(241d) A Framework for Global Sensitivity Analysis Via Machine Learning for Dimensionality Reduction

Large-scale, nonlinear complex models can be challenging to use to determine the effect of parameter uncertainty on model outputs. Global Sensitivity Analysis (GSA) is a well-established tool used for the identification and quantification of parametric uncertainty and its impact on the model outputs [1]. Variance-based decomposition using Sobol' indices is a commonly employed GSA method [2, 3]; however, the performance of GSA can be impacted by the dimensionality and nonlinearity of the model at hand. Researchers have developed and used different approaches for grouping parameters to execute smaller-scale GSA analyses [4] as a means to overcome challenges arising from the complexities associated with the full-scale model formulation. By reducing the dimensionality of high-fidelity models, the computational effort required for parameter analysis can be significantly decreased.

In this context, we present a framework that harnesses the advantages of classification algorithms to identify non-critical parameters and remove them from the set considered for GSA. Specifically, the presented framework comprises the following steps: (1) generation of the minimum necessary synthetic dataset within the ranges of interest, (2) use of ML-aided classification algorithms, such as Support Vector Machines (SVM) and Random Forest (RF), to assess the points generated in step (1) and reduce the parametric set to only include the parameters that have a significant impact on the model outputs to generate a reduced-order model, and (3) variance-based GSA for the quantification of the impact that uncertain parameters have on the model outputs.

The framework is tested and assessed on a mixed-integer supply chain model used to optimise the Chimeric Antigen Receptor (CAR) T cell supply chain [5, 6]. We assess the capabilities of SVM and RF as classification algorithms to reduce the dimensionality of the parametric set. The results provide an overview of the level of information retained after step (2) and a quantitative analysis of the percentage of critical relationships eliminated as a result of the dimensionality reduction. We also compare the level of information obtained from step (2) as opposed to Sobol’ GSA and assess this with respect to the computational time required in each case.

References

[1] E. Borgonovo and E. Plischke, “Sensitivity analysis: A review of recent advances,” European Journal of Operational Research, vol. 248, no. 3. Elsevier B.V., pp. 869–887, Feb. 01, 2016. doi: 10.1016/j.ejor.2015.06.032.

[2] H. Ma, E. P. Li, A. C. Cangellaris, and X. Chen, “Support Vector Regression-Based Active Subspace (SVR-AS) Modeling of High-Speed Links for Fast and Accurate Sensitivity Analysis,” IEEE Access, vol. 8, pp. 74339–74348, 2020, doi: 10.1109/ACCESS.2020.2988088.

[3] W. H. Jung and A. A. Taflanidis, “Efficient global sensitivity analysis for high-dimensional outputs combining data-driven probability models and dimensionality reduction,” Reliab Eng Syst Saf, vol. 231, Mar. 2023, doi: 10.1016/j.ress.2022.108805.

[4] A. Kiparissides, S. S. Kucherenko, A. Mantalaris, and E. N. Pistikopoulos, “Global sensitivity analysis challenges in biological systems modeling,” Ind Eng Chem Res, vol. 48, no. 15, pp. 7168–7180, Aug. 2009, doi: 10.1021/ie900139x.

[5] A. Bernardi, M. Sarkis, N. Triantafyllou, M. Lakelin, N. Shah, and M. M. Papathanasiou, “Assessment of intermediate storage and distribution nodes in personalised medicine,” Comput Chem Eng, vol. 157, p. 107582, Jan. 2022, doi: 10.1016/J.COMPCHEMENG.2021.107582.

[6] N. Triantafyllou, A. Bernardi, M. Lakelin, N. Shah, and M. M. Papathanasiou, “A digital platform for the design of patient-centric supply chains,” Sci Rep, vol. 12, no. 1, Dec. 2022, doi: 10.1038/s41598-022-21290-5.