Formulation graph convolution networks (f-GCN) for predicting performance of formulated products
A formulation graph convolution network (F-GCN) with multiple GCNs assembled in parallel and connected to filters and an external learning architecture is able to predict the effectiveness of a formulation. Input into the multiple GCNs are molecular structures of formulants, which are processed as molecular graphs and output as molecular descriptors. The molecular descriptors are filtered by normalized ratios or fractions of the ingredient molecules in a formulation, such as a battery electrolyte or solvent. A formulation descriptor combines the filtered molecular descriptors to arrive at a predicted performance for the formulation, such as the battery capacity for an electrolyte formulation, by an external learning architecture. F-GCN may use a pre-trained GCN with physico-chemical properties of known molecular structures.
1 . A system comprising:
a processor set;
one or more computer readable storage media; and
program instructions stored on the one or more computer readable storage media to cause the processor set to perform operations comprising:
inputting multiple molecular structures into multiple respective graph convolution networks (GCNs) assembled in parallel, wherein input into each GCN is a molecular structure of the multiple molecular structures and output from each GCN is a molecular descriptor of the corresponding input molecular structure;
scaling each molecular descriptor based on a normalized ratio of each respective molecular structure in a formulation;
combining each scaled molecular descriptor to form a single formulation descriptor representing each molecular structure in the formulation; and
obtaining a performance label for the formulation from a neural network, wherein the neural network generates the performance label as output based on the single formulation descriptor as input.
2 . The system of claim 1 , wherein the molecular structures are input into the GCN via a symbolic line notation of the molecular structure.
3 . The system of claim 1 , wherein the molecular structures input into each GCN comprise a node matrix and an adjacency matrix, wherein the node matrix comprises information about atoms in the corresponding input molecular structure and the adjacency matrix comprises information about bonds within the corresponding input molecular structure.
4 . The system of claim 3 , wherein the node matrix comprises one-hot encoded electronegativities of atoms in the corresponding input molecular structure.
5 . The system of claim 1 , wherein the multiple GCNs are pre-trained with physico-chemical properties of known molecular structures.
6 . A system comprising:
a processor set;
one or more computer readable storage media; and
program instructions stored on the one or more computer readable storage media to cause the processor set to perform operations comprising:
inputting multiple molecular structures into multiple respective graph convolution networks (GCNs) assembled in parallel of a formulation graph convolution network (F-GCN), wherein input into each GCN is a molecular structure of the multiple molecular structures for a compound in an electrolyte formulation and output from each GCN is a molecular descriptor of the corresponding input molecular structure;
scaling, by an intermediate layer of the F-GCN, the molecular descriptor of each of the multiple molecular structures according to a normalized ratio of each respective molecular structure in the electrolyte formulation;
combining, by an output layer of the F-GCN, each scaled molecular descriptor to form a single formulation descriptor representing each molecular structure in the electrolyte formulation; and
obtaining a battery capacitance performance label for the electrolyte formulation from a neural network as output based on the single formulation descriptor as input.
7 . The system of claim 6 , wherein the molecular structures are input into the GCN via a symbolic line notation selected from the group consisting of SMILES, INCHI Keys, SELFIES, and BigSMILES.
8 . The system of claim 6 , wherein the molecular structures input into each GCN comprise a node matrix and an adjacency matrix, wherein the node matrix comprises information about atoms in the corresponding input molecular structure and the adjacency matrix comprises information about bonds within the corresponding input molecular structure.
9 . The system of claim 8 , wherein the node matrix comprises one-hot encoded electronegativities of atoms in the corresponding input molecular structure.
10 . The system of claim 6 , wherein the multiple GCNs are pre-trained with homo-lumo energy levels from known molecular structures.
11 . A computer-implemented method comprising:
inputting multiple molecular structures into multiple respective graph convolution networks (GCNs) arranged in parallel of a formulation graph convolutional network (F-GCN), wherein input into each GCN is a molecular structure of the multiple molecular structures and output from each GCN is a molecular descriptor of the corresponding input molecular structure;
scaling, by an intermediate layer of the F-GCN, each of the molecular descriptors based on a normalized ratio of each respective molecular structure in a formulation;
combining, by an output layer of the F-GCN, each scaled molecular descriptor to form a single formulation descriptor representing each molecular structure in the formulation; and
feeding the single formulation descriptor into an external learning architecture comprising a neural network to obtain a performance label of the formulation based on the single formulation descriptor.
