Session: 03-04-02: Artificial Intelligent Applications in Manufacturing II

Paper Number: 172521

Towards Real-Time Stochastic Melt Pool Modeling via Scientific Deep Learning in Laser Powder Bed Fusion 

The introduction of in-situ monitoring systems in the Laser Powder Bed Fusion (L-PBF) process has played a crucial role in capturing in-process stochasticity and identifying defects such as porosity, keyholing, and balling during metal additive manufacturing. These monitoring tools—typically infrared cameras, photodiodes, and optical sensors—allow for real-time feedback by recording surface temperatures, melt pool morphology, and spatter events. However, the data they collect is inherently limited in both spatial and temporal resolution, confined to surface-level information and specific detection points. Additionally, these measurements are often affected by significant noise and physical disturbances such as vapor plume, powder dynamics, and surface reflectivity changes, which obscure the true physical state of the melt pool. In contrast, physics-based simulations, particularly those using high-fidelity thermal-fluid models, can provide detailed insights into melt pool behavior under controlled conditions. These models can capture temperature gradients, fluid flow, and phase transformations with high accuracy. However, their computational expense and inability to account for stochastic disturbances limit their practical utility for online prediction or real-time control during fabrication. Moreover, these deterministic models struggle to reflect the real-world variability intrinsic to the L-PBF process, making them insufficient for robust decision-making in dynamic manufacturing environments. To overcome these limitations, we propose a novel Scientific Deep Learning (Sci-DL) framework that integrates multi-modal in-situ monitoring data with high-fidelity simulation results to enable real-time, stochastic modeling of melt pool evolution. Our approach employs a physics-informed encoder-decoder architecture that fuses noisy, sparse sensor measurements with deterministic simulation outputs within a shared latent embedding space. This embedding allows the model to learn stochastic mappings that represent the variability of real-world manufacturing conditions while preserving the underlying physical relationships governed by simulation. The encoder-decoder structure is built using convolutional neural networks, with the encoder designed to extract features from raw in-situ signals and the decoder trained to reconstruct full-field 3D melt pool geometries. A calibrated thermal-fluid simulation model is used to generate training data under a range of process parameters, such as laser power, scan speed, and hatch spacing. To simulate realistic variability, we define a stochastic parameter space that introduces controlled perturbations representing common sources of uncertainty in the L-PBF process. Preliminary results demonstrate that the Sci-DL model can accurately recover spatial melt pool characteristics and temporal evolution across a variety of processing conditions, using only partial in-situ observations as input. The model achieves near real-time inference speeds, making it well-suited for integration into closed-loop control systems. This work contributes a promising AI-augmented framework for advancing real-time monitoring, defect prediction, and adaptive control in metal additive manufacturing, representing a step toward intelligent, self-correcting L-PBF systems.

Presenting Author: Lin Cheng University of Maryland

Presenting Author Biography: Lin Cheng is currently an assistant professor in the Department of Mechanical Engineering at University of Maryland, College Park. Dr. Cheng received his B.S. degree from Xi’an Jiao Tong University and M.S. degree from Shanghai Jiao Tong University. He holds a Ph.D. in mechanical engineering from the University of Pittsburgh and worked as a postdoctoral researcher at Northwestern University from 2019 to 2021. His research interests lie in scientific artificial intelligence and computational design for metal additive manufacturing. The design optimization methods developed by him have been adopted and implemented by ANSYS in their engineering simulation software.

Authors:

Lin Cheng University of Maryland
Yunhao Zhang Worcester Polytechnic Institute
Yao Fu Viriginia Tech