The MICDE PhD Student Seminar Series showcases the research of students in the Ph.D. in Scientific Computing. These events are open to the public, but we request that all who plan to attend register in advance.

If you have any questions, please email [email protected].

Accelerating Fock exact exchange calculations using Tucker Tensor techniques

Density Functional Theory (DFT) is widely used to predict the electronic structure and properties of a broad range of materials. Although exact in theory, DFT simulations rely on exchange-correlation (Exc) functionals that are approximated in practice. The accuracy of DFT calculations is solely dependent on the accuracy of the Exc functionals. Hybrid exchange-correlation functionals are a class of functionals that have been shown to match experimental observations more closely compared to other Exc functionals. However, the use of hybrid Exc functionals necessitates the computation of Fock exact exchange, significantly increasing the computational cost. Furthermore, the nature of Fock exact exchange demands a substantial increase in memory requirements and communication across processors. The latter is a serious issue as it affects the scalability of the code, restricting routine simulations to a few tens of atoms. In this work, we have developed a Tucker Tensor-based approach that significantly reduces the computational cost of Fock exact exchange calculations. We have incorporated an innovative communication pattern that reduces communication without significantly increasing peak memory usage. Consequently, we have developed a robust, efficient, and scalable algorithm that achieves an order-of-magnitude speedup over the current state of the art.

Vishal Subramanian (Materials Science & Engineering and Scientific Computing)

Vishal Subramanian is a PhD candidate in the Materials Science and Engineering department. He is interested in harnessing the power of linear algebra and high-performance computing to develop robust, and efficient algorithms that can compute material properties accurately. His work with Prof. Gavini’s group developing algorithms and scalable implementations for fast density functional theory (DFT) calculations on large-scale systems earned him the 2023 Gordon Bell Prize – the highest honor given in high-performance computing.

Topology Optimization for Die Casting with Nonplanar Parting Surfaces

This talk presents a density-based topology optimization method for the simultaneous design of die-castable geometry, die drawing directions, and arbitrarily nonplanar parting surface. Viewing a die casted part as a two-component system consisting of the cavities of die halves, an arbitrarily nonplanar parting surface is represented as the boundaries between adjacent partitioned domains similar to the joints in multi-component topology optimization (MTO). The draw direction of each die half is represented as a probability distribution to avoid premature convergence, and the undercut of a part geometry in the draw direction is evaluated using the gradient of the density field. Several numerical examples are presented to demonstrate the advantages of the inclusion of nonplanar parting surfaces as optimization variables.

Heting Fu (Mechanical Engineering and Scientific Computing)

Heting Fu is a Ph.D. candidate under the guidance of Professor Kazuhiro Saitou in Mechanical Engineering. His research involves multi-component, multi-material, and multi-process topology optimization.