Graph neural networks (GNNs) are emerging in chemical engineering for the end-to-end learning of physicochemical properties based on molecular graphs. A key element of GNNs is the pooling function which combines atom feature vectors into molecular fingerprints. Most previous works use a standard pooling function to predict a variety of properties. However, unsuitable pooling functions can lead to unphysical GNNs that poorly generalize. We compare and select meaningful GNN pooling methods based on physical knowledge about the learned properties. The impact of physical pooling functions is demonstrated with molecular properties calculated from quantum mechanical computations. We also compare our results to the recent set2set pooling approach. We recommend using sum pooling for the prediction of properties that depend on molecular size and compare pooling functions for properties that are molecular size-independent. Overall, we show that the use of physical pooling functions significantly enhances generalization.

Co-design of alternative fuels and future spark-ignition (SI) engines allows very high engine efficiencies to be achieved. To tailor the fuel’s molecular structure to the needs of SI engines with very high compression ratios, computer-aided molecular design (CAMD) of renewable fuels has received considerable attention over the past decade. To date, CAMD for fuels is typically performed by computationally screening the physicochemical properties of single molecules against property targets. However, achievable SI engine efficiency is the result of the combined effect of various fuel properties, and molecules should not be discarded because of individual unfavorable properties that can be compensated for. Therefore, we present an optimization-based fuel design method directly targeting SI engine efficiency as the objective function. Specifically, we employ an empirical model to assess the achievable relative engine efficiency increase compared to conventional RON95 gasoline for each candidate fuel as a function of fuel properties. For this purpose, we integrate the automated prediction of various fuel properties into the fuel design method: Thermodynamic properties are calculated by COSMO-RS; combustion properties, indicators for environment, health and safety, and synthesizability are predicted using machine learning models. The method is applied to design pure-component fuels and binary ethanol-containing fuel blends. The optimal pure-component fuel tert-butyl formate is predicted to yield a relative efficiency increase of approximately 8% and the optimal fuel blend with ethanol and 3,4-dimethyl-3-propan-2-yl-1-pentene of 19%.