This paper introduces a methodology for optimizing neural network models in precipitation forecasting using a combination of continuous and categorical binary indices. The authors propose a novel multi-objective loss function that integrates these indices, which are typically not differentiable and therefore cannot be directly optimized using gradient descent. By using the sigmoid function to approximate the logical comparison operators, they create differentiable versions of popular categorical indices such as Probability of Detection (POD) and False Alarm Rate (FAR). The experimental section tests this methodology by training a U-net neural network to estimate precipitation using geopotential heights as input. The results, evaluated using metrics like ROC curves, demonstrate how the proposed indices can be used to optimize the skill of precipitation models. This is the first proposal of a methodology for optimizing weather neural network models based on categorical indices.This paper introduces a methodology for optimizing neural network models in precipitation forecasting using a combination of continuous and categorical binary indices. The authors propose a novel multi-objective loss function that integrates these indices, which are typically not differentiable and therefore cannot be directly optimized using gradient descent. By using the sigmoid function to approximate the logical comparison operators, they create differentiable versions of popular categorical indices such as Probability of Detection (POD) and False Alarm Rate (FAR). The experimental section tests this methodology by training a U-net neural network to estimate precipitation using geopotential heights as input. The results, evaluated using metrics like ROC curves, demonstrate how the proposed indices can be used to optimize the skill of precipitation models. This is the first proposal of a methodology for optimizing weather neural network models based on categorical indices.