AP08856141 “Development of a topological optimization method based on Deep Learning and GPU-accelerated computations for creating aerodynamic structures”
Relevance
Recent advances in additive manufacturing are now opening up new design opportunities in many industries, including aerospace and robotics, where multi-component designs developed using old-fashioned manufacturing processes can be easily replaced with lightweight one-component designs using cost-effective single-stage 3D printing technology.
Today's conventional topological optimization and generative design methods used to develop design solutions for additive manufacturing are mostly non-reversible in terms of end product application. Given the design space and constraints, these methods either optimize the material layout to minimize structural compliance (increase in stiffness) by applying rigid body mechanics, or improve the heat exchanger by applying heat transfer. Moreover, these methods are computationally expensive and require long iterations. Thus, the proposed topological optimization based on adjoint aerodynamics-rigid body modeling (ASTO) will solve the topology problem: how to distribute the material within the design area in such a way that the final structure has an optimally reduced mass / volume, increased rigidity, aerodynamically improved shape while observing the prescribed restrictions.
Objective of the project
Development of a new TO method that will include accelerated GPU and Deep Learning methods to develop 3D printed structures optimized for multiphysics applications, in particular aerodynamic interaction applications with rigid bodies. The new method will allow designers to iterate faster with fewer errors and explore the development space more deeply.
The expected results of the project are
- Analysis of ordinary TO algorithms from the point of view of their computational complexity. Particular attention will be paid to the time and memory requirements of TO algorithms. Such an analysis will help to understand how to integrate parallel computing methods.
- Development of accelerators based on graphic processors for ordinary TO algorithms using the hardware and software architecture of parallel computing CUDA. Graphics processing units (GPUs) have become an integral part of existing supercomputers. Due to power limitations, they will remain critical components of future systems. CUDA is a well established approach to GPU computing. GPU-based parallel computing provides an effective means of reducing the computational intensity of TO. The large number of cores in GPUs makes them ideal for massive parallelism.
- Incorporating deep learning (DL) methods, in particular Convolutional Neural Networks (ConvNet), as a data-driven approach to quickly generate accurate TO solutions without having to complete all the iteration steps. DL power will be used in TO as an efficient pixel image processing technology.
- Combining DL and GPU-accelerated computations with a traditional TO algorithm and performing validation tests with a single physics (solid mechanics) such as Messerschmitt-Belkov-Bloom (MBB) TO and cantilever beams. Four different combinations will be considered in the test tasks: simple TO, with ConvNet support, with GPU-acceleration and with ConvNet / GPU-acceleration support for benchmarking.
- Understanding the conjugation of the interaction of aerodynamics and rigid body, and further development of the single-physics TO algorithm into topological optimization based on adjoint modeling of aerodynamics-rigid body (ASTO). Validation of ASTO using various test tasks and checking the suitability for 3D printing of the resulting structures. In the current project, ASTO will deal with fluid flow, aerodynamic efficiency and rigid body mechanics, with the exception of the thermal effect. Thus, the proposed ASTO will solve the problem of topology: how to distribute the material within the design area in such a way that the final structure has an optimally reduced mass / volume, increased stiffness, aerodynamically improved shape while respecting the prescribed restrictions.
- Implementation of DL and GPU-accelerated computations in ASTO algorithm. GPU-based calculations will include aerodynamic efficiency calculation, numerical solution of fluid motion equation and rigid body mechanics. Moreover, it will also speed up the computational tasks of ConvNet. In the meantime, ConvNet will be trained to generate TO solutions, thus aiming for higher accuracy with more training data and test runs. Test problems that include interactions between aerodynamics and a rigid body will be solved using ASTO, taking into account four different combinations as in Task 4 for comparative analysis.
- Training of interdisciplinary (DL/TO/GPU-accelerated computing) young professionals for the academic/industrial sector of the Republic of Kazakhstan. Undergraduates/doctoral students will be involved in the project
Results achieved:
A literature review was carried out to study the modern results of TO/ASTO research, DL applications and GPU computing in the field of TO. The results of the classified reviews were obtained mainly in three areas: l) ASTO; 2) application of DL methods in TO; and 3) the use of GPU for maintenance tasks. A review of the literature has shown that today's conventional topological optimization and generative design methods used to develop design solutions for additive manufacturing are mostly non-reversible in terms of end product application. Given the design space and constraints, these methods either optimize the material layout to increase stiffness by applying rigid body mechanics, or improve the heat exchanger by applying heat transfer. Moreover, these methods are computationally expensive and require long iterations. Until now, the aerodynamic structure associated with TO, working with GPUs and with the help of DL, has not been studied. There are only a few papers where multiphysics TO is considered. Basic research has been carried out: analysis of conventional TO algorithms from the point of view of their computational complexity. Also test studies on (single physics) conventional TH of BCM beams and cantilever beams with frequently used geometries and load/constraint schemes. The computational complexity of conventional TO algorithms is estimated. The results of test studies on conventional maintenance of MBB and cantilever beams were obtained.
Names and surnames of the study group members with their identifiers:
- Akhmetov Bakytzhan, scientific adviser
- Rustamov Samir, Chief Researcher
- Kaltaev Aidarkhan, Chief Researcher
- Inkarbekov Medet Karkynbekovich, Senior Researcher
- Maksum Elaman Armanuly, Senior Researcher
- Bekzat Ajan, Performer
- Elmin Gasimov, Performer
- Rustem Aisariyev, Performer