Nidhi Mishra

Background and Motivation
The design and engineering of components for robotic systems in the supplier industry remain labor-intensive and heavily dependent on domain expertise and iterative prototyping. As customer requirements grow increasingly complex and time-to-market pressures intensify, there is a critical need for data-driven approaches that can accelerate the design cycle while maintaining engineering rigor.

Research Objectives
Recent advances in foundation models have demonstrated remarkable capabilities in learning from heterogeneous data sources, presenting new opportunities for automated engineering design workflows. This master's thesis investigates the application of foundation models as autonomous agents for automated component design. The research aims to:

- Develop a unified machine learning architecture capable of ingesting heterogeneous data sources (technical datasheets, experimental test results, simulation outputs)

- Learn implicit design principles from historical engineering data

- Generate novel, performance-optimized component geometries

- Evaluate the effectiveness of AI-generated designs against human expert baselines

Methodology
The thesis will explore a multi-modal foundation model architecture with the following key component

Data Integration Layer

- Encoder networks for processing structured data (CAD parameters, material properties)

- Processing of unstructured text (technical documentation)

- Handling of numerical sequences (FEM/CFD simulation results, experimental performance metrics)

Cross-Modal Learning

- Transformer blocks enabling the model to learn correlations between design parameters, physical constraints, and performance outcomes across modalities

- Self-supervised pre-training on unlabeled historical design data

Generative Design Agent

- Decoder architecture (e.g., conditional VAE or diffusion model approaches) to generate parametric design candidates conditioned on requirement specifications

- Physics-informed constraints integrated into the loss function to ensure generated designs respect manufacturability and operational limits

Optimization and Refinement

- Agent-based optimization, where the model iteratively refines designs based on simulated performance feedback

- Multi-objective optimization considering performance, manufacturability, and cost constraints

Expected Contributions

Scientific Contributions:

- Novel multi-modal architecture for engineering design tasks integrating heterogeneous technical data

- Demonstration of transfer learning from historical designs to novel application scenarios

- Analysis of interpretability mechanisms to extract actionable design insights from learned representations

Practical Contributions:

- Quantitative evaluation framework comparing model-generated designs against human expert baselines

- Proof-of-concept implementation demonstrating feasibility in industrial robotics component design

- Guidelines for integrating foundation models into existing engineering workflows

Requirements

- Strong background in machine learning

- Programming proficiency

- Familiarity with CAD systems and engineering design principles (beneficial but not mandatory)

- Interest in bridging AI research with industrial applications

Literature