Nanjing Tech AI Breakthrough: Real-Time Temperature Forecasting Revolutionizes Metal Additive Manufacturing

Researchers affiliated with Nanjing Tech University have unveiled an innovative artificial intelligence model, marking a significant leap forward in the field of additive manufacturing. This novel system, which is fundamentally grounded in physical principles, is specifically engineered to provide operational temperature forecasting during the Wire Arc Additive Manufacturing (WAAM) process.

The publication of these findings in the prestigious journal Communications Engineering signifies a major milestone toward enhancing the quality and stability of metallic printed components. The core of this development lies in the creation of a physics-informed geometric recurrent neural network. This sophisticated hybrid methodology successfully merges fundamental laws of physics with the computational power of deep learning, enabling the dynamic modeling of material thermal behavior.

A crucial achievement of this research is its operational speed. While conventional techniques, such as finite element modeling, can require up to an hour to yield accurate results, the new AI model demonstrates astonishing efficiency, delivering temperature predictions in just 12 milliseconds. This rapid response capability is absolutely vital for implementing effective closed-loop feedback systems within demanding industrial environments.

The team, spearheaded by Minxuan Tian from the School of Mechanical and Power Engineering, tackled a two-fold challenge: overcoming the computational sluggishness inherent in classical simulations and minimizing the typical error accumulation found in purely data-driven models. In practical experiments, which involved the layer-by-layer printing of thin-walled steel structures using a robotic WAAM setup, the model exhibited a maximum prediction error of approximately 4.5% in simulations and 13.9% during actual testing.

Furthermore, the system reliably forecasts temperature evolution for a duration of up to 10 seconds. This specific timeframe represents a key window for actively managing thermal flows and mitigating residual stresses. Temperature control is the cornerstone of metallic 3D printing; uneven heating or cooling invariably leads to critical defects, such as cracking and deformation, thereby compromising the structural integrity of the final product.

By embedding physical constraints directly into the neural network's architecture, the researchers ensured that the predictions remain physically sound. This design choice allows the system to generalize effectively across diverse geometric shapes and various process parameters. The team also highlighted that utilizing transfer learning significantly reduces training time, boosting the technology's practical applicability and adaptability across varied manufacturing conditions. This methodological advance paves the way for implementing feed-forward control systems in additive manufacturing, allowing machinery to adjust parameters—like heat input or wire feed rate—before potential issues even materialize.