Machine learning for parameter development
How artificial intelligence ends the trial and error of new 3D printing alloys.
1. Introduction: The parameter labyrinth
Anyone who wants to print a completely new material (e.g. an experimental high-entropy alloy) using Laser Powder Bed Fusion (PBF-LB) is faced with a mathematical nightmare. Laser power, scanning speed, hatch distance (track spacing), layer thickness and hundreds of other parameters influence each other. Conventionally, it often takes months and costs tens of thousands of euros to find the perfect parameter set for 99.9% density through trial and error.
Artificial intelligence (AI) and machine learning (ML) are revolutionizing this process. Algorithms find the perfect recipe for new metals in weeks instead of months.
What is Design of Experiments (DoE)?
DoE is the statistical planning of experiments. Instead of blindly guessing parameters, the software lets the machine specifically print a grid of dozens of small cubes. Each cube has slightly different parameters (e.g. cube A: 200W/800mm/s, cube B: 210W/850mm/s). After printing, the cubes are checked for porosity and the data is reported back to the software.
2. How machine learning accelerates the process
Classic DoE reaches its limits with 15+ variables. Machine learning takes over here.
- Bayesian optimization: An algorithm is trained with historical data (e.g. parameters from Titan or Inconel). He prints a small series of test specimens (DoE) from the new material. The algorithm “learns” the physical relationships based on the results (density, cracks). He then suggests completely new, unconventional parameter combinations that a human engineer would never have thought of.
- Digital twin: Using AI-supported thermal simulation (e.g. Physics-Informed Neural Networks), the melt pool is created virtually in the computer. The ML model tests millions of parameters purely virtually and identifies the optimal "process window" without wasting a single gram of powder.
3. Parameters for special requirements
There is no “one” perfect parameter set.
Sometimes you need maximum build speed (core parameters for the interior of the component, where density is important but surface area is irrelevant). Sometimes you need mirror-smooth surfaces (skin parameters for the outer shell). Machine learning helps manufacturers to generate so-called “parameter profiles” that change dynamically locally in the component (e.g. extra energy for support structures, less energy for delicate overhangs).
4. Correction of in-situ data
The next step in AI development is to link parameter development with in-situ monitoring (live cameras in the printer).
If the ML model detects live from the camera feedback (melt pool heat) when printing the test cubes that the temperature is getting too hot (risk of keyhole porosity), it stops printing this cube and automatically tests cooler parameters in the next layer. The printer essentially parameters itself while it is running.
5. Conclusion: democratization of new materials
Until now, only aviation giants could afford to develop custom alloys. Machine learning (software solutions such as Alchemite from Intellegens) reduces the development costs for PBF parameters massively. AI is the key to transferring thousands of new, ultra-efficient light metal alloys from the laboratory to industrial 3D printing.