Abstract
In Directed Energy Deposition (DED), melt pool geometry governs part quality, but its depth lies below the surface and cannot be measured directly in real time, a long-standing barrier to closed-loop control. This paper presents a Sim-to-Real transfer learning framework that predicts melt pool width and depth in real time from infrared (IR) imagery and process parameters, bridging physics-based thermal simulation with a small set of experimental data. A multi-input hybrid CNN is pretrained on a large, simulated corpus of surface temperature maps paired with laser power, scan speed, and spot size, then fine-tuned, with all layers unfrozen, on a limited set of labeled IR frames from single-track builds. Against a baseline trained on experiments alone, the fine-tuned model is more accurate and converges faster and more stably, with the clearest gains for depth in the low-data regime (for small training sets it reaches R² > 0.7 where the baseline stays below 0.2). On six previously unseen single-track builds, frame-by-frame real-time prediction yields a mean absolute percentage error (MAPE) of 7.41% for width and 22.45% for depth, versus 29.35% and 74.66% for the baseline. An input ablation confirms that pairing thermal imagery with process parameters gives the best accuracy. Together these results show the approach is practical for sensor-driven monitoring and points toward closed-loop control in metal additive manufacturing.

This publication has 21 references indexed in Scilit: