The aim of this study is to replace the stress-strain constitutive relation within a finite element method (FEM) implementation with an adaptive neuro-fuzzy inference system (ANFIS). The computation is supported by the Compute Unified Device Architecture (CUDA) technology available on modern Nvidia graphics processing units (GPUs). The multi-threaded multiple input / multiple output (MIMO) CUDA-ANFIS model is compared with experimental results of torsional shear (TS) tests. The cylindrical non-cohesive soil samples were tested within the small strain range (0.001–0.1%). Particular attention was paid to a good reflection of the results of laboratory tests in numerical simulations based on FEM with embedded ANFIS. The obtained results confirm the possibility of replacing explicit constitutive relations with a neuro-fuzzy inference system and using this model to determine the soil mechanical properties in engineering practice. The most important finding is that the decrease in the discretization mesh density as well as the decrease in the number of linguistic labels result in better overall system performance. In addition, the created ANFIS algorithm adapted to the GK110 graphics processor architecture speeds up calculations 27 times compared to single-threaded implementation. The presented approach can be used to construct implicit constitutive models based on the numerical interpretation of the results obtained from laboratory and/or field tests.

这项研究的目的是用自适应神经模糊推理系统（ANFIS）替代有限元方法（FEM）实施中的应力-应变本构关系。现代Nvidia图形处理单元（GPU）上可用的计算统一设备体系结构（CUDA）技术支持该计算。将多线程多输入/多输出（MIMO）CUDA-ANFIS模型与扭剪（TS）测试的实验结果进行了比较。在小应变范围（0.001–0.1％）内测试了圆柱形非粘性土壤样品。在基于嵌入式ANFIS的FEM数值模拟中，要特别注意实验室测试结果的良好反映。获得的结果证实了用神经模糊推理系统代替显式本构关系并在工程实践中使用该模型确定土壤力学性能的可能性。最重要的发现是离散化网格密度的减少以及语言标签数量的减少会导致更好的整体系统性能。此外，与单线程实现相比，适用于GK110图形处理器体系结构的ANFIS算法可将计算速度提高27倍。可以基于从实验室和/或现场测试获得的结果的数值解释，使用所提出的方法来构造隐式本构模型。最重要的发现是离散化网格密度的减少以及语言标签数量的减少会导致更好的整体系统性能。此外，与单线程实现相比，适用于GK110图形处理器体系结构的ANFIS算法可将计算速度提高27倍。可以基于从实验室和/或现场测试获得的结果的数值解释，使用所提出的方法来构造隐式本构模型。最重要的发现是离散化网格密度的减少以及语言标签数量的减少会导致更好的整体系统性能。此外，与单线程实现相比，适用于GK110图形处理器架构的ANFIS算法可将计算速度提高27倍。可以基于从实验室和/或现场测试获得的结果的数值解释，使用所提出的方法来构造隐式本构模型。