Meta-Learning Enables Precise Geometry Prediction in Laser-Directed Energy Deposition

Global: Meta-Learning Enables Precise Geometry Prediction in Laser-Directed Energy Deposition

Background

Researchers have identified that predicting bead geometry in laser-directed energy deposition (L‑DED) is complicated by limited and diverse experimental data, which vary across materials, machine settings, and process parameters.

Proposed Approach

The study introduces a cross‑dataset knowledge‑transfer framework that leverages gradient‑based meta‑learning algorithms—Model‑Agnostic Meta‑Learning (MAML) and Reptile—to rapidly adapt to new deposition conditions using only a few training examples.

Experimental Setup

Multiple datasets were compiled from peer‑reviewed literature and in‑house experiments, encompassing powder‑fed, wire‑fed, and hybrid wire‑powder L‑DED processes. The meta‑learning models were trained on a subset of these datasets and evaluated on unseen target tasks.

Performance Outcomes

Both MAML and Reptile achieved accurate bead‑height predictions on previously unseen tasks with as few as three to nine training samples, consistently outperforming conventional feedforward neural networks trained under comparable data constraints.

Across diverse target tasks, the meta‑learning models attained R‑squared values up to approximately 0.9 and mean absolute errors ranging from 0.03 mm to 0.08 mm, indicating strong generalization across heterogeneous L‑DED settings.

Significance

The findings demonstrate that meta‑learning can effectively transfer knowledge across heterogeneous additive‑manufacturing datasets, potentially reducing the data burden for new material or machine configurations and accelerating process optimization.

This report is based on information from arXiv, licensed under Academic Preprint / Open Access. Based on the abstract of the research paper. Full text available via ArXiv.