⚡ electrical engineering

Multi-scale closed-loop melt pool control for LPBF via policy optimization

This paper presents a dual-loop, data-driven control strategy for laser powder bed fusion that combines policy gradient-optimized in-layer feedback with layer-to-layer feedforward control to achieve robust, near-optimal temperature stabilization and significantly reduce tracking errors and constraint violations without requiring manual tuning.

Original authors: Junan Lin, Riccardo Zuliani, Baris Kavas, Markus Bambach, John Lygeros, Efe C. Balta

Published 2026-07-21

📖 7 min read🧠 Deep dive

CC BY 4.0

Original authors: Junan Lin, Riccardo Zuliani, Baris Kavas, Markus Bambach, John Lygeros, Efe C. Balta

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine a factory where robots build complex metal objects not by cutting them out of a block, but by painting them with a super-hot laser, one microscopic layer of powder at a time. This process, called Laser Powder Bed Fusion (LPBF), is like a high-tech version of 3D printing, but instead of plastic, it uses molten steel or titanium. The magic happens when a laser beam melts a tiny spot of powder, fusing it to the layer below. If the laser is too weak, the metal doesn't stick together (a defect called "lack of fusion"). If it's too strong, it boils the metal, creating tiny bubbles or even cracking the part (a defect called "keyholing").

The problem is that this process is incredibly sensitive. As the robot prints, the metal gets hotter and hotter, like a pan on a stove that never cools down. If the robot doesn't adjust the heat instantly, the part warps, cracks, or turns into a useless lump. Traditionally, engineers have to manually tune the robot's settings, guessing the right balance of power and speed. It's a slow, expensive game of trial and error, especially for complex shapes. This paper tackles that headache by teaching the robot to "feel" the heat and adjust itself in real-time, using a smart, two-part control system that learns from mistakes and adapts to chaos.

The Two-Brain Solution for Melting Metal

The researchers at ETH Zürich and inspire AG have developed a clever new way to control this melting process, acting like a dual-brain system for the 3D printer. Instead of just reacting to the heat as it happens, their system has two distinct "brains" working together: one that thinks fast (in-layer control) and one that thinks strategically (layer-to-layer control).

The Fast Brain: Reacting in the Moment
Imagine you are driving a car on a bumpy road. Your reflexes are the "fast brain." When you hit a pothole, you instantly steer or brake to keep the car steady. In the printer, this is the in-layer controller. As the laser scans across a single layer of powder, a high-speed sensor (a pyrometer) measures the temperature thousands of times a second. If the metal gets too hot, this controller instantly dials back the laser power; if it's too cold, it cranks it up.

However, tuning these reflexes is tricky. If you set them too sensitive, the laser might jitter wildly; too slow, and the part burns. The authors used a method called policy optimization (a type of artificial intelligence training) to find the perfect "reflexes." They didn't just guess; they ran thousands of simulations in a computer, throwing in random errors and noise to see what worked best. They found a set of rules that allows the printer to stay stable even when the real world doesn't perfectly match the computer model.

The Strategic Brain: Learning from Past Layers
Now, imagine you are driving that same car, but you also have a map of the road ahead. You know a steep hill is coming, so you slow down before you hit it. This is the layer-to-layer controller. In 3D printing, once a layer is finished, the robot waits for a new layer of powder to be spread. This pause gives the "strategic brain" time to think. It looks at how the previous layer turned out. Did it get too hot? Did the metal swell up? It then calculates a new "plan" for the next layer, adjusting the baseline laser power to compensate for the heat that built up.

This part of the system uses a technique called Iterative Learning Control (ILC). It's like a student taking a test, getting a grade, and then studying the mistakes to do better on the next one. By combining this long-term planning with the fast reflexes of the first brain, the system creates a "closed-loop" that is incredibly robust.

What They Found: Smoother Melts and Fewer Mistakes

The team tested their idea in two ways: first, in a detailed computer simulation, and second, on a real metal 3D printer in a lab.

