The inversion of the geometric parameters of the ellipsoidal particle group through the extinction coefficient is an important problem in the field of particle measurement. The traditional inversion technology based on evolutionary algorithm requires multiple numerical integration to solve the extinction coefficient, which is low in efficiency. To solve this problem, an acceleration method based on machine learning is proposed in this work. First, the particle size and shape are expressed parametrically; second, the training and testing datasets of ellipsoidal particle extinction coefficient are established based on the anomalous diffraction approximation theory; finally, the mapping between particle parameters and extinction coefficient is realized by using multilayer perceptron artificial neural network, and the effects of the number of neurons, wavelength, particle group distribution model and other factors on the prediction accuracy and efficiency are studied. Experimental results show that when the number of hidden layer neurons is 20, the average prediction error is less than 0.05%, and the single machine prediction time is about 0.6 μs. The technology provides an efficient and accurate extinction coefficient calculation tool. With further employment of evolutionary algorithms, it is expected to realize the real-time inversion of spherical and ellipsoidal particle size and shape parameters.