Lifetime Prediction of Semiconductor Devices Based on Deep Learning

The reliability of power semiconductor devices, such as insulated gate bipolar transistors (IGBTs), is crucial for aerospace and industrial applications. Traditional prediction methods face challenges in multimodal data, integration physical constraints, and feature extraction accuracy. This study proposes a physics-informed deep learning framework for remaining useful life(RUL) prediction, by fusing high-frequency transient waveforms, steady-state thermal measurements, and electrical characterization data from source measurement units(SMUs). A hybrid architecture combines dilated convolutional neural networks (Dilated CNNs), to capture multi-scale transient features, long short-term memory (LSTM) networks with attention mechanisms for thermal sequence modeling and physics-guided loss functions incorporating the Coffin-Manson fatigue model. Experimental validation on NASAs accelerated aging dataset devices 2-5 demonstrated rapid convergence, with validation loss decreasing from 8362.9460 to 0.0224, and training loss from 0.4585 to 0.1121 over 100 epochs. The model achieved an RMSE of 0.0536 and an MAE of 0.0523, significantly outperforming non-dilated CNN baselines in convergence speed and stability.