Organic Intermolecular Coupling Prediction with Graph Neural Networks
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for determining intermolecular electronic couplings in organic semiconductors are time-consuming and require extensive computational resources, limiting the exploration of vast chemical spaces and are not transferable across different molecules.
Innovation Solution
A machine learning model using graph neural networks (GNNs) predicts intermolecular electronic couplings from 3D molecular geometries, leveraging graph representations of organic molecules to achieve rapid and transferable predictions.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If density functional theory (DFT) calculations are used to determine intermolecular electronic couplings, then high accuracy is achieved, but computational time increases significantly
Solution Approach 1:
The patent creates a machine learning model that copies the predictive capability of expensive DFT calculations but operates much faster. The ML model is trained on a dataset of intermolecular electronic couplings calculated using DFT, allowing it to reproduce high-accuracy results at a fraction of the computational cost. This copying approach enables rapid prediction without requiring repeated expensive DFT calculations.
Solution Approach 2:
The patent performs preliminary DFT calculations on a training dataset to generate the machine learning model. This upfront computational effort creates a transferable model that can then rapidly predict intermolecular electronic couplings for new molecules without requiring time-consuming DFT calculations during actual use. The costly computational work is done in advance during model training rather than during prediction.
2Adaptability or versatility
If molecular dimer geometries are transformed to machine learning model input using coulomb matrix, then transferable model is achieved, but model performance decreases due to sparsity from large variation in atom numbers
Solution Approach 1:
The patent employs graph neural networks that process molecular structures through local neighborhood relationships. Each atom node in the molecular graph maintains its own features and interactions with neighboring atoms, allowing the model to handle variable molecular sizes without sparsity issues. This local processing approach preserves information density regardless of the number of atoms in the molecule.
Solution Approach 2:
The patent transitions from using coulomb matrices (which have dimensions based on atom count) to graph representations where the structure is defined by connectivity relationships rather than fixed dimensional arrays. This dimensional transformation allows the model to accommodate molecules with different numbers of atoms while maintaining consistent representation quality, eliminating the sparsity problem inherent in fixed-size matrix representations.
3Ease of manufacture
If trial-and-error design is used for organic semiconductor development, then simplicity of approach is maintained, but productivity decreases due to slow exploration of chemical space
Solution Approach 1:
The patent implements a machine learning pipeline that automatically predicts intermolecular electronic couplings and screens organic molecules for semiconductor applications. The system self-services the evaluation process by taking molecular inputs and automatically generating predictions without requiring manual computational analysis. This automation dramatically increases the throughput of chemical space exploration while maintaining the simplicity of the overall workflow.
Solution Approach 2:
The patent replaces manual trial-and-error computational analysis with an automated machine learning system. Instead of manually performing DFT calculations and analyzing results for each candidate molecule, the system uses pre-trained ML models to automatically evaluate thousands of molecules, substituting the mechanical process of manual computation with automated algorithmic prediction.
Data Source
AI summary
This disclosure is directed to machine learning-based methods for modelling the intermolecular electronic couplings of organic molecules. The method comprises inputting synthetically generated graph representations of at least two organic molecules into a machine learning system. The machine learning system predicts an intermolecular coupling property (V) between the at least two organic molecules from the molecular graph representations. The machine learning system further determines an anisotropic charge-carrier mobility value for the organic molecule from the predicted intermolecular coupling property. The machine learning system then determines whether the anisotropic charge-carrier mobility value meets or exceeds a predetermined threshold anisotropic charge-carrier mobility value. Machine learning systems for modelling intermolecular electronic couplings of at least two organic molecules, and methods for training the machine learning systems, are described also.