Embedded Files
MINED Research Group
IMECE Poster
Winer_Dylan_IMECE_Poster.pdfAbout the Lab
The MINED research group, led by Professor Surya Kalidindi, focuses on creating a data-driven ecosystem to accelerate materials innovation, reducing the cost and time required to deploy new materials in high-performance applications. By combining high-throughput measurements, multiscale modeling, and data science, the lab studies hierarchical materials across various scales. Our goal is to develop an open-access platform that fosters cross-disciplinary collaboration to optimize material design, processing, and performance for diverse applications.
MINED Research Group
Project Objective
The objective of this currently in development project (started January 2024) is to develop an inverse mapping model for shape memory alloys (SMAs) that predicts composition and thermal processing history based on thermal properties. By leveraging advanced machine learning techniques, such as Gaussian Process Regression (GPR) and Conditional Invertible Neural Networks (INN), the project aims to optimize material design and processing parameters. The goal is to create a predictive model that can guide the development of SMAs with desired thermal characteristics, facilitating the efficient design of high-performance materials for various applications.
Impact of Work
This work has the potential to significantly accelerate the design and optimization of SMAs, which are crucial in a wide range of applications, including medical devices, aerospace, and robotics. By providing a predictive tool for SMA thermal properties, the project enhances the ability to tailor material compositions and processing conditions for specific performance needs, improving efficiency and reducing potential for failure. The impact extends to industries relying on advanced materials, enabling innovations in systems that require precise control over shape transformation under varying conditions.
My Role
As an undergraduate researcher, I am gaining academic course credit for my work in the MINED Lab. My role involves data preprocessing, feature engineering, and building machine learning models, including Gaussian Process Regression and Inverse Neural Networks, to predict material properties. This experience thus far has allowed me to apply theoretical knowledge to real-world challenges and further deepen my understanding of materials science and data-driven research.
I am working closely under my mentor, PhD candidate Michael Buzzy, and consulting with Professor Kalidindi to eventually publish this work in an academic journal.
My Contribution
Data Processing
I began in Spring 2024 by processing the NASA SMA dataset, converting the original Excel workbook into a single, consistent DataFrame. This involved aggregating individual sheets into one Pandas table, removing extraneous columns, and ensuring that each trial had filled rows with consistent formatting. I standardized the composition data, ensuring that the percentages for each element summed to 100%, and organized the thermal and mechanical properties into columns suitable for machine learning analysis.
Excel Workbook
Conversion to DataFrame
Feature Engineering
Analyzed the Density of the NASA Provided Data
For feature engineering, I identified and categorized the relevant input parameters, including element percentages, heat treatment conditions, cycle parameters, and thermal processing details. Additionally, I incorporated atomic properties and lattice structure data as new features to enhance the model's predictive capabilities. This process involved refining the data structure and ensuring all inputs were consistent and ready for use in the model, with a focus on ternary alloys as the primary dataset for model training.
Forward Model (GPR) Building
I am currently working on building a forward Gaussian Process Regression (GPR) model to predict thermal properties based on input variables like composition and thermal processing history. This model uses a multitask GPR approach with an RBF-ARD kernel, and I am iteratively fine-tuning it using stochastic gradient descent with the Adam optimizer. Parity plots are being generated to visualize model performance, with a focus on refining the accuracy of thermal property predictions through continued training and optimization.
Preliminary GPR Parity Plots
INN Building
In parallel, I am developing an Inverse Neural Network (INN) to generate new material designs based on the thermal properties predicted by the GPR model. The INN utilizes a conditional invertible architecture to enable reversible mappings between process parameters and thermal properties. I am optimizing the model by training it and adjusting the architecture to reduce the number of trainable parameters while ensuring accurate predictions for new SMA designs.
Preliminary INN Predictions
Challenges & Solutions
Data Processing
Challenge: Converting the NASA Excel spreadsheet into a unified DataFrame due to inconsistent formatting.
Solution: I converted each individual sheet of the Excel workbook into a standardized template and then compiled all these standardized templates into a unified dataframe, ensuring consistency across the data to be used for machine learning applications.
Prediction Uncertainty
Challenge: Uncertainty in the predictions made by the original neural network model for the four thermal properties.
Solution: I switched from the neural network model to Gaussian Process Regression (GPR), which allows for the quantification of prediction uncertainty, providing more reliable and interpretable results for the thermal property predictions.
Technical Language
Challenge: Learning the technical language of both materials science, specifically related to SMAs, and machine learning algorithms and data processing techniques, which I had limited experience with prior to this project.
Solution: I leveraged online articles and scientific literature to deepen my understanding of SMAs and machine learning techniques, while collaborating closely with my mentor Michael Buzzy and Professor Kalidindi, who provided guidance and expertise to bridge the knowledge gap and facilitate my learning in both fields.
Results
The project is currently in progress, and final results are not yet available. These will be added once the work is complete and the models have been fully evaluated and validated.
Shaping Future Experience
Although this research project is still ongoing, it has already reinforced my desire to pursue a PhD in mechanical engineering, with a particular interest in aerospace robotics. The computational skills I’ve developed—especially in machine learning, data processing, and modeling—have been invaluable in shaping my academic and career goals. Additionally, working on this project has provided me with a deeper understanding of the intricacies of conducting academic research, from problem formulation to data analysis and model development. These experiences have solidified my passion for research and innovation in engineering, guiding me toward further academic pursuits.
Google Sites
Report abuse