Knowledge-based neural network (KBNN) modeling of HBT junction temperature and thermal resistance from electrical measurements

A knowledge-based neural network (KBNN) modeling and analysis method is presented for determining junction temperature, Tj, and thermal resistance, Rth, from simple electrical measurements of HBTs. The method retains sound physical principles of classical approaches but provides significant additional practical benefits for modeling and prediction based on the mathematical properties of neural networks when endowed with additional a priori “knowledge”. The method returns explicit, infinitely differentiable approximations for Tj and Rth as functions of ambient temperature, Tamb, and power dissipation, Pdiss. The method enables reliable predictions of Tj over a very wide range (e.g. 30°C to 260°C) by working with any complete set of experimental variables. The method also provides an automatically trained, measurement-based DC electro-thermal transistor model as a function of bias and temperature.