A physics-informed knowledge distillation model with spatial–temporal attention for energy consumption pre-assessment in sustainable additive manufacturing

Abstract

To achieve sustainable additive manufacturing (AM), a novel data-driven model termed, physics-informed knowledge distillation (PKD), is proposed to accurately pre-assess the energy consumption at the product design stage. The energy consumption pre-assessment modeling is first constructed in light of the operating characteristics of AM technology. The modeling analysis indicates that physics data plays a pivotal role in determining energy consumption, while current data-driven approaches did not take full advantage of it. Hence, we design a physics-informed architecture with spatial–temporal attention in our PKD to jointly integrate the geometric and physical information from the product design process. Different from classical knowledge distillation, our teacher network and student network utilize an asymmetric physics-informed architecture respectively, encouraging the student to span the knowledge gap over its teacher. A knowledge distillation framework is further devised to enable our PKD to reduce computational costs while retaining pre-assessment performance. Extensive experimental findings show our PKD yields Root Mean Squared Error (RMSE), Mean Error Ratio (MER), and Total Error Rate (TER) as low as 300.16, 5.93%, 1.76%, respectively. It demonstrates that the proposed PKD outperforms state-of-the-art energy consumption pre-assessment methods. The well-distilled PKD is capable of providing referenced energy consumption value for designers and manufacturers to make decisions and manage energy for achieving sustainable additive manufacturing.