PiGRAND: New AI Framework Uses Graph Neural Networks to Model 3D Printing Heat Transfer

Researchers at Leipzig University have published PiGRAND, a new machine learning framework that combines physics-informed modeling with graph neural networks to better predict heat transfer in 3D printing processes.

The Heat Transfer Challenge

Heat transport is critical in additive manufacturing. The quality of 3D printed parts depends heavily on how heat flows through the material during printing. Traditional finite element methods are computationally expensive, and purely data-driven approaches often lack physical consistency.

PiGRAND Combines Physics with Graph Neural Networks

PiGRAND addresses these limitations by embedding physical principles derived from partial differential equations directly into the learning model. The framework uses graph neural diffusion to model continuous heat transport across the geometry being printed.

The approach is inspired by both explicit Euler and implicit Crank-Nicolson numerical methods for modeling heat equations. By combining these with sub-learning models, PiGRAND secures accurate diffusion across graph nodes representing different points in the print geometry.

Key Innovations

• Efficient graph construction: A new procedure reduces computational complexity for graph learning
• Physics-constrained: Incorporates actual physical laws rather than learning purely from data
• Transfer learning: Combined with efficient transfer learning for improved computational performance
• Open source: Code available on GitHub

Results

Evaluated on thermal images from actual 3D printing, PiGRAND showed significant improvements in prediction accuracy and computational performance compared to traditional graph neural diffusion (GRAND) and physics-informed neural networks (PINNs).

This research represents a significant step toward intelligent, real-time process control in additive manufacturing where quality predictions could be made during printing rather than after.