AI-based optimization of vehicle crashworthiness design: Taking randomness out of design optimization

The primary goal of the project is to develop an AI-based intelligent design optimisation framework for static and dynamic optimisation problems that is interpretable, accurate, versatile, and energy-efficient. These are achieved through explainable AI surrogates, robust, novel optimisation strategies, and running algorithms on customised inference hardware optimised for accuracy and energy consumption.

• Implement an explainable and interpretable AI-based design optimisation method.
• Develop a Graph network-based FE-Surrogate-based design assistant to replace FE-Simulations for boundary value problems with rich past simulation data.
• Develop an intelligent design assistant closely incorporated into the Finite Element solver to replace computations at an element level to accelerate FE simulations for BVPs with less training data.
• Develop an AI-based design assistant with a novel reward strategy based on continuous residuals rather than discrete steps and a novel agent architecture for dealing with problems in mechanics and dynamics.
• Develop a GUI application to enable flexible coupling between different analysis programmes and proposed design assistants. Reduction of computational time and energy consumption through a customised AI inference chip, leading to a sustainable AI-based optimisation framework.
• Deploy the developed models on a customised, in-house AI inference chip built on an FPGA to achieve a plug-and-play extension to conventional workstations and a hardware-based design assistant supporting design optimisation.
• Validate proposed design assistants on various benchmark problems, dynamic problems from public data banks and test cases from collaborators in the SPP.

• Independent design assistants for FE-Surrogates, intelligent Finite Elements and Design optimisation.
• Provide benchmark problems for other collaborators.
• Support deployment on sustainable hardware.

• Provide test cases to evaluate the proposed framework and strategies further.

R. Gulakala, B. Markert, M. Stoffel, Generative adversarial network based data augmentation for CNN based detection of Covid-19. Sci Rep 12, 19186, s41598-022-23692-x, 2022. 

R. Gulakala, B. Markert, M. Stoffel (2022), Graph Neural Network enhanced Finite Element modelling. PAMM, e202200306, 2022. 

A. Borse, R. Gulakala, M. Stoffel, Machine Learning Enhanced Optimization of Crash Box Design for Crashworthiness Analysis, PAMM e202300145, 2023.

R. Gulakala, B. Markert, M. Stoffel, Generative learning-based model for the prediction of 2D stress distribution, PAMM; e202300201, 2023.

Z. Ding, K. Flaßkamp, R. Gulakala, M.K. Hoffmann, J. Mühlenhoff, T. Sattel, M. Stoffel, Data augmentation for design of concentric tube continuum robots by generative adversarial networks, PAMM; e202300278, 2023.

A. Borse, R. Gulakala, M. Stoffel, Development of a machine learning-based design optimisation method for crashworthiness analysis, Archives of Mechanics, 76, 2024.

R. Gulakala, V. Bhaskaran, M. Stoffel, Generative learning and graph-based framework for computing field variables in Finite Element Simulations, CMAME; 428, 117111. 2024

A. Borse, R. Gulakala, M. Stoffel, Multi-parameter design optimisation of crash box for crashworthiness analysis, PAMM 202400096, 2024.