Research & Projects

As a mechanical engineer, my research focuses on deep and machine learning methods that operate on 3D mechanical geometries and simulation data to power generative design and surrogate modeling, so that engineering design teams can explore design spaces, build prototypes, and validate product performance more effectively.

My research interests include:

Generative AI for Engineering Design

Engineering Design is changing all around us. AI/ML methods are transforming the way we look at design workflows. I'm specifically interested in exploring how we can better integrate the use of ML methods within this workflow to assist design engineers, explore design spaces under valid constraints.

Geometric Representations for ML Models

I work on ways to represent 3D mechanical parts and assemblies so they can be used effectively by learning algorithms. This includes using CAD features, meshes, graphs, and other geometric encodings that preserve how a design behaves under load and how it is manufactured. The aim is to create representations that are useful both to design engineers and to deep learning models trained on geometric data.

Data-Driven and Physics-Based Surrogate Modeling

I build surrogate models that approximate simulation results for mechanical and automotive components. These models learn from existing FEA or other simulation runs to predict quantities like stress, deflection, or mass much faster than running a full analysis. I am especially interested in combining these data‑driven surrogates with basic physics knowledge so that the predictions remain physically reasonable over the design space.

AI Foundations for Computational Design

I am interested in core methods that make AI more useful for engineering design problems. This includes combining geometry‑based models, surrogate models, and optimization so that they can be applied together in real product development workflows. My long‑term goal is to develop reliable tools that mechanical engineers can use alongside CAD and CAE to make better design and testing decisions, not to replace existing engineering practice.

PROJECTS

Here's a small description of the projects I've worked on in the past. The full report or paper written is attached on the left side.

Feature-Based Graphs Enabling CAD-ML Design Iteration

Collaborators: Satchit Ramnath (Clemson University); Ibraheem Alawadhi (Ohio State University); Jami J. Shah (Ohio State University)

Overview

Introduced Attributed Feature Graphs (AFGs), a feature-based graph representation encoding CAD design features (e.g., extrusions, ribs, pockets) as nodes with parametric attributes and geometric/dependency edges, bridging traditional CAD workflows with graph neural networks for interpretable ML in iterative engineering design on the CarHoods10K automotive hood frame dataset.

Objectives

Introduce AFGs to preserve CAD feature semantics, parametric structure, and bidirectional editability for AI/ML tasks like surrogate modeling.
Demonstrate AFG extraction and CAD regeneration for stamped hood frame features (depressions, pockets, ribs).
Train GNN-based evaluation engine to predict structural metrics (stress, mass, deflection) and compare against image/point cloud baselines.
Enable traceable predictions mapped back to editable CAD features for rapid design iteration.

Results

AFG-GNN achieved competitive test R² scores (stress 0.87, mass 0.89, deflection 0.87; overall MAPE ~9.5%) on CarHoods10K, outperforming CNN images (8-15% MAPE) and matching multimodal baselines while enabling direct CAD re-entry for feature edits. 93.6% of predictions within 20% of FEA ground truth, with mass near-perfect (98.7% within 20%), confirming AFGs' value for interpretable, CAD-native surrogate modeling.

Multimodal Machine Learning Model for Performance Prediction

Collaborators: Satchit Ramnath

Overview

Developed a multimodal machine learning framework to rapidly predict the structural performance of automotive hood frames in early design, fusing image, geometric, and parametric data from the CarHoods10K dataset to reduce reliance on time-intensive FEA while enabling faster, more informed design exploration.

Objectives

Develop a custom multimodal architecture that combines images, cross-sections, and parametric features to predict hood frame performance.
Benchmark multimodal predictions against a unimodal (image-only) baseline.
Test generalization to hood frame geometries not included in training.
Illustrate how multimodal ML can support simulation-driven conceptual design workflows.

