To start, would you mind introducing yourself, what you study, and the main question your current research focuses on?

My name is Eunjeong Park, and I am a graduate student at Penn State studying mechanical engineering, with a focus on materials science, advanced manufacturing, and artificial intelligence.

My research mainly explores how we can use AI to better understand and predict the behavior of materials. In particular, I am interested in whether AI models can predict phase formation, microstructure, and material properties from information such as chemical composition, temperature, and manufacturing conditions. The larger goal is to reduce the time and cost required for experiments and physics-based simulations while still producing scientifically meaningful predictions.

What first made you interested in materials science and mechanical engineering?

I became interested in materials science because the properties of a material can change dramatically depending on its composition, processing conditions, and internal structure. Even when two components look similar from the outside, their performance can be completely different because of what is happening at the microscopic level.

I found that connection between processing, microstructure, and properties very interesting. Mechanical engineering also gave me a way to connect fundamental materials knowledge to practical problems, such as manufacturing stronger, lighter, or more reliable components.

How did artificial intelligence become part of your interest in materials and manufacturing?

Artificial intelligence became part of my research because traditional materials research can be very computationally and experimentally expensive. For example, investigating every possible alloy composition or manufacturing condition through experiments would take a tremendous amount of time and resources.

AI models can learn patterns from existing simulation or experimental data and then make predictions for new material systems. I became interested in using AI not as a replacement for materials science, but as a tool that works together with physical knowledge. My goal is to develop models that are both computationally efficient and scientifically reasonable.

You currently mentor a high school student in microstructural image processing. What does that project involve, and how have you grown through being a mentor?

I am currently working with a high school junior on a project that combines microstructure simulation, image processing, and machine learning.

First, we generate phase-field-like microstructural data using Monte Carlo simulations. We then process the simulated images to extract quantitative information, such as phase fraction, grain size, grain boundaries, and the spatial distribution of different materials. Using these microstructural descriptors, we are developing a machine-learning framework that predicts the effective properties of composite materials depending on their material composition and microstructure.

The student has followed the project very well, and we are now approaching the final stage. This has been especially meaningful to me because it is my first time formally guiding someone through a research project. Mentoring has taught me how important it is to explain technical ideas clearly and to divide a complicated research problem into manageable steps. It has also been rewarding to watch the student become more confident and independent, and I am excited to see how they continue to grow.

Could you walk us through your recent phase-diagram project, from the original problem to the model’s final predictions?

The original problem was that predicting which phases will form in a multicomponent alloy can be very difficult. As the number of elements increases, the possible composition space becomes extremely large, and conventional thermodynamic calculations can become expensive when predictions are needed for many different conditions.

For this project, I focused on multicomponent solder alloys. We first generated a dataset containing alloy compositions, temperatures, elemental properties, and the corresponding stable phases obtained from thermodynamic calculations.

We then converted the material information into structured text prompts and fine-tuned a large language model to predict the set of stable phases. Instead of asking the model to produce a long written explanation, we trained it to provide a specific phase assemblage for each composition and temperature.

However, the initial language-model predictions were not always complete. For example, the model sometimes omitted a phase or predicted an extra phase. To address this, I developed a second correction stage using tree-based machine-learning models, including Random Forest and XGBoost. The corrector evaluates the initial phase set and decides whether a phase should be added, removed, or retained.

The final framework therefore combines the general pattern-recognition ability of a large language model with the more structured decision-making ability of tree-based models. We evaluate the model not only on alloy systems similar to the training data, but also on more complex quaternary and quinary systems that represent out-of-distribution conditions.

What does a normal work session on this project look like, from opening the data or code to producing a result you can evaluate?

A normal work session usually begins with checking the dataset and confirming which experimental or computational condition I want to study. I may examine the distribution of alloy compositions, temperatures, and phase labels to make sure that the training and test datasets are properly separated.

Next, I preprocess the data and convert each material condition into the input format required by the model. Depending on the experiment, I may train a new model, change a prompt format, adjust the model parameters, or test a different correction strategy.

After training, I run the model on validation and test datasets and compare its predictions with the thermodynamic ground truth. I calculate metrics such as exact phase-set accuracy and intersection-over-union, which measures how much the predicted and true phase sets overlap.

I also examine individual errors rather than relying only on the overall accuracy. For example, I check whether the model predicted the wrong phase name, missed one phase, or added an unnecessary phase. That error analysis usually determines what I work on next.

What have been the biggest challenges with creating your AI models, and how do you troubleshoot and fix those problems?

One of the biggest challenges is deciding what to do when the results are different from what I expected. Sometimes a model that I expected to improve performance produces worse results, and it can initially feel overwhelming because there are many possible causes.

At the same time, AI is developing extremely quickly, and many new models and methods are constantly being introduced. I often get ideas by reading recent papers and studying how researchers in different fields solve similar problems.

However, another challenge is that it is easy to combine too many ideas at once. When too many models, loss functions, features, or correction methods are added together, the project can lose its original direction. Making efficient decisions about which ideas are actually necessary is very difficult, and honestly, it is something I am still learning.

When troubleshooting, I try to return to the simplest version of the model and change one component at a time. I compare the results carefully, examine individual prediction errors, and determine whether each additional component provides a measurable benefit.

What is one realistic step a high school student could take this month to explore materials science or artificial intelligence for engineering?

I believe that, as with any field, understanding the domain itself should come before trying to apply artificial intelligence to it. To use AI meaningfully in materials science, a student should first learn some basic concepts about materials, such as how composition, processing, microstructure, and properties are connected.

A realistic first step would be to choose one materials-related topic that interests them, such as batteries, metals, semiconductors, composites, or additive manufacturing, and spend a few weeks learning the basic science behind it. After that, they could look for examples of how AI is being applied to that particular area.

Many technology companies, especially companies such as NVIDIA, are currently investing in domain-specific AI for science and engineering. Following recent research projects, company publications, and technology trends can help students understand what kinds of problems researchers are trying to solve. They do not need to begin by building a complicated model. Learning the scientific background and developing the habit of following current research trends would already be a strong first step.