project geneng

content/talk/2024_geneng_pan_sun_wilson_skillen/index.md

@@ -0,0 +1,27 @@
+---
+title: "Generative Modeling for Engineering Design"
+author: ["Wei Pan"]
+lastmod: 2024-01-24T22:47:03+00:00
+draft: false
+weight: 4001
+event: "Projects 2024"
+authors: ["weipan","mingfeisun","carlwilson","alexskillen"]
+summary: "Explore the future of engineering design with our project on 'Generative Modeling for Engineering Design.' This cutting-edge research combines AI and machine learning to revolutionise CAD data analysis, automating feature extraction and classification. Partnering with industry leaders Cummins, we aim to transform engineering practices, enhancing efficiency and innovation. Join us in shaping advanced AI applications in real-world engineering!"
+---
+
+Background
+Cummins, a prominent mechanical engineering organisation, manages an extensive repository of technical product data, mainly in the form of 3D Computer-Aided Design (CAD) geometry and 2D design drawings. This data, representing a significant investment, is pivotal to the company's product design specifications. However, a major challenge lies in the semi-structured to unstructured nature of these CAD models and drawings, making extracting valuable insights for product performance enhancement difficult. Traditional methods of translating these designs into structured formats are manual, labour-intensive, and error-prone. This complexity poses a significant barrier, especially in the context of applying machine learning algorithms, which necessitate large, accurately labelled datasets.
+
+Project Aims
+The overarching aim of this PhD project is to revolutionise the way Cummins interacts with and leverages its vast CAD data. The project seeks to develop an innovative AI-based tool capable of automated extraction of user-defined geometric feature dimensions from 3D CAD models. This tool aims to be component agnostic, minimising the need for manual labelling while emphasising user interaction primarily for verification purposes. The project also aims to enhance the accessibility and usability of Cummins' locked data sources, thus providing a pathway for more effective application of machine learning algorithms. A key deliverable is creating a software tool for automated CAD model labelling and classification, aiming to provide accurate estimations of labelling accuracy and thereby improve the reliability of data used in product development and performance analysis.
+
+Methodology
+The methodology of this project involves integrating several cutting-edge AI technologies. Firstly, generative modelling will be employed for the computer-aided design, mapping CAD models to appropriate labels and classifications. This approach will address the current limitations in data structuring. Secondly, the project will utilise reinforcement learning with human-in-the-loop feedback, drawing on techniques used in large language models like GPT-3. This aspect will ensure that the AI's output aligns with human expertise and preferences, facilitating intuitive and practical AI systems in engineering. Lastly, the project will explore large-scale parallel and differentiable simulation, which allows for end-to-end differentiability in training neural networks. This will enable the tool to learn directly from simulation environments, significantly enhancing the efficiency of AI training in complex engineering tasks. Together, these methodologies aim to create a transformative tool for Cummins, automating and refining the process of extracting and utilising data from CAD models for advanced engineering applications.
+
+
+***References***
+
+[1] Deterministic and Discriminative Imitation (D2-Imitation): Revisiting Adversarial Imitation for Sample Efficiency, AAAI 2022, https://ojs.aaai.org/index.php/AAAI/article/view/20813 
+[2] Trust Region Bounds for Decentralized PPO Under Non-stationarity, AAMAS 2023, https://dl.acm.org/doi/abs/10.5555/3545946.3598613 
+[3] Imitation human behaviour with diffusion models, ICLR 2023, https://openreview.net/pdf?id=Pv1GPQzRrC8
+[4] Uni[MASK]: Unified Inference in Sequential Decision Problems, NeurIPS 2022, https://proceedings.neurips.cc/paper_files/paper/2022/file/e58fa6a7b431e634e0fd125e225ad10c-Paper-Conference.pdf