Real-time digital twin inference for melt pool dynamics using deep neural operators

This study explores melt pool dynamics in metal additive manufacturing, which serve as a critical determinant of microstructure evolution, material uniformity, process stability, and defect mitigation in fabricated components. Although conventional numerical methods such as computational fluid dynamics and emerging physics-informed neural networks have demonstrated predictive capabilities, their substantial computational cost presents a significant constraint, particularly under varying material properties or boundary conditions that necessitate costly retraining procedures. This investigation presents the Deep Operator Network, denoted as DeepONet, as a computationally efficient surrogate model for real-time melt pool prediction within digital twin frameworks. The proposed model accurately captures velocity, pressure, and temperature fields across diverse process conditions and time increments, achieving high predictive performance with an R² value of 0.99, as verified through extensive benchmarking. Ablation studies further confirm the model's robustness to variations in network depth and neuron configuration, underscoring its adaptability across a range of scenarios. These attributes establish DeepONet as a highly effective virtual sensor for melt pool monitoring in additive manufacturing digital twins, offering a favorable balance between computational efficiency and physical fidelity.