In power electronics, ultrasonic wire bonding is

used to connect the electrical terminals of power modules.

To implement a self-optimization technique for ultrasonic wire

bonding machines, a model of the process is essential. This

model needs to include the so called ultrasonic softening effect.

It is a key effect within the wire bonding process primarily

enabling the robust interconnection between the wire and a

substrate. However, the physical modeling of the ultrasonic

softening effect is notoriously difficult because of its highly

non-linear character and the absence of a proper measurement

method. In a first step, this paper validates the importance of

modeling the ultrasonic softening by showing its impact on the

wire deformation characteristic experimentally. In a second step,

the paper presents a data-driven model of the ultrasonic softening

effect which is constructed from data using machine learning

techniques. A typical caveat of data-driven modeling is the need

for training data that cover the considered domain of process

parameters in order to achieve accurate generalization of the

trained model to new process configurations. In practice, however,

the space of process parameters can only be sampled sparsely.

In this paper, a novel technique is applied which enables the

integration of prior knowledge about the process into the datadriven

modeling process. It turns out that this approach results in

accurate generalization of the data-driven model to novel process

parameters from sparse data.