IN718 Nickel-Based Superalloy Based...

Area of Science:

• Materials Science and Engineering
• Additive Manufacturing
• Computational Materials Science

Background:

• Modeling the non-linear relationship between process parameters and relative density in selective laser melting (SLM) is challenging, especially with limited experimental data.
• IN718 nickel-based superalloy is a critical material for high-performance applications, necessitating precise control over its SLM process.
• Small-sample conditions hinder the development of robust predictive models for SLM.

Purpose of the Study:

• To develop and validate a data-driven prediction framework for modeling the relative density of IN718 superalloy produced by SLM under small-sample conditions.
• To integrate data augmentation, physics-informed feature engineering, machine learning, and interpretability analysis for robust SLM process optimization.
• To overcome the limitations of small datasets in achieving accurate predictions of SLM part quality.

Main Methods:

• Collected 14 experimental datasets varying laser power, scan speed, and hatch spacing for IN718 SLM in vertical and horizontal directions.
• Employed data augmentation strategies (RBF, GAN, KNN) under physical constraints (local perturbation, volumetric energy density filtering) to address small-sample limitations.
• Engineered eight physics-informed features (including volumetric and line energy density) and trained/optimized SVR, RF, and ANN models using exhaustive grid search and LOO-CV.

Main Results:

• The artificial neural network (ANN) model achieved the highest average R² of 0.9269, significantly outperforming a physics-based baseline model (R² = 0.2534).
• Volumetric energy density (E_vol) was the most influential feature, contributing 51.58% to the predictive power.
• Physics-derived features improved model accuracy (average R² gain of 0.0246) compared to using raw process parameters alone.

Conclusions:

• The developed data-driven framework provides a reliable and interpretable solution for intelligent SLM process optimization with limited experimental data.
• Machine learning models, particularly ANN, demonstrate superior performance over empirical formulas for predicting relative density in SLM.
• Incorporating physics-informed features enhances model accuracy and provides physical insights into the SLM process.