The new system predicts microstructure formation in real-time, slashing simulation time and boosting print precision.
According to Arizona State University (ASU), researchers are utilizing artificial intelligence to significantly enhance metal 3D printing, making it faster, more accurate, and less wasteful. Professors Aviral Shrivastava and Ashif Iquebal are leading the charge through their NSF-funded project, ‘CompAM: Enabling Computational Additive Manufacturing’.
The team’s challenge was to 3D print a five-axis 316L stainless steel naval propeller – a part with demanding geometry and performance requirements – while precisely controlling the metal’s microstructure. Their target was to achieve grain sizes under one micron – smaller than a spider’s silk strand – which dramatically enhances material properties.
“When we do metal printing, the quality of metal is actually dependent on the cooling curve,” said Shrivastava. Traditional methods require either months-long simulations on supercomputers or expensive trial-and-error. The team aims to cut that down drastically by developing a physics-informed, AI-powered system that learns how metal forms in real time during printing. Rather than brute-force simulations, their model intelligently identifies and focuses only on critical zones – skipping parts that stay stable. This not only slashes simulation time, but also boosts accuracy.
“Physics is just a set of rules that are obeyed in the real world,” said Shrivastava. By combining those rules with data-driven learning, the AI adapts without relying on massive datasets.
“The real value of this work is its ability to bridge research and industrial need,” said Iquebal. In sectors like aerospace or defense, where material performance is non-negotiable, the ability to predict and fine-tune material properties in advance is a game-changer.
Using ASU’s state-of-the-art 3D printer – equipped with lasers and a six-axis robotic arm – the team will compare predicted versus actual microstructures in a printed propeller. Their results will be benchmarked against traditional methods, and the tools they develop will be made open-source.
Read the rest of this story at VoxelMatters
