Source-linked AI summary
Thermal Control of Laser Powder Bed Fusion Using Deep Reinforcement Learning
Francis Ogoke, Amir Barati Farimani
TL;DR
Defects associated with melt-pool behavior and overheating challenge Laser Powder Bed Fusion. This paper uses a DRL framework and continuum temperature simulations to learn velocity-based control policies; in simulation, the policies reduce melt-pool variability and overheating across different scan paths.
Problem
Melt-pool behavior and overheating are linked to defects, motivating control that can adjust process parameters as temperature distribution changes during a scan.
Source: §1 · p. 1Method
A Proximal Policy Optimization-enabled DRL framework learns laser-velocity control policies for melt-pool depth using efficient continuum heat-conduction simulations.
Source: §1 · p. 2, §2.1 · p. 3, §2.4 · p. 6Results
In simulation, learned policies reduce melt-pool variability and overheating across different scan paths.
Source: §4 · p. 9Takeaways & Limitations
The simulations demonstrate effective velocity-based control policies for reducing overheating and melt-pool depth variation across the tested scan paths.
Source: §4 · p. 9Velocity bounds may be imposed to avoid defects not examined in this work, such as spatter caused by high laser-motion jerk.
Source: §3 · p. 9
Abstract
from arXiv · showhide
Powder-based additive manufacturing techniques provide tools to construct intricate structures that are difficult to manufacture using conventional methods. In Laser Powder Bed Fusion, components are built by selectively melting specific areas of the powder bed, to form the two-dimensional cross-section of the specific part. However, the high occurrence of defects impacts the adoption of this method for precision applications. Therefore, a control policy for dynamically altering process parameters to avoid phenomena that lead to defect occurrences is necessary. A Deep Reinforcement Learning (DRL) framework that derives a versatile control strategy for minimizing the likelihood of these defects is presented. The generated control policy alters the velocity of the laser during the melting process to ensure the consistency of the melt pool and reduce overheating in the generated product. The control policy is trained and validated on efficient simulations of the continuum temperature distribution of the powder bed layer under various laser trajectories.
1 Introduction
Laser Powder Bed Fusion can produce complex parts, but defects associated with melt-pool behavior and residual stresses limit adoption in precision applications. This work presents a Deep Reinforcement Learning framework to control melt-pool variation, using a continuum heat-conduction simulation to support computationally feasible training.
1 Introduction
- Laser Powder Bed Fusion produces complex parts, but defects and residual stresses can impair physical properties and limit adoption in precision applications.Reported defects include poor surface finish, porosity, delamination, and cracking.Source: §1 · p. 1
- Melt-pool characteristics and temperature gradients contribute to defects, while the continuum-scale model neglects convection and radiation to enable data-intensive DRL.The cited defects include keyhole porosities and cracks.Source: §1 · p. 1
- The simulation advances temperature by blurring the prior field to represent distant heat diffusion, then superimposing an Eagar-Tsai solution at the laser position.The figure’s colorbar maximum represents the powder melting point.Source: Figure 1 · p. 2
- Classical optimization can require severe model-order reduction and incur computational expense, while PID control requires tuning and open-loop control cannot respond dynamically to feedback.DRL is presented as a way to learn a control policy with minimal manual tuning.Source: §1 · p. 2
- The framework applies Deep Reinforcement Learning to control melt-pool dimension variation with minimal manual parameter searching.The authors identify this variation as a key mechanism of defect formation.Source: §1 · p. 2
2 Methodology
The method pairs an efficient continuum heat simulation with PPO control, using local temperature observations to adjust laser velocity and reward melt-depth consistency.
2.1 Simulation Description
- The simulation models heat conduction from a moving source on a rectangular plate and reuses translated and rotated Eagar-Tsai line solutions to efficiently build temperature distributions.This makes it feasible to iterate over candidate control policies with lower computational expense than standard finite-element analysis.Source: §2.1 · pp. 2–3
2.2 Convolution and Boundary Conditions
- Boundary heat-source effects are handled with mirrored virtual sources, while diffusion history is approximated by Gaussian convolution with boundary-specific padding.Melt-pool depth is measured by locating where the temperature first exceeds the material's melting temperature below the hottest surface point.Source: §2.2 · p. 4
2.3 Reinforcement Learning Framework
- The policy observes recent local temperature cross-sections, selects laser velocity updates, and is rewarded for target melt depth while penalized for depth variation.The observations comprise nine heat maps from three planes across the current and two previous timesteps; the reward penalizes the episode's minimum-to-maximum melt-depth distance.Source: §2.3 · pp. 4–5
2.4 Proximal Policy Optimization
- PPO optimizes the policy using advantage-based gradient updates with bounded step sizes, and parallel agents collect episode experience to accelerate learning.PPO is selected for continuous control and lower hyper-parameter tuning requirements than comparable methods described in the paper.Source: §2.4 · p. 6
3 Results and Discussion
Across horizontal cross-hatching and concentric-triangle scans, learned velocity policies limit overheating-related melt-depth variation relative to constant velocity. They increase speed near reversals and, for the triangular path, toward the center.
3 Results and Discussion
- PPO was trained for 15,000 updates on two scan paths, using 50-µm action intervals and a target melt-pool depth of 55 µm.The paths were horizontal cross-hatching and concentric triangles.Source: §3 · pp. 7–8
- For horizontal cross-hatching, learned velocity control narrows melt-depth variation versus constant speed, despite slight target undershoots.The constant-velocity case has melt-depth peaks of up to 20 µm at quarter intervals.Source: §3 · p. 8
- For concentric triangles, learned control avoids the constant-speed case’s 40 µm melt-depth jump in the final 20% of the trajectory.The policy increases velocity at direction changes and raises average velocity toward the scan-path center.Source: Figure 4 · p. 7, §3 · p. 8
- For both scan paths, control-policy melt depths have means closer to target and smaller standard deviations than constant-velocity melt depths.The melt pool depth is also described as more stable under the policy for both paths.Source: Figure 5 · p. 8, Figure 6 · p. 9
- Velocity bounds can be imposed to avoid unexamined defects such as spatter, which can result from high laser-motion jerk.Source: §3 · p. 9
4 Conclusion
The paper presents PPO-based velocity control to reduce defect formation in laser powder bed fusion. Simulations show reduced melt-pool variability across scan paths, and the framework is described as supporting closed-loop control and generalization across trajectories and boundary conditions.
4 Conclusion
- PPO optimizes a policy network to maximize expected reward during melting and derive a control policy intended to reduce defect formation.Source: §4 · p. 9
- In simulation, velocity-based policies decrease melt-pool depth variation across scan paths and reduce overheating associated with domain and trajectory geometry.The policy increases velocity where heat may accumulate.Source: §4 · p. 9
- The framework is described as enabling closed-loop control and generalizing to different trajectories and boundary conditions without domain-specific modifications.Source: §4 · p. 9
- Experimental extensions could incorporate physical effects not accounted for in the Eagar-Tsai model into a computationally feasible reinforcement-learning environment.Source: §4 · p. 9