Analyzing Temperature-Dependent Thermal Properties of Composite Materials Using Machine Learning Methods
I tiakina i:
| I whakaputaina i: | ProQuest Dissertations and Theses (2025) |
|---|---|
| Kaituhi matua: | |
| I whakaputaina: |
ProQuest Dissertations & Theses
|
| Ngā marau: | |
| Urunga tuihono: | Citation/Abstract Full Text - PDF |
| Ngā Tūtohu: |
Kāore He Tūtohu, Me noho koe te mea tuatahi ki te tūtohu i tēnei pūkete!
|
| Whakarāpopotonga: | This research focuses on the prediction of material properties for Ti-based biomaterials using machine learning methods. To begin, a web scraper was employed to collect data from relevant articles, and Plot Digitizer was utilized to extract the data, which was then exported to an Excel file for further analysis. After obtaining the data, various machine learning techniques, including Artificial Neural Networks (ANN), Random Forest, and Decision Trees, were applied to predict key material properties such as Young's modulus, density, thermal conductivity, and specific heat at different temperatures. The predicted values were validated against literature data to assess the accuracy of the models.Additionally, the Rule of Mixtures was employed to estimate composite material properties by combining the individual properties of the constituent materials (such as Ti, Al, and Cu) based on their volume fractions. This rule provides a theoretical estimate of the overall material property by assuming a linear relationship between the properties of the components and their respective volume fractions. The rule of mixtures was used as an additional benchmark to evaluate the accuracy of the predicted results for composite materials like TiAl, TiCu, and TiO2, aiding in the interpretation and comparison of the machine learning model outputs.The initial phase involved the creation and validation of predictive code, ensuring accuracy by comparing model predictions with actual data at specified temperatures. This validation was crucial for establishing the reliability of the models before advancing to the development phase.In model development, the ANN model showed high accuracy for TiAl with an R² score of 0.980874, effectively predicting density and Young’s modulus. Minor deviations in band gap and energy absorption suggest potential areas for improvement. For TiCu, the ANN model achieved an impressive R² score of 0.997607, with small deviations in band gap and specific heat, indicating strong performance but also areas for refinement. The ANN model for TiO2 demonstrated strong prediction capabilities but exhibited significant deviations in band gap and specific heat, highlighting the need for further adjustments to better capture temperature-dependent behaviors.The Random Forest model also demonstrated strong performance, with an R² score of 0.998168 for TiAl and a low Mean Squared Error (MSE) of 0.0054, though it encountered challenges with density and band gap predictions. For TiCu, the model showed accurate predictions but had notable deviations in band gap and density. The Random Forest model for TiO2 excelled in predicting band gap with an R² score of 0.9994 but faced issues with density predictions at higher temperatures.The Decision Tree model provided valuable insights with an R² score of 0.993841 for TiAl, accurately predicting specific heat and Young’s modulus. For TiCu, it predicted trends consistent with thermal expansion but had deviations in band gap and thermal conductivity. The model for TiO2 accurately predicted band gap trends but showed problematic decreases in density, indicating potential model fitting issues. Overall, the study demonstrates that while all three models effectively predict material properties, each has specific strengths and areas for improvement. The ANN model is effective in capturing non-linear relationships, the Random Forest model offers robust predictions, and the Decision Tree model provides clear hierarchical insights.The predicted values were validated against both literature data and experimental results to assess the accuracy and reliability of the machine learning models. For validation, experimental results were sourced from peer-reviewed literature and data repositories, which provided a robust benchmark for evaluating the performance of the models. The validation process involved comparing the predictions from the machine learning models (ANN, Random Forest, and Decision Tree) with actual experimental data collected at specified temperatures. These comparisons helped establish the models' predictive power and identify potential areas for refinement.The next phase involves expanding the database through innovative data acquisition methods, including the development of an advanced web scraping tool. This tool automates the process of searching, downloading, and extracting relevant scientific papers from repositories like arXiv, which are then processed to extract material-specific plots and numerical data. To improve data handling, Python scripts are utilized to convert PDF files into images, extract graphs, and convert the extracted visual data into structured CSV files for analysis. These advancements aim to create a more comprehensive and reliable predictive model, facilitating a broader understanding of material properties across various conditions.This study addresses challenges in comparing web-scraped data with existing material property datasets, particularly focusing on inconsistencies and errors in publicly available resources. The Berkeley dataset, chosen as a benchmark, was found to contain significant gaps and physically implausible values, such as negative entries for properties like specific heat and band gaps. To improve reliability, a thorough data cleaning process was applied, involving the handling of missing values, correcting outliers, and standardizing units. The cleaned dataset, though reduced in size, provided a more reliable basis for comparison with web-scraped data. A Random Forest model was employed to evaluate the accuracy of both datasets, revealing systematic differences in predicted material properties. These findings highlight the importance of data quality and careful validation in material science research, emphasizing the need for reliable datasets for accurate predictive modeling.This study expands the evaluation of composite material properties, incorporating Electrical Conductivity (S/cm) alongside traditional properties like density, thermal conductivity, specific heat, and Young’s modulus. Machine learning models, including Artificial Neural Networks (ANN), Random Forest (RF), and Decision Trees (DT), are applied to predict material behavior for Titanium (Ti), Aluminum (Al), and Copper (Cu) across a temperature range of 298.15K to 618.15K, with a 4K increment. New properties such as Bulk Modulus, Shear Modulus, Thermal Diffusivity, Electrical Resistivity, and Thermal Stress are calculated based on material parameters sourced using a web scraper. The study aims to provide a comprehensive understanding of the materials' performance, with the models ensuring reliable and accurate predictions. This approach supports material selection and design by evaluating the mechanical, thermal, and electrical behavior under varying conditions. This research contributes to advancing material property prediction by employing Artificial Neural Network (ANN), Random Forest, and Decision Tree models to accurately forecast the thermal and mechanical properties of TiAl, TiCu, and TiO2 across varying temperatures. The improved predictive capability reduces the need for extensive experimental testing, enhancing material selection and design processes. These insights support the development of durable and stable biomedical implants, surgical instruments, and antimicrobial coatings, improving performance in clinical applications. Additionally, research aids industries such as aerospace, automotive, and electronics by providing data-driven solutions for enhanced thermal management. This comprehensive approach ensures improved material performance under diverse operating conditions. |
|---|---|
| ISBN: | 9798293829309 |
| Puna: | ProQuest Dissertations & Theses Global |