In the Simulation:
They simulated printing six layers of metal with a "square-spiral" path, which is notoriously difficult because the laser has to stop and turn sharply at the corners, causing heat to pile up. They tested their dual-brain system against two other methods: one that only had the fast brain (in-layer only) and one that only had the strategic brain (layer-to-layer only).
The results were clear. The single-brain systems struggled. The "fast brain" alone couldn't stop the heat from accumulating over multiple layers, and the "strategic brain" alone was too slow to fix sudden temperature spikes. But the dual-loop controller kept the temperature steady, hovering within about 20 Kelvin of the target temperature, even when they introduced significant errors into the simulation to mimic real-world messiness.

In the Real World:
Moving to the physical machine, the researchers had to simplify their system because the actual printer hardware couldn't change settings as fast as the computer simulation could. They adapted the system to use a standard "PI controller" (a common type of industrial controller) with a fixed "feedforward" plan. They tuned this simplified version using their AI method, but only using data from a single, uncontrolled test print.

When they printed actual metal wedges (a shape designed to get very hot), the results were impressive:

Better Accuracy: Their method reduced the average tracking error (how far off the temperature was from the target) by 3.4% compared to a state-of-the-art method called Bayesian Optimization (BO).
Fewer Dangerous Spikes: Most importantly, they reduced the average "input-constraint violation" by 47.5%. This means the laser power stayed much closer to the safe zone. In the BO method, the laser power frequently jumped too high (risking keyholing) or too low (risking lack of fusion). Their controller kept the laser in the "Goldilocks" zone much more often.
Smoother Surfaces: When they looked at the printed metal under a microscope, they saw something fascinating. The uncontrolled prints and the BO prints showed "vector head swelling"—a bulge at the start of every laser line where the metal piled up. Their controller almost eliminated this swelling, resulting in parts that looked much more like the original digital design.

A Surprising Discovery: The Power of "Jitter"

One of the most playful and unexpected findings came when they tested a version of their controller that didn't penalize rapid changes in the laser power. This caused the laser to oscillate (jitter) at a very high frequency, between 2 and 4 kHz.

Usually, engineers try to keep laser power smooth and steady. But here, the "jittery" laser actually reduced the swelling at the start of the lines even more than the smooth version. The authors suggest that these rapid fluctuations might be messing with the fluid dynamics of the molten metal, preventing it from piling up at the start of the scan. While they don't fully understand the physics behind this yet, it opens a new door: maybe a little bit of controlled chaos is exactly what a melting metal pool needs to stay calm.

The Bottom Line

This paper demonstrates that you don't need a massive amount of real-world data to teach a metal printer how to behave. By using a smart simulation to train the controller and then transferring that knowledge to the real machine, the researchers achieved results that match or beat the best existing methods. They proved that combining a fast-reacting reflex with a slow-thinking strategy creates a system that is not only accurate but also robust against the messy, unpredictable nature of melting metal.

While the "jittery" laser effect is still a mystery waiting to be solved, the main takeaway is clear: a dual-brain approach to controlling heat can make 3D printed metal parts stronger, more precise, and less likely to fail. It's a step toward printers that don't just follow instructions, but actually understand the physics of the material they are building with.

Technical Summary: Multi-scale Closed-Loop Melt Pool Control for LPBF via Policy Optimization

Problem Statement
Laser Powder Bed Fusion (LPBF) is a critical additive manufacturing process for producing complex metal geometries. However, the quality of printed parts is highly sensitive to thermal variations, which can lead to defects such as distortion, cracking, and inconsistent microstructures. Existing control strategies generally fall into two categories: in-layer feedback control (adjusting laser power in real-time during a single layer) and layer-to-layer feedforward control (optimizing inputs for subsequent layers based on previous measurements).
Current limitations include:

Manual Tuning: Most methods require extensive manual tuning or large datasets, making them time-consuming and expensive.
Simulation vs. Reality Gap: Many approaches are validated only in simulation or fail to account for model mismatch and measurement noise when deployed on physical systems.
Lack of Integration: Few studies successfully integrate in-layer and layer-to-layer control. Existing combined approaches often impose high computational loads or rely on proprietary modules, leaving questions regarding robustness, reproducibility, and transferability unresolved.
State Estimation: Reliable state estimation for in-layer processes is difficult; relying solely on output measurements often proves insufficient under noise and model uncertainty.

