Simulation software for laser DED with wire or powder
We simulate the whole bead in one run.
The software models melt pool and bead layer by layer, with your geometry and your material. Because the melt consists of freely moving particles, the mesh-free method needs no computational mesh, which makes the simulations efficient.
A computational mesh that follows the part shape has to be adapted continuously to the changing geometry. The particles, by contrast, move with the material and so save computing time.
A computational mesh on the left, free particles on the right. While four layers are deposited, only the mesh on the left has to be adapted to the growing part; the particles follow the material.
The SPH particle method
SPH (smoothed particle hydrodynamics) represents the material with moving computational particles. Each particle carries temperature and phase, solid or liquid, and moves with the melt. Heat transfer, melt flow and the shape of the free surface follow from how the particles interact.
Schematic, illustration
- Process input
- Laser and material feed set how much energy and mass enter the modeled region.
- Particle state
- Heat transfer and phase change update the state of each particle.
- Free surface
- Surface tension shapes the melt. At the contact line, wetting determines how far the melt spreads.
- Geometry
- Bead width, layer height and melt depth follow from the particle field.
Model quantities and limits
The model works at bead scale and simulates melt-pool flow and the geometry of the deposited bead. From these quantities the simulation computes the parameters along the toolpath that your CAM defines. Thermal history and distortion of the whole part belong to part-scale CAE.
Schematic, illustration
The model computes
- Bead width
- how wide the deposited bead lies on the surface
- Layer height
- how high the bead stands above the surface
- Melt depth
- how deep the melt reaches into the material below
- Peak surface temperature
- how hot the melt pool gets at its surface
- The model does not establish
- Residual stress, part distortion, microstructure and process windows outside the matched range.
The model takes the beam profile as an input: Gaussian, ring, top-hat.
Your team checks the adjusted machine file, runs the print and is responsible for the qualification of process and part.
Matching with experiments
The founder validated the model for pure titanium with single-track experiments in his doctoral thesis (2023).
In a pilot project we calibrate the model to your individual machine, for example with a cross-section. We adjust the model until simulated and experimental results agree as closely as possible.
Every comparison names the measurements we calibrated with.
From CPU cluster to GPU
Our GPU software is built on highly optimized CPU code. It spreads a simulation across many nodes of a compute cluster and scales almost linearly. We carried its data layout for fast memory access and its parallel structure over to GPUs; today we compute on GPUs only.
Measurement, compute cluster
Normalized core count r: the number of cores divided by the cores of the smallest run. r = 8 means eight times as many cores.
- Within one node
- almost ideal with few cores, then memory bandwidth slows it, because all cores of a node share the same memory and wait for data
- Across nodes
- each node brings its own memory, so almost linear again up to about five hundred times the cores of the smallest run, flatter beyond
Measured on a CPU cluster: AMD EPYC nodes with 128 cores, InfiniBand interconnect, November 2023
The simulation can be spread across many processor cores that compute at the same time. The speed-up states how much faster a run is than the smallest run. The normalized core count r states how many cores a run uses relative to the smallest run. If the curve lay on the diagonal, every doubling of the cores would halve the computing time. Across several nodes, meaning individual computers of the cluster, the curve runs almost parallel to this diagonal up to about r = 512. It flattens only beyond that. The speed-up therefore grows almost in proportion to the core count, but stays below the ideal value.
CPU code, optimized and scaled
Data layout and parallel structure carried over
GPU software, today
Sources
- Mesh and particlesSchematic, illustration: Synafloat method illustration.
Shows the principle, mesh-based versus mesh-free. Not a computed result.
- Three beam profilesSimulation, demonstrator: three single tracks, three beam profiles.
Supports the beam profile as an input of the model, not a comparison of the profiles.
- Speed-up over the normalized core countMeasurement, CPU cluster, November 2023: own measurement series.
Supports the relative speed-up over the normalized core count.
Supports the formulation of surface tension and wetting at free surfaces in SPH.
Matching of the model for pure titanium with single-track experiments.
Technical articles
The articles go deeper into bead geometry and the choice of simulation method.
Let’s talk about your part
In a call we clarify whether the simulation answers your parameter question.
Book a callYou pick the slot on book.synafloat.com.
You talk to Dr. Michael BlankThe founder has researched particle-based melt-pool simulation since 2017 and earned his doctorate in the field.
