⚡ electrical engineering

Machine Learning-based Surrogate Model for Melt Pool Control in Laser-based Direct Energy Deposition

This paper presents a novel machine learning-based surrogate controller coupled with a finite element solver that enables real-time, autonomous adjustment of laser power in Laser-based Direct Energy Deposition to maintain stable melt pool penetration, thereby enhancing geometric precision and reducing computation time while demonstrating robust generalization across diverse scanning sequences.

Original authors: Runeal Ramma, Carlos Moreira, Michele Chiumenti, Manuel Alejandro Caicedo, Raj Das, Zhijun Ji, Andrey Molotnikov

Published 2026-07-21

📖 3 min read☕ Coffee break read

CC BY 4.0

Original authors: Runeal Ramma, Carlos Moreira, Michele Chiumenti, Manuel Alejandro Caicedo, Raj Das, Zhijun Ji, Andrey Molotnikov

Original paper licensed under CC BY 4.0 (https://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 you are trying to build a castle out of molten chocolate using a high-powered heat gun. You want the layers to stick together perfectly, but if you hold the heat gun at the same setting the whole time, the bottom of your castle gets too hot and starts to melt into a gooey mess, while the top stays too cold to stick. This is the tricky reality of a manufacturing technique called Laser-based Direct Energy Deposition (DED-LB). It's like 3D printing with metal, where a laser melts powder to build parts layer by layer. The big problem is "heat accumulation": as the machine builds up, the part gets hotter and hotter, making the melted metal pool grow too big and deep, which ruins the shape and strength of the final object. For years, engineers have tried to fix this by either stopping to cool down or by guessing the right settings, but these methods are slow or imprecise. The goal is to find a way to automatically tweak the laser's power in real-time, like a smart thermostat that knows exactly when to turn the heat up or down to keep the "melt pool" at the perfect depth.

This paper introduces a clever new solution: a "surrogate controller" powered by Machine Learning (ML) that acts like a super-fast, super-smart co-pilot for the laser. Instead of waiting for the slow, heavy calculations of traditional computer simulations to figure out the right power settings, the researchers trained an AI model to predict the perfect laser power instantly. They first used a high-fidelity computer simulation (a very detailed digital twin of the process) to generate a massive library of "what-if" scenarios. They taught the AI to recognize patterns in the heat and the shape of the melted metal, specifically focusing on keeping the "penetration depth" (how deep the laser melts into the metal) constant at 0.40 mm. Once trained, this AI model was hooked up to the simulation to control the laser in real-time.

The results show that this AI-driven approach is a game-changer for speed and stability. In tests involving different shapes like thin walls and hollow cubes, the AI controller managed to keep the melt pool depth steady, just as well as the slow, traditional method, but it did it roughly three times faster. For example, while the old method took about 141 hours of computer time to figure out the power settings for a hollow cube, the AI did it in about 57 hours. Even more impressively, the AI successfully controlled complex shapes it had never seen before during its training, proving it can generalize its knowledge to new jobs. The study suggests that by using this ML-based surrogate, manufacturers can reduce heat buildup, prevent defects, and print complex metal parts with much greater geometric precision, all while cutting down the time needed to plan the process. The authors note that while the AI is currently a simulation-based tool, it offers a robust path toward autonomous, real-time control for the future of metal 3D printing.

Technical Summary: Machine Learning-based Surrogate Model for Melt Pool Control in Laser-based Direct Energy Deposition

Problem Statement

Laser-based Direct Energy Deposition (DED-LB) is a critical technology for metal additive manufacturing (AM), particularly for component repair and complex geometry fabrication. However, its broader adoption is hindered by challenges in dynamic process control, specifically heat accumulation. Employing constant process parameters, such as a fixed laser power, leads to significant thermal distortion, elevated residual stresses, and substrate cracking. While numerical approaches and controllers exist to optimize parameters, traditional methods often rely on iterative offline optimization or parameter-dependent controllers that lack adaptability to new scanning sequences or environmental changes. Furthermore, there is a gap in frameworks that can efficiently compute optimal processing parameters while rapidly adapting to new scenarios to ensure geometric accuracy and repeatability.

