Parameter search method

A parameter search method using a classification model, a neural network, a parameter extraction portion, a circuit simulator, and a control portion, comprising:

• a step of supplying measurement data of a semiconductor element and a data set including a process parameter to the parameter extraction portion,
• a step of extracting a model parameter by the parameter extraction portion,
• a step of providing a first netlist from the control portion to the circuit simulator,
• a step of outputting a first output result by the circuit simulator using the model parameter and the first netlist,
• a step of learning the model parameter and the first output result by the classification model, classifying the model parameter, and outputting a first model parameter,
• a step of providing a second netlist and a second model parameter from the control portion to the circuit simulator,
• a step of supplying a first model parameter variable included in the second model parameter from the control portion to the neural network,
• a step of calculating a first action value function Q from the first model parameter variable by the neural network,
• a step of updating the first model parameter variable to a second model parameter variable by the control portion using the first action value function Q and outputting a third model parameter,
• a step of performing simulation by the circuit simulator using the second netlist and the third model parameter and outputting a second output result,
• a step of judging the second output result by the control portion using a convergence condition given to the second netlist,
• a step of setting a high reward by the control portion in the case where the second output result is close to the convergence condition and setting a low reward by the control portion in the case where the second output result is far from the convergence condition when the second output result is judged not to satisfy required characteristics of the second netlist,
• a step of calculating a second action value function Q from the second model parameter variable by the neural network, and
• a step of updating a weight coefficient of the neural network by the neural network using an error calculated using the reward, the first action value function Q, and the second action value function Q,
• wherein when the second output result is judged to satisfy the required characteristics of the second netlist, the first model parameter variable is judged to be the best candidate for the second netlist.