Whilst new biologically derived molecules are being suggested as biofuels, screening large numbers of these molecules for their combustion performance is a challenge. Methods to predict combustion performance provide a useful indication of a molecule's suitability as a fuel. Cetane is an important combustion performance metric for diesel-like combustion regimes. However, a variety of techniques exist for measuring cetane in pure compounds. These include both highly variable methods, (e.g, blending or ignition delay methods) and highly accurate gold-standard measurements (e.g., ASTM D6890 and D7170 methods). Traditionally, machine learned classification and prediction of cetane in pure compounds from known structural features or other chemical features has been forced to choose between accepting highly variable and potentially inaccurate measurements in the training data, or removing potentially useful data. In this study we use cetane measurements from 387 unique compounds, which cover a variety of measurement techniques, to train a Random Forest classifier for predicting high/low cetane in 50 recently measured hydrocarbons/oxygenates. This technique uses sampling of decision trees to weight values based on their measurement method. We use previous assessments of cetane measurement accuracy and reproducibility to preferentially weight each training compound measurement. After assessing the accuracy of this technique in 50 validation compounds, we use structural and physical/chemical data to rank 8701 potentially biologically produced hydrocarbons/oxygenates for cetane. This machine learning tool provides not only a means of ranking pure compounds by cetane, but also provides computational chemists and cheminformaticists with a new method for including measurements of varying quality in their predictive models.<br/>

Developing a high- and low-cetane classifier for biologically produced chemicals using variable quality training data

253rd National Meeting (2017)

The official archive and repository for all ACS Meetings & Expositions abstracts and select posters or presentations.