Supervised machine learning models for accurate prediction of micro pump displacement
Article : Articles dans des revues internationales ou nationales avec comité de lecture
This study presents a supervised learning framework for predicting the maximum membrane displacement
of a shape memory alloy (SMA) antagonistic micro pump subjected to material and geometric
uncertainties. SMA micro pumps are indeed promising candidates for precise fluid control in
microfluidics, but their nonlinear and hysteretic thermomechanical behavior makes parametric
analyses and uncertainty evaluation using finite element analysis (FEA) methods particularly
computationally expensive.
To overcome this difficulty, a high-fidelity dataset comprising 500 design points was generated
by three-dimensional simulation using a uniform randomized experimental design. Eight input
parameters were investigated: five properties of the NiTi material (Young’s modulus E, Poisson’s
ratio ν , hardening parameter h, yield strength R, and martensite end temperature Mf) and three
geometric dimensions (thickness emm, membrane radius Rmm and spacer length Lmm). After
appropriate preprocessing, five regression algorithms XGBoost, Gradient Boosting, LightGBM,
Random Forest and Support Vector Regression were trained and optimized using RandomizedSearchCV
and 5-fold cross-validation. Among these, XGBoost offered the best predictive performance, with
R2 = 0.9547, RMSE = 0.0444, MSE = 0.0020, and MAE = 0.0323 across the entire test
set. SHAP analysis showed that membrane thickness and radius are the most influential factors
corroborating the physical consistency of the learned correlations. Learning curves and residual
diagnostics confirmed the robustness of the XGBoost model.
The proposed surrogate model significantly reduces computation time relative to
direct finite element methods based uncertainty quantification methods or reliability
based design optimization while maintaining high predictive accuracy. These results
demonstrate that machine learning based surrogate modeling constitutes an efficient
and reliable complement to FEA for the design of SMA micro pumps in advanced
microsystems.