Physics Informed Neural Network (PINN) to Predict Probe Card Key Performance Indicators and Physics of Degradation in Microelectronics

Probe cards are a fundamental component of a wafer testing cell. It is an interface between an electronic test system (ETS) and a semiconductor wafer whose main purpose is to provide an electrical path between the test system equipment and the circuits on the wafer, hence permitting the testing and validation of the circuits at the wafer level, usually before they are diced and packaged. Several key performance indicators affect the reliability of these probe cards, which is critical for a testing setup. Some of these indicators refer specifically to probes, e.g. buckling load, maximum stress at nominal overtravel, probes lateral force, current carrying capability (CCC), scrub mark dimensions, and average lifetime (expressed in touch downs). Normally, the test that is used to evaluate the reliability of probes is a full scale lifetime assessment on over 1M+ touch downs; however, an optimization of the design parameters through this technique results in huge loss of time and finances. It thus results that a reduction of the number of experiments becomes a critical point. In the recent past, multiple studies have adapted the finite element-based design-on-simulation technology for the assessment of the reliability of these probe cards. As this analysis is time consuming, Artificial intelligence (AI) can be of critical importance to overcome these shortcomings. This research work intends to propose an AI-assisted model based on a Physics-Informed Neural Network (PINN) to predict probe card key performance indicators, devising a strategy for weakest link physics of degradation. As a first step the focus is directed towards a relevant performance indicator: the probe buckling force. A PINN has been employed to determine the nominal buckling load at which probe card pins suffer a malfunction. Then, the system reliability is assessed by incorporating the proper governing equations and the boundary/loading conditions to affect it. Another motivation behind this adapted road map is to sidestep a frequently encountered obstacle associated with almost every AI model, and that is to require extensive data for training. The solution of the governing (mechanical) equations is used as a training parameter for the AI-based model, which means the model will not require any synthetic or experimental data for training.