We compare two machine-learning-based approaches, artificial neural network (ANN) and micromechanical model with automatic Bayesian identification of the model parameters, in application to mimicking the deformation behavior of nanoporous aluminum extracted from molecular dynamics (MD) simulations. Reference data are generated by means of MD simulation of both hydrostatic and uniaxial deformation with compression of representative volume elements of aluminum single crystal with nanopores of spherical, cubic and cylindrical shapes at the temperatures of 300, 500, 700 and 900 K. Several typical sizes of the nanopores are considered: The smallest one corresponds to the initial porosity less than 1%, while the largest one gives the initial porosity in the range of 30–50%. Plastic collapse of nanopores in all cases occurred by the mechanism of emission of partial Shockley dislocation half-loops from the pore surface. The emission of initial loop occurred earlier in the case of cylindrical pore; however, this did not lead to an explosive increase in the number of dislocations in the system at this stage. In general, flat free surfaces of pores are less subjected to the dislocation nucleation than the rounded ones. A new physically-based micromechanical model of the plastic compaction of nanoporous metal is formulated with accounting of different pore shapes and anisotropy of the compaction process. Both tested machine-learning approaches show an adequate approximation of MD data. The developed ANN and parameterized micromechanical model are applied to simulate the propagation of a shock wave in nanoporous aluminum in comparison with direct MD simulations of this process; this comparison shows an adequate description of the shock wave structure by means of both approaches incorporated into continuum mechanics modeling. Thus, the developed machine-learning-based approaches can be applied as constitutive equations of nanoporous aluminum in macroscopic simulations of the dynamic compaction and shock-wave processes in this material.

我们比较了两种基于机器学习的方法，人工神经网络 (ANN) 和微机械模型以及模型参数的自动贝叶斯识别，用于模拟从分子动力学 (MD) 模拟中提取的纳米多孔铝的变形行为。参考数据是通过流体静力和单轴变形的 MD 模拟生成的，在 300、500、700 和 900 K 的温度下压缩具有球形、立方和圆柱形纳米孔的铝单晶的代表性体积元素。几种典型尺寸考虑纳米孔的数量：最小的对应于小于 1% 的初始孔隙率，而最大的对应于 30-50% 范围内的初始孔隙率。在所有情况下，纳米孔的塑性塌缩都是通过从孔表面发射部分肖克利位错半环的机制发生的。在圆柱孔的情况下，初始环的发射发生得较早；然而，这并没有导致现阶段系统中位错数量的爆炸性增加。一般来说，平坦的孔的自由表面比圆形的更不容易受到位错成核的影响。建立了一种新的基于物理的纳米多孔金属塑性压实的微观力学模型，其中考虑了不同的孔形状和压实过程的各向异性。两种经过测试的机器学习方法都显示出对 MD 数据的充分近似。将开发的人工神经网络和参数化微机械模型应用于模拟冲击波在纳米多孔铝中的传播，并与该过程的直接 MD 模拟进行比较；这种比较显示了通过结合到连续介质力学建模中的两种方法对冲击波结构的充分描述。因此，所开发的基于机器学习的方法可以用作纳米多孔铝的本构方程，用于对该材料的动态压实和冲击波过程进行宏观模拟。