Methodology

The authors propose a novel strategy integrating a Machine Learning (ML) model as a surrogate controller, coupled with a Finite Element (FE) solver, to achieve run-time process control. The methodology is executed in two distinct phases:

1. Offline Optimization and Data Generation

Simulation Framework: High-Fidelity (HF) thermal analysis is performed using FEMUSS, an in-house multiphysics software. The model is calibrated against experimental data from a Ti6Al4V thin-wall deposition, minimizing errors between measured and simulated thermal profiles to determine heat transfer coefficients (HTCs).
A-Posteriori Optimization: An iterative offline loop adjusts the laser power supply to maintain a user-defined target melt pool penetration depth ( $D_{target} = 0.40$ mm). This target was selected to ensure remelting of the previous layer (preventing delamination) while minimizing lack of fusion.
Control Logic: A proportional feedback control approach ( $K_p$ ) iteratively corrects the power supply based on the difference between observed and target penetration depths. This process generates a dataset of optimized power profiles and corresponding thermal/geometric signatures.

2. Machine Learning Model Development

Dataset Construction: Synthetic data was generated from 42 simulated cases involving unidirectional and bidirectional scanning sequences of thin walls (10–110 mm).
Feature Selection: A Spearman's rank correlation analysis identified key input features, including maximum temperature, melt pool temperature, pulse temperature, next region temperature, melt pool volume, and penetration depth. The "next region penetration depth" was pre-set to the target value (0.40 mm) to act as a safeguard during training.
Architecture: A Regression Artificial Neural Network (R-ANN) was designed using Bayesian hyperparameter tuning (via Keras). The optimal architecture includes one hidden layer with 15 neurons, ReLU activation, and Mean Absolute Error (MAE) as the loss function. The model was trained using the Adam optimizer with early stopping to prevent overfitting.
Closed-Loop Integration: The trained ML model is integrated with the FE solver in a closed loop. During run-time, the solver provides current thermal and geometric states to the ML model, which instantly predicts the corrective laser power to maintain the target penetration depth.

Key Contributions

Rapid Surrogate Control: The framework derives a variable laser power profile ensuring constant melt pool penetration depth in a single iteration, eliminating the need for iterative offline optimization or experimental trial-and-error.
Computational Efficiency: The ML-based approach achieves results approximately 3 times faster than the offline optimization approach (e.g., reducing computation time from ~40 hours to ~12 hours for a thin wall).
Generalization Capability: The model successfully predicts optimized power profiles for scanning sequences (hollow cubic and infill-partitioned cubic) that were not included in the training dataset, demonstrating robust generalization.
Thermal Management: The system effectively generates necessary power reductions to minimize heat accumulation, maintaining consistent melt pool morphology (depth and length) across different geometries.
Feature Interpretability: Using SHapley Additive exPlanations (SHAP), the study quantifies the importance of input features, revealing that thermal characteristics (substrate temperature, melt pool length) are paramount in influencing power predictions, often outweighing geometric features.

Results

The performance of the ML-based surrogate controller was evaluated against constant power supply and offline optimization strategies across three numerical examples: a unidirectional thin wall, a hollow cubic deposition, and an infill-partitioned cubic deposition.

Penetration Depth Stability:
- Constant Power: Penetration depth grew uncontrollably.
- Offline Optimization: 98.51% of penetration depth values fell within the target window (0.38–0.42 mm) for the hollow cubic case.
- ML Surrogate: 91.77% of values fell within the target window for the hollow cubic case, and 87.68% for the infill-partitioned cubic case.
Melt Pool Length: The ML controller maintained melt pool length within the target range (3.18–3.55 mm) with 82.81% (hollow) and 87.54% (infill) adherence, comparable to or exceeding the offline optimization in some metrics.
Error Metrics: On an unseen validation case (cylinder), the ML model achieved a Mean Absolute Error (MAE) of 100.4 W and a Mean Absolute Percentage Error (MAPE) of 5.55% compared to the offline optimized power profile.
Thermal Profiles: Both offline and ML strategies significantly reduced thermal accumulation compared to constant power, with the ML strategy showing smoother temperature transitions at the start of hatches.

Significance

The paper claims that this work provides a robust framework for the precise and scalable control of metal AM processes. By coupling an ML surrogate with an FE solver, the system enables autonomous, run-time control that adapts to dynamic thermal conditions without human intervention. The ability to generalize to unseen scanning sequences suggests the method can serve as a reliable tool for rapid prototyping and the manufacturing of complex structural parts, significantly reducing the time and cost associated with parameter optimization. The authors note that while the current model uses a regression network, future work will investigate Recurrent Neural Networks (RNNs) to better capture temporal dependencies in time-series data for even more responsive control.