使用深度学习与核碰撞判断中子皮类型的尝试

Determining the neutron skin types using deep learning and nuclear collisions:An attempt

中子在原子核内的分布对确定中子皮厚度、原子核的对称能、致密核物质状态方程以及中子星质量半径关系都极为重要.但提取中子在原子核内的分布异常困难.理论上,重原子核的多体量子从头算方法遭遇维数灾难.实验上,中子不带电,不像电荷分布(质子分布)那样可以直接测量.杰佛逊实验室的宇称破缺电子散射实验,通过宇称破缺测量弱荷(中子)分布.我们尝试从重离子碰撞的末态强子分布提取中子在原子核内的分布.原则上,质子和中子在原子核内的分布对应着核碰撞的初态涨落和关联,这些信息经过碰撞后的相对论流体力学演化和强子输运会最终转化为末态强子在动量空间的关联.如何建立核结构的初态涨落与末态关联之间的映射,是利用重离子碰撞确定初态核结构的关键.正向过程中,我们从halo类型、skin类型以及无中子皮(noskin)三种不同的中子分布抽样208Pb原子核,并使用相对论分子动力学模型SMASH进行halo-halo,skin-skin以及noskin-noskin三类核核碰撞模拟,得到末态强子的动量信息.人工智能技术中的深度神经网络拥有非常强大的模式识别能力和在不同数据类型之间建立映射的能力.我们使用点云神经网络和卷积神经网络,分别根据抽样出的初态核子分布和碰撞后的末态强子分布,判断208Pb原子核的中子皮类型.因为涨落效应的存在,深度神经网络使用初态核子的坐标仅可在halo和skin类型两分类任务中达到62%的分类精度.使用末态强子的动量信息无法在统计误差范围内成功分类halo-halo和skin-skin两种类型的核碰撞.但在三分类任务中,使用末态强子信息,分类准确度达到37.47%,超过3万个测试样本三分类随机猜测准确度(33.3%)约4个百分点.即人工神经网络虽然无法区分中子皮的种类,但能以一定的概率从核碰撞末态确定208Pb原子核中是否存在中子皮.如果筛选出超级擦边碰撞事件,使用多样本混合的方法和参与碰撞粒子的电荷信息,人工神经网络在halo-halo和skin-skin分类中拥有大约56%的分类精度.

The distribution of neutrons inside the nucleus is essential for determining the neutron skin thickness,symmetry energy,equation of state of dense nuclear matter,and the mass-radius relationship of neutron stars.However,obtaining the neutron distribution is extremely challenging.Theoretically,the quantum many-body ab-initio calculations of heavy nucleus suffer the curse of dimensionality.Experimentally,measuring neutrons is not as easy as protons because protons explicitly carry electric charges.Parity violating electron scattering experiment at Jefferson laboratory determines the neutron distribution through the weak charge distribution.We tried to extract the neutron distribution inside the nucleus from the final state hadrons of heavy ion collisions.In principle,the proton and neutron distributions inside the nucleus correspond to the initial state fluctuations and correlations.The initial state fluctuations are converted to correlations of the final state hadrons in momentum space.Mapping the final state to the initial state is the key to determining nuclear structure from heavy ion collision.We sampled208Pb nucleons from halo-type,skin-type,and no-neutron-skin distributions and performed halo-halo,skin-skin,and noskin-noskin heavy-ion-collision simulations using the relativistic molecular dynamics program SMASH.Artificial intelligence and deep neural networks have powerful pattern recognition abilities,which can map data from two different domains.Using a point cloud network and convolutional neural network,we classified the neutron skin types using the initially sampled nuclei and final state hadrons after the collision.Because of the strong fluctuations from Monte Carlo sampling,the deep neural network only attained 62%prediction accuracy on the initial state.It failed to predict the neutron skin types using the final state hadrons.However,the network achieved a prediction accuracy of 37.47%on the 3-types classification,which was 4%higher than the random-guess accuracy of 33.3%with 30000 testing samples.This study shows that the network might not determine the neutron skin types from the final state hadrons.However,it can identify whether there is a neutron skin using the final state hadrons with non-zero probability.If ultraperipheral collisions are selected,the network can classify the halo-halo and skin-skin types with a testing accuracy of 56%,using the charge information and multi-event mixing method.