Conquering Generalization Challenges in Al-enhanced Computational Mechanics: Thoughts and Practices
Prof Xu Guo
Dalian University of Technology; Ningbo Institute of Dalian University of Technology
Abstract: Artificial Intelligence (AI) has become an active research frontier in computational mechanics, a field characterized by complex physical phenomena and diverse engineering scenarios. Conventional end-to-end AI models often demonstrate strong performance on specific datasets; however, they tend to suffer a marked decrease in generalization capability when confronted with unfamiliar boundary conditions, material properties, or geometric configurations. To overcome this limitation, a problem-independent machine learning (PIML) framework has been developed to enhance large-scale structural analysis and topology optimization. The core idea centers on revisiting the foundation of the finite element method—the shape function. Specifically, machine learning is employed to construct an implicit mapping between the material distribution within coarse mesh elements and their corresponding numerical Green’s functions. As a result, the proposed PIML method is genuinely independent of particular analysis or optimization settings, since the numerical shape functions of coarse mesh elements are uniquely defined by their internal material layout, independent of external loads, boundary conditions, or the design domain’s geometry and topology. Numerical experiments confirm that the algorithm improves optimization efficiency by two orders of magnitude for topology optimization problems involving millions of degrees of freedom in three dimensions, compared to mainstream commercial software under equal computational resource constraints. In a parallel computing environment using 6750 cores, each iteration of a 3D topology optimization with 10 billion degrees of freedom requires only 42 seconds. Looking forward, this approach lays the groundwork for a next-generation universal CAE software platform that integrates AI with classical numerical methods, paving the way for more efficient and intelligent engineering simulation and design.
Bio: Professor Xu Guo, a member of the Chinese Academy of Sciences, is from the Dalian University of Technology, P.R. China. He once served as the Vice President of the Chinese Society of Theoretical and Applied Mechanics and the President of the Chinese Association of Computational Mechanics. Currently, he is the Vice President of International Society for Structural and Multidisciplinary Optimization, and one of the editorial board members of Computer Methods in Applied Mechanics and Engineering and International Journal for Numerical Methods in Engineering.
Xu Guo has been working in the field of computational mechanics, solid mechanics and structural optimization for more than 25 years. He has published more than 270 SCI papers in renowned scientific journals including JMPS, CMAME, PRL, Nature Material and Science. He is the recipient of numerous academic awards and honors, including the ASSMO Award, ICACM Award, etc. He is also the Plenary/Semi-Plenary speaker in numerous prestigious international conferences/workshops/symposiums, including the 2016 World Congress on Computational Mechanics (WCCM).
