Explainable Machine Learning dengan Optimasi Genetic Algorithm untuk Prediksi Porositas dan Pemetaan Parameter Proses Cold Spray

Setyobudi, Bima Bagus (2026) Explainable Machine Learning dengan Optimasi Genetic Algorithm untuk Prediksi Porositas dan Pemetaan Parameter Proses Cold Spray. Other thesis, Institut Teknologi Sepuluh Nopember.

Text 5007221221-Undergraduate_Thesis.pdf Restricted to Repository staff only Download (11MB) | Request a copy

Abstract

Cold spray merupakan teknologi pelapisan berbasis solid-state yang banyak digunakan untuk menghasilkan coating dengan degradasi termal rendah. Namun, pengendalian porositas masih menjadi tantangan karena kualitas coating dipengaruhi oleh banyak parameter proses yang saling berinteraksi secara kompleks dan nonlinier. Penelitian ini mengembangkan pendekatan explainable machine learning berbasis optimasi Genetic Algorithm untuk memprediksi porositas dan memetakan parameter proses cold spray. Dataset yang digunakan terdiri dari 287 data eksperimen yang dikompilasi dari literatur, mencakup proses cold spray berbasis single powder, multi-powder, dan high entropy alloy. Model yang dikembangkan meliputi Artificial Neural Network, Random Forest, XGBoost, serta stacking ensemble dengan variasi preprocessing, penanganan missing value, encoding data kategorikal, outlier handling, feature selection, dan optimasi hyperparameter menggunakan Genetic Algorithm. Secara keseluruhan, 129 variasi model dievaluasi berdasarkan RMSE dan R² pada data pengujian. Hasil penelitian menunjukkan bahwa model terbaik berdasarkan evaluasi testing diperoleh pada stacking ensemble dengan XGBoost sebagai base learner dan ANN sebagai meta-learner, menghasilkan Test RMSE sebesar 1,2635 dan Test R² sebesar 0,9589. Hasil robustness menunjukkan bahwa model tersebut juga memiliki kemampuan generalisasi yang stabil, ditunjukkan oleh gap training loss dan testing loss yang lebih terkendali dibandingkan model ysng lain. Stabilitas ini juga diperkuat oleh hasil Gaussian noise sensitivity, di mana performa model tetap terjaga dengan nilai R² pada kisaran 0,956–0,959 meskipun data uji diberi noise hingga 5%. Analisis SHAP mengidentifikasi gas temperature, gas pressure, dan powder feeder rate sebagai tiga parameter paling berpengaruh terhadap prediksi porositas. Hasil PDP–ICE dan windowing menunjukkan bahwa porositas rendah cenderung diperoleh pada temperatur gas tinggi, tekanan gas yang cukup tinggi, serta laju umpan serbuk yang tidak berlebihan. Dengan demikian, pendekatan yang dikembangkan tidak hanya mampu memprediksi porositas secara akurat, tetapi juga memberikan interpretasi hubungan antarparameter proses sehingga dapat digunakan sebagai dasar decision-support system untuk mempercepat screening parameter cold spray sebelum eksperimen validasi dilakukan.
=======================================================================================================================================
Cold spray is a solid-state coating technology widely used to produce coatings with minimal thermal degradation. However, controlling porosity remains a major challenge because coating quality is governed by multiple process parameters that interact in complex and nonlinear ways. This study develops an explainable machine learning approach optimized using a Genetic Algorithm to predict porosity and map the process parameters of cold spray. The dataset consists of 287 experimental data points compiled from the literature, covering single-powder, multi-powder, and high-entropy alloy cold spray processes. The developed models include Artificial Neural Network, Random Forest, XGBoost, and stacking ensemble models, combined with variations in preprocessing, missing value handling, categorical data encoding, outlier handling, feature selection, and hyperparameter optimization using a Genetic Algorithm. In total, 129 model variations were evaluated based on RMSE and R² on the testing data. The results show that the best performance was achieved by a stacking ensemble model using XGBoost as the base learner and ANN as the meta-learner, with a test RMSE of 1.2635 and a test R² of 0.9589. The robustness results indicate that the best model provides stable generalization, as shown by a more controlled gap between training loss and testing loss compared with the other models. This stability is further supported by the Gaussian noise sensitivity test, where the model maintained its performance with R² values ranging from 0.956 to 0.959, even when the testing data were subjected to noise levels of up to 5%. SHAP analysis identified gas temperature, gas pressure, and powder feeder rate as the three most influential parameters in predicting porosity. The PDP–ICE and windowing results indicate that lower porosity is generally associated with higher gas temperature, sufficiently high gas pressure, and a moderate powder feed rate. Therefore, the proposed approach not only provides accurate porosity prediction but also offers interpretable insights into the relationships among cold spray process parameters. This makes it a potential basis for a decision-support system to accelerate cold spray parameter screening before experimental validation.

Item Type:	Thesis (Other)
Uncontrolled Keywords:	Cold Spray, Porositas, Explainable Machine Learning, Genetic Algorithm, SHAP, PDP-ICE Plot, Windowing Parameter, Cold Spray, Porosity, Explainable Machine Learning, Genetic Algorithm, SHAP, PDP–ICE Plot, Windowing Parameter
Subjects:	T Technology > TJ Mechanical engineering and machinery > TJ217.6 Predictive Control
Divisions:	Faculty of Industrial Technology and Systems Engineering (INDSYS) > Mechanical Engineering > 21201-(S1) Undergraduate Thesis
Depositing User:	Bima Bagus Setyobudi
Date Deposited:	04 Aug 2026 06:25
Last Modified:	04 Aug 2026 06:25
URI:	http://repository.its.ac.id/id/eprint/143189

Actions (login required)

View Item