Real-time prediction of temperature field during welding by data-mechanism driving

Research question

How can process observations support fast temperature-field estimates when full-field measurement is difficult and finite-element simulation is slow?

Method

Finite-element temperature fields provide training data. LSTM estimates heat-source parameters from pool-width sequences, and DeepONet uses an updated heat-source function together with current and voltage information to approximate the heat-transfer solution.

Main findings

Against finite-element reference data, the reported DeepONet test achieves a mean absolute temperature error of about 1.12 K and an inference time of about 2.94 ms. These quantify surrogate-model prediction, rather than field temperature-sensor accuracy.