Evaluating causal‐based feature selection for fuel property prediction models
- Sandia National Laboratories Livermore California USA
- Sandia National Laboratories Livermore California USA, Department of Immunology, University of Washington Seattle Washington USA
Abstract
In‐silico screening of novel biofuel molecules based on chemical and fuel properties is a critical first step in the biofuel evaluation process due to the significant volumes of samples required for experimental testing, the destructive nature of engine tests, and the costs associated with bench‐scale synthesis of novel fuels. Predictive models are limited by training sets of few existing measurements, often containing similar classes of molecules that represent just a subset of the potential molecular fuel space. Software tools can be used to generate every possible molecular descriptor for use as input features, but most of these features are largely irrelevant and training models on datasets with higher dimensionality than size tends to yield poor predictive performance. Feature selection has been shown to improve machine learning models, but correlation‐based feature selection fails to provide scientific insight into the underlying mechanisms that determine structure–property relationships. The implementation of causal discovery in feature selection could potentially inform the biofuel design process while also improving model prediction accuracy and robustness to new data. In this study, we investigate the benefits causal‐based feature selection might have on both model performance and identification of key molecular substructures. We found that causal‐based feature selection performed on par with alternative filtration methods, and that a structural causal model provides valuable scientific insights into the relationships between molecular substructures and fuel properties.
- Sponsoring Organization:
- USDOE
- Grant/Contract Number:
- NA0003525
- OSTI ID:
- 1786607
- Journal Information:
- Statistical Analysis and Data Mining, Journal Name: Statistical Analysis and Data Mining Journal Issue: 6 Vol. 14; ISSN 1932-1864
- Publisher:
- Wiley Blackwell (John Wiley & Sons)Copyright Statement
- Country of Publication:
- United States
- Language:
- English
Similar Records
Can machine learning predict fuel properties accurately?
Artificial Neural Network Models for Octane Number and Octane Sensitivity: A Quantitative Structure Property Relationship Approach to Fuel Design