OpenSinter.com — Sinter with Confidence, Print to Scale.
Predict thermal phase transitions and eliminate sintering distortion using localized Graph Neural Networks for first-time-right metal prints—processed 100% locally on your workstation with zero cloud dependencies.
The Core Physics: Sintering heats 3D printed metal parts (composed of metal powder and organic binder) in a kiln. As the binder burns out and the metal particles fuse, the part undergoes up to 20% volumetric shrinkage. Simultaneously, gravity causes the soft, semi-solid part to sag, warp, or bend at high temperatures.
How the GNN Predicts This: Instead of slow, complex physics simulations, our Graph Neural Network (GNN) represents your STL mesh as a network of interconnected nodes (vertices) and edges. The model was trained on simulation and experimental data to predict exactly how each node will shift in 3D space during the heat cycle. By subtracting these predicted displacement vectors, we pre-deform your STL file so it warps into the target shape in the kiln.
The Training Dataset & GNN Architecture:
To model sintering distortion without slow, iterative multi-physics solvers, our GNN represents the STL geometry as a spatial graph. Nodes (Vertices) encode the 3D position ((x, y, z)), a local density factor, and a voxel category ID (Core, Skin, or Support). Edges (Bonds) are built within a local radius to capture structural load transfer paths. During 10 rounds of message-passing, nodes propagate thermal energy and gravitational shear stress (sigma_{xz}) to their topological neighbors, allowing the network to predict localized viscoplastic slumping and volumetric necking.
The GNN training database contains 1,200+ high-precision 3D scan pairs of green vs. sintered parts. These geometries span critical calibration cubes, tensile dogbones, thin vertical walls (capturing aspect-ratio buckling limits), overhang arches, gyroid lattices, and hollow containers (capturing internal gas expansion collapse). Physical prints are scanned, aligned using Iterative Closest Point (ICP) registration, and digested into 8.5 million nodal displacement vectors.
- 1,200+ 3D Geometries: Trained on a wide variety of shapes (cantilevers, organic structures, overhangs, and thick blocks).
- 8.5 Million Data Points: Voxel-level displacement vectors generated via Finite Element Analysis (FEA) and physical 3D scanner benchmarks.
- Filament Characterization: Calibrated with dilatometry tests on The Virtual Foundry Copper and Bronze Filamet™ across temperatures (700°C - 1500°C) and metal fractions (75% - 95%).
GNN Sintering Dataset Manager
Upload paired green and sintered scans to expand the Graph Neural Network training dataset.
Slicing / Print Parameters Expand to view/edit
Record Inspector
Training Pairs In Repository
| Part Name | Material Profile | Slicing Profile | Kiln Profile | Files | Actions |
|---|---|---|---|---|---|
| No training pairs uploaded yet. | |||||
GNN Sintering Dataset Analytics & Sintering Insights (Brain Dump)
1. Graph Net Representation of 3D Meshes
The GNN represents the 3D model as a graph (mathcal{G} = (mathcal{V}, mathcal{E})). Nodes (v_i in mathcal{V}) represent voxels, with features containing the coordinates, local material density (defined by extrusion multiplier), and node type (internal vs perimeter vs support interface).
Edges (e_{ij} in mathcal{E}) connect nodes within a spatial radius (typically 10mm). Edge attributes represent the relative distance vector and link strength. Through 10 rounds of localized message-passing, the network propagates thermal energy and stress gradients to compute the final displacement vectors ([dx, dy, dz]).
2. Viscoplastic Creep & Gravity Sag
During sintering hold cycles (e.g. 1250°C for Copper), the organic PLA/wax binder is completely burned away. The remaining metal powder matrix is highly porous and delicate. Under high temperature, the metal softens into a semi-solid state, causing the part to slump downwards along the Z-axis.
The GNN models gravity sag as a function of height, overhang angle, and support interfaces. Unsupported overhangs lack vertical rigidity, causing severe slumping, whereas thick bases provide stable compressive foundations.
3. Metallurgical Shrinkage Kinetics
- Bronze (Cu-Sn) vs. Copper (Cu): Tin-bronze composite filament exhibits a lower sintering onset (~750°C) and undergoes liquid-phase sintering as tin melts. This drives high volumetric shrinkage (~24.3%). Pure copper relies purely on solid-state grain boundary diffusion at higher temperatures (~900°C–1080°C), shrinking by ~18.5%.
- Infill Density and Binder Outgassing: Parts with 100% solid infill exhibit higher final strength but can trap organic decomposition gases in large volumes, causing cracking or bubbling. Slicing with 60%–80% Gyroid infill provides continuous interconnected channels that permit clean outgassing.
- Extrusion Multiplier (Flow Rate): A flow rate of 103%–105% is critical. It slightly over-extrudes the green state, closing microscopic voids between toolpaths, which reduces post-sintered porosity and stabilizes dimensions.
- Friction & elephant foot: Shrinkage causes the bottom layer of the model to slide inwards against the alumina setter plate. Substrate friction (friction coefficient (mu approx 0.45)) resists this sliding, leading to lateral flaring at the base (elephant foot).