Laser ablation is critical in material processing, but predicting outcomes like ablation depth is challenging, especially when materials or experimental conditions vary. Machine learning (ML) models can provide accurate predictions but often struggle with generalization to different conditions, especially when data is limited. This work demonstrates how combining a physics-based model with a neural network can overcome these limitations even with limited data.

Our dataset consists of depth map images, with 500–1000 datapoints for each material. We begin with the physics-based Two-Temperature Model (TTM), using three adjustable parameters: d (equivalent to penetration depth), Fth (fluence threshold), and S (incubation coefficient). A neural network refines the model's predictions by adjusting discrepancies between the TTM and experimental results. The hybrid TTM+NN model was initially trained on silicon data, showing significant improvement over the pure TTM model.

When applied to stainless steel, the hybrid model's performance dropped, as expected with new materials and limited data. To address this, we used transfer learning, fine-tuning the silicon-trained model with only 10 carefully selected datapoints from steel, chosen via Bayesian optimization. This approach improved the model's R² from -0.40 to 0.67, demonstrating the effectiveness of both the hybrid and transfer learning approaches.

This methodology is particularly valuable in industrial applications where data acquisition is costly and time-consuming, enabling faster tuning and deployment of laser processes on new materials without large datasets. By combining physics-based modeling with machine learning, this approach offers predictive accuracy and robustness, making it a practical solution for real-world laser processing applications.