Methodology
The authors propose a computationally efficient, data-driven, dual-loop control strategy to stabilize melt pool temperature using in-situ pyrometer measurements. The architecture consists of two nested loops:

Outer Loop (Layer-to-Layer Control):
- This loop generates optimized feedforward input profiles ( $u^f_N$ ) for each new layer based on measurements from previous layers.
- It combines Temperature Trajectory Optimization (TTO) and Iterative Learning Control (ILC).
- Formulated as a convex quadratic program (QP), it minimizes tracking error and control effort while correcting for model reduction errors, parameter estimation inaccuracies, and measurement noise using a learning gain ( $L_y$ ).
- It updates the reference trajectory for the upcoming layer, ensuring performance improves iteratively.
Inner Loop (In-Layer Control):
- This loop provides real-time feedback ( $u^b_N$ ) during the printing of a single layer using pyrometer signals.
- It employs a linear output feedback controller. Unlike traditional PID tuning, the feedback gains ( $K$ ) are optimized offline using a policy gradient method (specifically, a stochastic gradient descent with the Adam algorithm).
- The optimization problem minimizes the expected tracking error over a distribution of uncertain parameters (e.g., thermal conductivity, porosity) and measurement noise, ensuring robustness before deployment.
- The method leverages a simplified process model with domain randomization to facilitate sim-to-real transfer.

Modeling and Calibration

A control-oriented multi-layer thermal model is derived from a simplified heat equation, discretized into a linear time-varying (LTV) system.
To bridge the simulation-to-reality gap, the authors calibrate the model parameters (absorbed power fraction, beam radius, porosity, thermal conductivities) and an affine sensor map using data from a single uncontrolled layer printed on the physical machine. This offline calibration avoids the need for extensive real-world trial-and-error.

Key Contributions

Policy-Gradient Optimized Gains: The in-layer controller gains are derived via simulation using a policy gradient method, requiring minimal real-world tuning or print data.
Computational Efficiency: The feedback structure relies solely on scalar output measurements and a single-layer trajectory optimization, making it suitable for real-time operation.
Robustness: The controller is explicitly designed to handle significant model mismatch and sensor noise, demonstrating near-optimal performance in both simulation and physical experiments.
Sim-to-Real Transfer: The work presents one of the first successful applications of sim-to-real policy optimization in LPBF, where the controller is tuned entirely offline using data from a single calibration layer.

Results

Simulation: In a fully simulated environment with 729 test models (varying parameters and noise), the dual-loop controller significantly outperformed single-loop in-layer and layer-to-layer controllers. It maintained a mean absolute error of approximately 20 K (3% of the operating range) despite large perturbations.
Physical Experiments: The method was tested on an Aconity3D Midi+ system printing 316L stainless steel wedges.
- Performance: The proposed controller matched the state-of-the-art performance of a Bayesian Optimization (BO) tuned baseline.
- Error Reduction: It reduced the mean tracking error by 3.4% and the mean input-constraint violation by 47.5% compared to the BO baseline.
- Defect Mitigation: The controller effectively eliminated severe overheating and reduced "vector head swelling" (geometric distortion at the start of scan vectors).
- Microstructure: While porosity was slightly higher than in uncontrolled builds (due to symmetric penalization of input constraints), the controlled parts showed reduced pore size and count compared to the BO baseline.
- High-Frequency Dynamics: Experiments revealed that controllers without rate penalization induced high-frequency (2–4 kHz) oscillations in laser power. Surprisingly, this oscillatory behavior correlated with reduced vector head swelling, suggesting a new avenue for research into melt pool dynamics, though the physical mechanism remains to be fully resolved.

Significance and Claims
The paper claims that this work marks a significant step toward robust, automated control in LPBF by successfully integrating in-layer and layer-to-layer strategies via policy optimization. The authors emphasize that their approach achieves state-of-the-art performance with minimal experimental data (a single calibration layer) and demonstrates that sim-to-real policy optimization is viable for complex thermal processes. The discovery of high-frequency excitation dynamics reducing geometric swelling is highlighted as a novel finding that opens new research avenues in additive manufacturing, even though the underlying physical mechanisms require further targeted investigation. The work underscores the importance of component-specific controller design and the potential of combining offline optimization with real-time feedback to enhance part quality and geometric fidelity.

The Two-Brain Solution for Melting Metal

What They Found: Smoother Melts and Fewer Mistakes

A Surprising Discovery: The Power of "Jitter"

The Bottom Line

More like this