AMT-FLOW-3D

This dataset consists of LPBF (Laser Powder Bed Fusion) simulations performed using FLOW-3D.

Simulation Plan

The goal is to create a comprehensive dataset covering a wide range of process parameters:

• Laser Power: 50 Watts to 400 Watts (8 steps)
• Scanning Speed: 0.3 m/s to 2.4 m/s (8 steps)
• Laser Tilt Angle: Subset of {0°, 5°, 10°, 15°, 20°, 25°} (Planned for 4 angles)
• Total Combinations: 256 (8 Power × 8 Speed × 4 Angle)

ThermoPore Paper Parameters

Additional combinations from the ThermoPore paper are included:

• Laser Power: 103 Watts (constant)
• Spacing Sample:
  • Velocity: 1.40 m/s (constant)
• Velocity Sample:
  • Velocities: 1.05, 1.12, 1.19, 1.26, 1.33, 1.47, 1.54, 1.61, 1.68, 1.75 m/s

Existing Simulations

The source/ directory contains initial simulations at different laser tilt angles (all at 350W and 0.3 m/s):

• P350W_V0.3mps_A0deg
• P350W_V0.3mps_A5deg
• P350W_V0.3mps_A10deg
• P350W_V0.3mps_A15deg
• P350W_V0.3mps_A20deg
• P350W_V0.3mps_A25deg

Technical Details

• Material: Stainless Steel 316L (SS316L)
  • Solidus Temperature: 1674.15 K
  • Liquidus Temperature: 1697.15 K
• Mesh Configuration:
  • Cell size: 20 µm (uniform)
  • Mesh dimensions: 215 x 50 x 50 (approx. 537,500 real cells)
• Physics Models:
  • Heat source: Welding model (Gaussian/Keyhole)
  • Phase change (Solidification/Melting)
  • Surface tension and Evaporation pressure

Data Structure

Each simulation directory contains:

• flsinp.*: FLOW-3D input configuration.
• runhyd.txt: Solver execution log.
• report.*: Detailed simulation report including mesh and mass/energy audits.
• flslnk_npz/: Post-processed data in .npz format.
  • Arrays include: temperature, pressure, fraction_of_fluid, liquid_label, velocity.
  • Stored as (Z, Y, X) 3D grids.
• visualize.py: Script for generating 2D cross-sections and 3D isometric visualizations.

Findings

• Initial simulations confirm stable keyhole formation at 350W and 0.3 m/s.
• Increasing the tilt angle affects the keyhole stability and melt pool morphology.
• Post-processing converts large FLOW-3D results (approx. 144 GB uncompressed per sim) into manageable NumPy arrays for ML training.