de Oliveira Rafael Gaspar Bessa, Yudi Jones, da Silva Edson Paulo, Bestard Guillermo Alvarez, Ribeiro Junior Luiz Antonio, Silva Alysson Martins Almeida
College of Technology, Department of Mechanical Engineering, University of Brasília, Federal District, Brasília 70910-900, Brazil.
Institute of Physics, University of Brasília, Federal District, Brasília 70910-900, Brazil.
ACS Omega. 2025 Jul 16;10(29):32246-32256. doi: 10.1021/acsomega.5c04133. eCollection 2025 Jul 29.
Freeze casting is a versatile manufacturing process for producing porous materials with tailored microstructures and properties. However, due to the complexity and variability involved, predicting porosity based on process parameters remains a challenging task. Accurate prediction is crucial for designing materials as porosity significantly impacts their applications. This study applies machine learning techniques, including CatBoost, Random Forest, and XGBoost, to predict porosity using experimental data from 252 research papers covering ceramics, polymers, and composites. The CatBoost model demonstrated the best predictive performance with an of 0.81 on the test set. Shapley Additive Explanations (SHAP) analysis revealed that solid loading had the most significant influence on predictions, with lower loading leading to increased porosity, as expected theoretically. The results highlight the potential of explainable machine learning to guide experimental design and optimize porosity in freeze casting materials.
冷冻铸造是一种用于生产具有定制微观结构和性能的多孔材料的通用制造工艺。然而，由于所涉及的复杂性和变异性，基于工艺参数预测孔隙率仍然是一项具有挑战性的任务。准确的预测对于材料设计至关重要，因为孔隙率会显著影响其应用。本研究应用机器学习技术，包括CatBoost、随机森林和XGBoost，利用来自252篇涵盖陶瓷、聚合物和复合材料的研究论文的实验数据来预测孔隙率。CatBoost模型在测试集上表现出最佳预测性能， 为0.81。夏普利加法解释（SHAP）分析表明，固相含量对预测影响最大，正如理论预期的那样，较低的含量会导致孔隙率增加。结果突出了可解释机器学习在指导实验设计和优化冷冻铸造材料孔隙率方面的潜力。
