Patent DE 10 2021 126 108

When simulating analogue, digital or analogue-digital circuits on the basis of behavioural models, it is essential how accurately the behavioural models describe the behaviour of the circuit. A model error can lead to a chip design being classified as correct, even though problems are present in the associated circuits and therefore remain undetected. It is therefore important for the user of the behavioural model to know the model error and to be able to adjust the model accuracy and thus increase the reliability of the simulation results.

The idea of the invention is an AI-based error estimator that reads the inputs and the output of the behavioural model in parallel to the simulation and estimates the model error based on this. The error can thus be read online. This makes it possible to better estimate the output of the model, react to the error and, for example, change simulation parameters or cancel the simulation.

The model (01) is made up of elements that represent at least groups of semiconductor components of the circuit to be simulated. The model (01) is used in a learning mode to learn an error estimator (08) formed by a machine-learning system by feeding first test signals to inputs of the circuit to be simulated and to corresponding inputs of the model (01) and by comparing at least one resulting output signal of the model (01) with at least one resulting corresponding output signal of the circuit to be simulated. The model (01) is extended by the trained error estimator (08). The model (01) extended by the learnt error estimator (08) is used in a simulation mode for simulating and verifying the electronic circuit.