Related Experiment Videos
Prediction of C60 solubilities from solvent molecular structures
1Department of Chemistry, The Pennsylvania State University, 152 Davey Laboratory, University Park, Pennsylvania 16801, USA.
Journal of Chemical Information and Computer Sciences
|March 30, 2001
Summary
Computational neural networks (CNN) accurately predict fullerene solubility across diverse solvents. These models, using molecular descriptors, offer reliable predictions for C60 solubility, aiding material science applications.
Area of Science:
- Physical Chemistry
- Computational Chemistry
- Materials Science
Background:
- Fullerene solubility is crucial for their application in various fields.
- Predicting solubility accurately is challenging due to complex solvent-fullerene interactions.
Purpose of the Study:
- To develop predictive models for fullerene (C60) solubility in a wide range of solvents.
- To compare the performance of multiple linear regression and computational neural networks (CNN) for solubility prediction.
Main Methods:
- Utilized a dataset of fullerene solubility in 96 solvents at 298 K.
- Employed multiple linear regression and feed-forward computational neural networks (CNN).
- Solvents were characterized by calculated molecular structure descriptors.
Main Results:
- Developed multiple linear regression models and CNN models based on these linear models.
- The best CNN model (9-3-1 architecture) achieved a root-mean-square error (RMSE) of 0.255 log units on the training set.
- Cross-validation and external prediction sets showed RMSEs of 0.253 and 0.346 log units, respectively.
Conclusions:
- Computational neural networks provide accurate predictions of fullerene solubility.
- The developed models can be valuable tools for estimating C60 solubility in untested solvents.
- Molecular descriptors are effective in representing solvents for solubility modeling.