Inexpensive High Fidelity Melt Pool Models in Additive Manufacturing Using Generative Deep Diffusion

Francis Ogoke, Quanliang Liu, Olabode Ajenifujah, Alexander Myers, Guadalupe Quirarte, Jack L. Beuth, Jonathan Malen, Amir Barati Farimani

TL;DR

This paper introduces a generative deep learning framework that uses conditional diffusion models to upscale low-fidelity, coarse-mesh simulations of additive manufacturing melt pools into high-fidelity equivalents. By bypassing the need for computationally expensive fine-mesh multi-physics simulations, the approach accurately predicts critical features like melt pool depth and keyhole vapor cavity oscillations while reducing analysis time by two orders of magnitude.

What Problem Does This Paper Address?

Laser Powder Bed Fusion (L-PBF) is a prominent additive manufacturing technique capable of printing complex metallic parts. However, its widespread adoption is hindered by random defects like porosity, residual stress, and surface roughness. These defects are deeply tied to the mesoscale dynamics of the melt pool—the localized region of molten alloy created by the laser.

When the laser delivers excess energy, the substrate vaporizes, creating a deep, unstable "keyhole" cavity. If this cavity collapses, it traps gas bubbles that solidify into pores. To prevent this, engineers rely on multi-physics Computational Fluid Dynamics (CFD) simulations to optimize printing parameters. Unfortunately, capturing these complex, rapidly fluctuating dynamics requires extremely fine mesh sizes, making high-fidelity simulations computationally prohibitive (often taking hours per single track run). Coarse simulations run quickly but fail to resolve the critical small-scale physical features responsible for defect formation.

Methods

The authors bridge this gap by framing the problem as a spatial super-resolution task, mapping 2D cross-sections of low-fidelity (LF) coarse simulations to high-fidelity (HF) counterparts. They propose a novel two-stage generative pipeline:

  1. CNN Encoder: A Residual in Residual Dense Network (RRDN), adapted from the ESRGAN architecture, processes the coarse LF input. It is trained using an $L_1$ loss to approximate the large-scale morphology and mean temperature field of the high-fidelity target.
  2. Conditional Diffusion Model: Standard deterministic CNNs tend to blur out high-frequency details. To capture the stochastic nature of the melt pool, the authors feed the CNN encoder's features as a conditioning input into a Denoising Diffusion Probabilistic Model (DDPM). This U-Net-based diffusion model refines a random Gaussian noise vector to reconstruct realistic, high-frequency details and sharp boundaries.
Figure 1: a) The goal of this model is to create a deep-learning method for upscaling simulations of laser powder bed fusion, reducing the computational time required to obtain high fidelity data given a dataset of coarsely resolved simulations. b) To do so, the physical fields of the low fidelity simulation first are passed through a ResNet based upscaling model, followed by a further refining process carried out by a conditional diffusion model.

To make the model practical for rapid engineering workflows, they replace the standard Markovian sampler with an accelerated non-Markovian deterministic sampler. This reduces the required denoising steps from 1000 down to just 20, yielding a 50× speedup during inference.

Datasets

The dataset was generated using the commercial CFD software FLOW-3D to simulate single-track laser runs on bare metal plates:

  • SS316L (Stainless Steel): 295 simulations evaluated on a 2× upscaling task (mapping a 20 $\mu m$ coarse mesh to a 10 $\mu m$ fine mesh).
  • Ti-6Al-4V (Titanium Alloy): 40 simulations evaluated on a harsher 4× upscaling task (mapping a 20 $\mu m$ coarse mesh to a 5 $\mu m$ fine mesh), specifically targeting complex, non-deterministic keyhole behavior.

The data was split into 75% for training, 10% for validation, and 15% for testing on unseen process conditions.

Key Insights and Ideas

  • Deterministic vs. Generative Upscaling: Traditional upscaling methods like bicubic interpolation or standard CNNs fail because they average out the solution space. The diffusion model treats super-resolution probabilistically, allowing it to sample from a distribution of physically plausible high-fidelity states.
  • Capturing Intrinsic Physical Variability: High energy density printing causes the keyhole cavity to oscillate violently. Because the diffusion model introduces stochasticity during the denoising process, it naturally recreates these rapid temporal oscillations and keyhole wall protrusions without needing explicit, complex PDE-derived physics loss functions.

Key Results

  • Massive Speedup: A full-field fine-mesh CFD simulation takes roughly 3 hours per run. In contrast, the diffusion upscaling framework performs inference in just 0.41 seconds per sample, cutting down analysis time by over two orders of magnitude.
  • High Quantitative Accuracy: For the challenging 4× upscaling task on Titanium, the model predicted the final melt pool depth within an average error of 3 $\mu m$.
  • Oscillation Tracking: The diffusion model successfully captured the amplitude of keyhole depth fluctuations over time, achieving a Pearson $R^2$ coefficient of 0.74, vastly outperforming the standalone CNN model which scored a meager $0.06$.
Visual comparison of temperature fields across different upscaling methods against the high-fidelity target.

Limitations and Future Work

While the results are a major step forward for accelerating additive manufacturing design loops, a notable limitation is that the model currently operates on 2D cross-sections taken along the laser's plane of travel rather than full 3D volumes.

The authors note that future work will focus on extending this conditional diffusion framework to three-dimensional training data. This would allow the model to integrate directly into large-scale, coarse 3D component simulations, providing localized, high-fidelity defect insights that would otherwise be computationally impossible to resolve.

Summary updated

Abstract

Defects in laser powder bed fusion (L-PBF) parts often result from the meso-scale dynamics of the molten alloy near the laser, known as the melt pool. For instance, the melt pool can directly contribute to the formation of undesirable porosity, residual stress, and surface roughness in the final part. Experimental in-situ monitoring of the three-dimensional melt pool physical fields is challenging, due to the short length and time scales involved in the process. Multi-physics simulation methods can describe the three-dimensional dynamics of the melt pool, but are computationally expensive at the mesh refinement required for accurate predictions of complex effects, such as the formation of keyhole porosity. Therefore, in this work, we develop a generative deep learning model based on the probabilistic diffusion framework to map low-fidelity, coarse-grained simulation information to the high-fidelity counterpart. By doing so, we bypass the computational expense of conducting multiple high-fidelity simulations for analysis by instead upscaling lightweight coarse mesh simulations. Specifically, we implement a 2-D diffusion model to spatially upscale cross-sections of the coarsely simulated melt pool to their high-fidelity equivalent. We demonstrate the preservation of key metrics of the melting process between the ground truth simulation data and the diffusion model output, such as the temperature field, the melt pool dimensions and the variability of the keyhole vapor cavity. Specifically, we predict the melt pool depth within 3 $\mu m$ based on low-fidelity input data 4$\times$ coarser than the high-fidelity simulations, reducing analysis time by two orders of magnitude.
BibTeX
@misc{ogoke2023inexpensive,
  author = {Francis Ogoke and Quanliang Liu and Olabode Ajenifujah and Alexander Myers and Guadalupe Quirarte and Jack L. Beuth and Jonathan Malen and Amir Barati Farimani},
  title = {{Inexpensive High Fidelity Melt Pool Models in Additive Manufacturing Using Generative Deep Diffusion}},
  year = {2023},
  eprint = {2311.16168},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2311.16168},
}
Figure