Comparison of feature importance measures and variance-based indices for sensitivity analysis: case study of radioactive waste disposal flow and transport model.
Saved in:
| Title: | Comparison of feature importance measures and variance-based indices for sensitivity analysis: case study of radioactive waste disposal flow and transport model. |
|---|---|
| Authors: | Svitelman, Valentina1 (AUTHOR) svitelman@ibrae.ac.ru, Saveleva, Elena1 (AUTHOR), Neuvazhaev, Georgy1 (AUTHOR) |
| Source: | Stochastic Environmental Research & Risk Assessment. Oct2025, Vol. 39 Issue 10, p4827-4847. 21p. |
| Subjects: | Sensitivity analysis, Radioactive wastes, Groundwater flow, Machine learning, Reduced-order models |
| Abstract: | Sensitivity analysis is a crucial step in the development of computational models for any complex system, as it allows for comparison of the relative influence of model parameters on the simulation results. Its role is even more critical in the context of numerical safety assessment for future geological repositories of radioactive waste since it provides insights into understanding relevant to safety processes, establishes grounds for prioritizing additional research, enhances confidence in the safety assessment results. Unfortunately, groundwater flow and radionuclide transport models for radioactive waste repositories often become quite detailed and computationally expensive during the iterative process of the safety assessment. Moreover, any meaningful statistical analysis of the simulation results requires hundreds, if not thousands, of model realizations at different parameter combinations. This necessity has led to the increasing popularity of sensitivity analysis methods coupled with metamodeling approaches where a portion of the points in the parametric space (model realizations) is obtained through numerical simulation of processes, and another portion is approximated from available realizations using less computationally demanding data-driven algorithms. Modern machine learning methods also frequently address the problem of predicting a response function at new points using data from known points. Feature importance measures for machine learning models serve a role akin to sensitivity analysis: they assist in quantifying the effect of model inputs (features) on outputs (predictions). In this paper, we compare the modern first-choice variant of sensitivity analysis, variance-based Sobol' indices, obtained using polynomial chaos expansion metamodel, with well-known importance measures from the machine learning world. [ABSTRACT FROM AUTHOR] |
| Copyright of Stochastic Environmental Research & Risk Assessment is the property of Springer Nature and its content may not be copied or emailed to multiple sites without the copyright holder's express written permission. Additionally, content may not be used with any artificial intelligence tools or machine learning technologies. However, users may print, download, or email articles for individual use. This abstract may be abridged. No warranty is given about the accuracy of the copy. Users should refer to the original published version of the material for the full abstract. (Copyright applies to all Abstracts.) | |
| Database: | Engineering Source |
|
Full text is not displayed to guests.
Login for full access.
|
|
| Abstract: | Sensitivity analysis is a crucial step in the development of computational models for any complex system, as it allows for comparison of the relative influence of model parameters on the simulation results. Its role is even more critical in the context of numerical safety assessment for future geological repositories of radioactive waste since it provides insights into understanding relevant to safety processes, establishes grounds for prioritizing additional research, enhances confidence in the safety assessment results. Unfortunately, groundwater flow and radionuclide transport models for radioactive waste repositories often become quite detailed and computationally expensive during the iterative process of the safety assessment. Moreover, any meaningful statistical analysis of the simulation results requires hundreds, if not thousands, of model realizations at different parameter combinations. This necessity has led to the increasing popularity of sensitivity analysis methods coupled with metamodeling approaches where a portion of the points in the parametric space (model realizations) is obtained through numerical simulation of processes, and another portion is approximated from available realizations using less computationally demanding data-driven algorithms. Modern machine learning methods also frequently address the problem of predicting a response function at new points using data from known points. Feature importance measures for machine learning models serve a role akin to sensitivity analysis: they assist in quantifying the effect of model inputs (features) on outputs (predictions). In this paper, we compare the modern first-choice variant of sensitivity analysis, variance-based Sobol' indices, obtained using polynomial chaos expansion metamodel, with well-known importance measures from the machine learning world. [ABSTRACT FROM AUTHOR] |
|---|---|
| ISSN: | 14363240 |
| DOI: | 10.1007/s00477-024-02869-y |