Results

The multimodal model achieved validation errors below 20% for von Mises stress, mass, and directional deflection on CarHoods10K, substantially outperforming the image-only baseline, especially for stress and deflection. When evaluated on entirely new hood frame designs, it produced physically reasonable, trend-consistent predictions, confirming improved robustness and generalization and demonstrating its value for rapid, simulation-assisted design iteration.

Development & Analysis of B-Spline Curves

Overview

Developed C++ interactive software framework for constructing, manipulating, and visualizing B-Spline/Bezier curves, NURBS/Bezier surfaces, and knot operations (insertion/refinement), leveraging recursive basis functions, control points, and knot vectors to support CAD/CAM design automation in mechanical engineering.

Objectives

Implement B-Spline curve generation using recursive basis functions, knot vectors, and control points for accurate shape representation.
Create an interactive console interface supporting curve (B-Spline, Bezier, Cubic), surface (NURBS, Bezier), and knot operations (insertion, refinement).
Enable visualization and modification of geometric entities to explore effects of parameters like order, knots, and indices.
Address challenges in parameterization, index bounds, and Bezier as B-Spline special cases for design automation.

Results

The software successfully generated and plotted various B-Spline curves (e.g., smooth refinements post-knot insertion), Bezier curves adhering to convex hull properties, NURBS surfaces with weight modifications, and other types like rational ruled surfaces, demonstrating precise control over geometric shapes and continuity while highlighting limitations in plotting smoothness and modularity for future enhancements.

Optimization of Surface Roughness in Electrodeposition Process.

Overview

Explored advanced optimization techniques to reduce surface roughness in nickel electrodeposition through simulation and analysis. Using COMSOL Multiphysics and Taguchi methods, the study examined how current densities, time proportion, and solution conductivity affect surface quality in micromanufacturing applications.

Objectives

Minimize surface roughness and variation in deposition height.
Analyze effects of forward/reverse currents, time ratio, and conductivity.
Develop a COMSOL model to optimize electrodeposition parameters.
Improve micromanufacturing efficiency across electronics, automotive, and biomedical industries.

Results

Optimization through simulated annealing identified ideal deposition parameters, achieving a surface roughness reduction (Δy = 0.258) and fewer surface defects. The improved model demonstrated smoother finishes and better uniformity, advancing precision manufacturing through optimized electrodeposition processes.

Material Selection for Lightweight Fire Extinguisher Levers

Overview

Focused on optimizing the material selection for a fire extinguisher lever to reduce mass, enhance performance, and improve sustainability. The lever’s design ensures reliable actuation and durability under varying environmental conditions.

Objectives

Reduce lever mass without compromising strength or function.
Withstand ≥75 lb of actuation force.
Maintain a Young’s modulus ≥65 GPa.
Operate reliably between –65°F and 120°F.
Use materials that are non-corrosive and ergonomic.

Results

After evaluating Titanium Alloys, Silicon Nitride, and CFRP, the study identified 409 Stainless Steel as the most efficient and sustainable material. It offers the best balance of strength, cost, manufacturability, and recyclability, meeting performance goals while supporting eco-friendly design principles. Future research will explore lighter alternatives with comparable durability.

Optimizing the Performance of a 3D-Printed Brake Pedal using Design of Experiments (DOE) and ASTM Standard Testing Methods.

Overview

Explored 3D printing applications in automotive design by optimizing a brake pedal for strength, durability, and performance. The project examined how printing parameters and materials influence the mechanical behavior of 3D-printed components.

Objectives

Analyze tensile, flexural, and impact strength using DOE and ASTM standards.
Evaluate the effects of infill pattern, raster angle, layer thickness, material type, and infill density.
Develop an optimized brake pedal design through generative design and reverse engineering.

Results

The optimized 3D-printed brake pedal showed superior mechanical performance, with infill density and material selection emerging as the most influential factors. Adjusting raster angle and print orientation further enhanced the strength and durability. The final design demonstrates significant potential for advanced braking systems and future additive manufacturing innovations in automotive engineering.

Page updated

Google Sites

Report abuse