How Machine Learning Reduces Redundancy in Additive Manufacturing

On April 30, 2025, researchers Jiarui Xie and Yaoyao Fiona Zhao published Redundancy Analysis and Mitigation for Machine Learning-Based Process Monitoring of Additive Manufacturing, addressing the critical issue of redundancy in ML systems used for AM. Their study introduces a multi-level framework that reduces latency by 91%, lowers error rates by 47%, and decreases storage requirements by 99.4%, making ML-based monitoring more efficient and cost-effective.

The deployment of machine learning (ML)-based systems in additive manufacturing (AM) has advanced real-time defect detection and process optimization but faces challenges from redundancy, increasing costs and computational demands. This paper defines redundancy as sample-level, feature-level, or model-level and proposes a multi-level redundancy mitigation (MLRM) framework using techniques like data registration, downscaling, cross-modality knowledge transfer, and model pruning. Validated through an in-situ defect detection case study for directed energy deposition (DED), the framework achieved a 91% latency reduction, 47% error rate decrease, and 99.4% storage reduction while lowering sensor costs and energy consumption, enabling efficient, scalable monitoring systems.

In the heart of modern manufacturing lies a quiet revolution, where machine learning (ML) is transforming additive manufacturing (AM), commonly known as 3D printing. This technology, once confined to prototyping, now stands on the brink of reshaping industrial production by addressing critical challenges such as quality consistency and process optimisation.

At the core of this transformation is machine learning, which excels at identifying patterns and anomalies within vast datasets. This capability is particularly valuable in laser powder bed fusion (L-PBF), a widely used AM technique. ML algorithms are now being employed to analyse melt pool dynamics and acoustic emissions, ensuring consistent part quality.

A significant advancement involves real-time monitoring of melt pools during L-PBF processes. The melt pool, crucial for determining the mechanical properties of the final product, is now under constant surveillance using sophisticated models like LSTM-autoencoders. These models analyse high-speed camera data to detect deviations from optimal behaviour, enabling preemptive adjustments and reducing defects.

Beyond visual monitoring, acoustic emission (AE) analysis offers another layer of insight. AE sensors capture sound waves generated during material processing, providing valuable information about the powder bed’s mechanical behaviour. By integrating this data with ML algorithms, researchers can detect early signs of defects, such as porosity or cracking, enhancing quality control.

The integration of ML into AM is driving scalability and efficiency across industries. Real-time monitoring and predictive maintenance reduce material waste and downtime, making AM more viable for large-scale production. This shift is particularly beneficial in sectors like aerospace and automotive, where precision and reliability are paramount.

As research progresses, new sensor technologies and AI advancements promise further enhancements. Collaborative efforts between academia, industry, and technology providers will be crucial in translating these innovations into practical solutions. Machine learning is not just enhancing AM; it’s positioning additive manufacturing as a cornerstone of modern industrial production.

In conclusion, machine learning is unlocking the full potential of additive manufacturing, paving the way for a future where precision and efficiency define industrial production.