Retrosynthesis Neural Network Parallel Training Optimization
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing retrosynthesis processing methods are inefficient due to low accuracy and long computation times, resulting in non-optimized retrosynthesis routes.
Innovation Solution
A method involving a target neural network trained using a sample cost dictionary generated by concurrent retrosynthesis reaction training on multiple sample molecules, implementing a distributed parallel computing scheme to optimize the training process and improve calculation speed.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If traditional sequential training methods are used for retrosynthesis reaction training, then training accuracy can be maintained, but training time becomes excessively long and computation efficiency is low
Solution Approach 1:
The patent segments the training process by dividing sample molecules into multiple groups that can be trained independently and concurrently. Each group is processed separately through the neural network, allowing parallel execution of training operations across different molecular subsets, thereby significantly reducing overall training time while maintaining comprehensive coverage of the training data
Solution Approach 2:
The patent transitions from sequential single-threaded training to parallel multi-threaded training by adding a temporal dimension to the processing architecture. Multiple training operations execute simultaneously across different computational threads, transforming the training process from a linear sequence into a parallel distributed system that achieves substantial speedup
2Productivity
If conventional retrosynthesis processing methods are used, then implementation simplicity is maintained, but calculation speed is slow and iteration efficiency is low
Solution Approach 1:
The patent creates multiple copies of the training and processing operations that can execute simultaneously. Each computational thread maintains its own instance of the neural network processing pipeline, allowing independent parallel execution of retrosynthesis calculations on different molecular subsets without interfering with other threads, thus achieving high-speed concurrent processing
3Measurement precision
If accurate retrosynthesis reaction training is performed on multiple sample molecules, then retrosynthesis route accuracy improves, but computation time increases significantly
Solution Approach 1:
The patent performs preliminary actions by pre-dividing the sample molecule dataset into multiple independent groups before training begins. This pre-segmentation allows subsequent parallel training operations to proceed simultaneously without requiring complex coordination during execution, maintaining high accuracy through comprehensive training while reducing overall computation time through efficient parallelization
Solution Approach 2:
The patent ensures continuous useful action by maintaining multiple parallel training threads that continuously process different molecular subsets simultaneously. This continuous parallel processing eliminates idle time between training operations and ensures that computational resources are fully utilized throughout the training process, achieving both high accuracy and efficient time utilization
Data Source
AI summary
Embodiments of this application relate to an electronic device for performing a retrosynthesis processing method, and an associated non-transitory computer-readable storage medium. The method includes determining molecular representation information of a target molecule; inputting the molecular representation information into a target neural network; and performing, via the target neural network, retrosynthesis processing on the target molecule based on the molecular representation information of the target molecule, to obtain a respective retrosynthesis reaction of the target molecule for each step of the retrosynthesis processing.