12 . The computer-implemented method of claim 11 , wherein the molecular structures are input into the GCN via symbolic line notation for the molecular structures.
13 . The computer-implemented method of claim 11 , wherein the molecular structures input into each GCN comprise a node matrix and an adjacency matrix, wherein the node matrix comprises information about atoms in the corresponding input molecular structure and the adjacency matrix comprises information about bonds within the corresponding input molecular structure.
14 . The computer-implemented method of claim 13 , wherein the node matrix comprises one-hot encoded electronegativities of atoms in the corresponding input molecular structure.
15 . The computer-implemented method of claim 11 , wherein the multiple GCNs are pre-trained with physical chemical properties of known molecular structures.
16 . The computer-implemented method of claim 11 , wherein the external learning architecture is selected from a group consisting of a dense neural network and a GCN.
17 . A computer-implemented method comprising:
inputting multiple molecular structures of a battery electrolyte formulation into multiple respective graph convolution networks (GCNs) arranged in parallel of a formulation graph convolution network (F-GCN), wherein input into each GCN is a molecular structure of the multiple molecular structures of the battery electrolyte formulation and output from each GCN is a molecular descriptor of the corresponding input molecular structure;
scaling, by an intermediate layer of the F-GCN, each of the molecular descriptors based on a normalized ratio of each respective molecular structure in the battery electrolyte formulation;
combining, by an output layer of the F-GCN, each scaled molecular descriptor to form a single formulation descriptor representing each molecular structure in the battery electrolyte formulation; and
feeding the single formulation descriptor into an external learning architecture comprising a neural network to cause generation of a battery capacitance prediction for the battery electrolyte formulation.
18 . The computer-implemented method of claim 17 , wherein the molecular structures are input into the GCN via a symbolic line notation selected from the group consisting of SMILES, INCHI Keys, SELFIES, and BigSMILES.
19 . The computer-implemented method of claim 17 , wherein the molecular structures input into each GCN comprise a node matrix and an adjacency matrix, wherein the node matrix comprises information about atoms in the corresponding input molecular structure and the adjacency matrix comprises information about bonds within the corresponding input molecular structure.
20 . The computer-implemented method of claim 19 , wherein the node matrix comprises one-hot encoded electronegativities of atoms in the corresponding input molecular structure.
21 . The computer-implemented method of claim 17 , wherein the multiple GCNs are pre-trained with homo-lumo energy levels from known molecular structures.
22 . A computer program product comprising:
one or more computer readable storage media; and
program instructions stored on the one or more computer readable storage media to perform operations comprising:
inputting multiple molecular structures into multiple respective graph convolution networks (GCNs) assembled in parallel of a formulation graph convolutional network (F-GCN), wherein input into each GCN is a molecular structure of the multiple molecular structures and output from each GCN is a molecular descriptor of the corresponding input molecular structure;
scaling, by an intermediate layer of the F-GCN, each the molecular descriptors based on a normalized ratio of each respective molecular structure in a formulation;
combining, by an output layer of the F-GCN, each scaled molecular descriptor into a single formulation descriptor representing each molecular structure in the formulation; and
feeding the single formulation descriptor into an external learning architecture comprising a neural network to obtain a prediction performance label for the formulation for an application.
23 . The computer program product of claim 22 , wherein the molecular structures are input into the GCN via symbolic line notations for the molecular structures.
24 . The computer program product of claim 22 , wherein the molecular structures input into each GCN comprises a node matrix and an adjacency matrix, wherein the node matrix comprises one-hot encoded electronegativities of atoms in the corresponding input molecular structure and the adjacency matrix comprises information about bonds within the corresponding input molecular structure.
25 . The system of claim 1 , wherein the multiple GCNs are pre-trained with homo-lumo energy levels of known molecular structures, wherein the molecular structures input into each GCN are in an embedded form comprising a node matrix and an adjacency matrix, wherein the node matrix comprises one-hot encoded electronegativities of atoms in the corresponding input molecular structure and the adjacency matrix comprises information about bonds within the corresponding input molecular structure, wherein the embedded forms of corresponding molecular structures are modified by a respective GCN of the multiple GCNs to obtain the corresponding molecular descriptor as output.