Cometary dust is refractory particles lifted and ejected by sublimation of volatiles from the surface of comets with typical diameters from sub-micrometers to centimeters. These particles distribute around and recede from cometary nuclei and are illuminated by the Sun, forming the observable dust comae and tails. The fundamental characteristics of cometary dust, such as their size and size distribution, ejection velocity from the nuclei, etc., are of great significance to understanding the formation and evolution of comets. This paper presents a method, the Machine Learning Based Dynamic Parameter Fitting method (MLDPF method), for deriving the fundamental properties of cometary dust based on dynamical simulation and machine learning. In this process, the Monte Carlo method is used to generate dust particles in the assumed parameter space and solve for the spatial distributions of dust at the times of observations according to dynamical models. Then, we use these simulations to train a CNN (Convolutional Neural Networks) model, and finally fit the observed photos to derive the parameters of the cometary dust. Using this approach, we analyzed the ground-based images of Comet 103P/Hartley collected in the visible-band. The dust ejected from 103P is dominated by micro-particle is 2.11 ± 0.41 μm in radius, and follows an assumed exponential dust size distribution with a coefficient of −4.26 ± 1.41. The initial ejection velocity of dust is 87 ± 13.4 m/s. The dust producing rate is about 11.35 × 10¹² s⁻¹ according to the Afρ parameter obtained from the optical photos and the best-fit parameters. The dust emission is likely to be solar insolation dependent. The machine learning and the fitting process were able to converge to a set of solutions that are in good agreement with the previous analyses in the literature. The MLDPF method can establish a cometary dust parameter-image model by machine learning, in which training and test set of the model were calculated based on simulated images; and then the model can be applied to real images. It has a wide range of applicability to different comets and can be used to predict comet morphology. This method also has the scalability of parameters and can be used for the study of cometary dust under more complex parameters and dynamics.

彗星尘埃是通过从彗星表面的挥发物升华而升起和喷出的难熔颗粒，典型直径从亚微米到厘米。这些粒子分布在彗核周围并远离彗核，并被太阳照亮，形成可观测的尘埃彗尾和彗尾。彗星尘埃的基本特征，如它们的大小和分布、原子核的喷射速度等，对理解彗星的形成和演化具有重要意义。本文提出了一种方法，即基于机器学习的动态参数拟合方法（MLDPF 方法），用于基于动力学模拟和机器学习推导彗星尘埃的基本特性。在这个过程中，蒙特卡罗方法用于在假设的参数空间中生成尘埃粒子，并根据动力学模型求解观测时尘埃的空间分布。然后，我们使用这些模拟来训练一个 CNN（卷积神经网络）模型，最后拟合观察到的照片来推导出彗星尘埃的参数。使用这种方法，我们分析了在可见光波段收集的彗星 103P/Hartley 的地面图像。从 103P 喷出的尘埃以半径为 2.11 ± 0.41 μm 的微粒为主，并遵循系数为 -4.26 ± 1.41 的假设指数级尘埃粒径分布。粉尘的初始喷射速度为 87±13.4 m/s。产尘率约11.35×10 我们使用这些模拟来训练一个 CNN（卷积神经网络）模型，最后拟合观察到的照片来推导出彗星尘埃的参数。使用这种方法，我们分析了在可见光波段收集的彗星 103P/Hartley 的地面图像。从 103P 喷出的尘埃以半径为 2.11 ± 0.41 μm 的微粒为主，并遵循系数为 -4.26 ± 1.41 的假设指数级尘埃粒径分布。粉尘的初始喷射速度为 87±13.4 m/s。产尘率约11.35×10 我们使用这些模拟来训练一个 CNN（卷积神经网络）模型，最后拟合观察到的照片来推导出彗星尘埃的参数。使用这种方法，我们分析了在可见光波段收集的彗星 103P/Hartley 的地面图像。从 103P 喷出的尘埃以半径为 2.11 ± 0.41 μm 的微粒为主，并遵循系数为 -4.26 ± 1.41 的假设指数级尘埃粒径分布。粉尘的初始喷射速度为 87±13.4 m/s。产尘率约11.35×10 从 103P 喷出的尘埃以半径为 2.11 ± 0.41 μm 的微粒为主，并遵循系数为 -4.26 ± 1.41 的假设指数级尘埃粒径分布。粉尘的初始喷射速度为 87±13.4 m/s。产尘率约11.35×10 从 103P 喷出的尘埃以半径为 2.11 ± 0.41 μm 的微粒为主，并遵循系数为 -4.26 ± 1.41 的假设指数级尘埃粒径分布。粉尘的初始喷射速度为 87±13.4 m/s。产尘率约11.35×10¹²  s ^-1根据从光学照片中获得的Afρ参数和最佳拟合参数。粉尘排放很可能与日照有关。机器学习和拟合过程能够收敛到一组与文献中先前的分析非常一致的解决方案。MLDPF方法可以通过机器学习建立彗星尘埃参数-图像模型，其中基于模拟图像计算模型的训练和测试集；然后该模型可以应用于真实图像。它对不同彗星具有广泛的适用性，可用于预测彗星形态。该方法还具有参数的可扩展性，可用于更复杂参数和动力学下的彗星尘埃研究。