Predicting Tensile Test Results

Contributors - Alhagie Boye Carter Ulsmich, Patrick Mcdonalds & Ryan Carter

ScotForge is a U.S.-based employee-owned metal forging company known for producing high-quality custom open-die and rolled ring forgings. With over a century of experience, they serve critical industries such as aerospace, defense, energy, and infrastructure where precision, safety, and material reliability are non-negotiable. Every forged part undergoes rigorous testing to meet mechanical property standards, including tensile strength, ductility, and toughness.

The Problem

When a new forging order comes in, engineers at ScotForge must determine the correct heat treatment process to achieve specific mechanical properties. But there’s no quick way to predict outcomes. Instead, they manually dig through historical databases to find “similar” past orders ones that used similar material grades, sizes, and heat treatments. This process is:

• • Manual and time-consuming
• • Inconsistent, depending on the engineer’s experience
• • Limited to exact or near-exact matches

The Core Challenge: How can we automate and improve this search for similar jobs and go beyond it to predict tensile properties directly from input parameters?

Why It Matters?

In forging, every second and material decision counts.Engineers need fast, accurate answers to design and production questions like: “If I use this material and heat treat it this way, what tensile strength should I expect?”. Answering this quickly:

• • Speeds up quoting and production decisions
• • Reduces the need for trial-and-error testing
• • Unlocks value from decades of past data
• • Supports innovation without risking quality

This project aimed to build a tool that makes that prediction process instant and reliable.

Our Solution

We developed a machine learning system that predicts four key tensile properties using only heat treatment parameters and basic material/process information. Instead of searching the database manually, engineers can input process specs and receive predicted values for: Ultimate Tensile Strength (UTS) Yield Strength Elongation Reduction of Area (RA) This bypasses the need for guesswork or time consuming searches and enables smarter, data-driven decision making.

How We Built It

• • Data Preparation:
  We worked with ScotForge’s historical dataset, using Python with pandas to clean and preprocess over 20 years of tensile test records and associated process parameters.
• • Modeling:
  We tested different machine learning algorithms and found that XGBoost, a gradient boosting method, provided the best accuracy and generalization for multi-target regression.
• • Feature Engineering:
  We extracted meaningful features like steel grade, cooling method, tempering temperature, and soak time transforming raw data into model-ready inputs.
• • Evaluation:
  The model was evaluated using R² and RMSE metrics. We achieved strong predictive performance, especially on UTS and Yield Strength.
• • Deployment:
  We designed a lightweight Flask-based web interface that allows engineers to input new process parameters and get instant predictions.

This project aimed to build a tool that makes that prediction process instant and reliable.

Results

Our best model achieved:

Instead of manually searching for similar jobs, ScotForge engineers can now receive fast and reliable estimates allowing them to respond quicker, reduce uncertainty, and leverage the full potential of their historical data.