Deep Learning DOSY Processing with Synthetic Data
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for processing two-dimensional Nuclear Magnetic Resonance (NMR) Diffusion Ordered Spectroscopy (DOSY) face challenges such as the need for prior knowledge of molecular components, complicated parameter adjustments, and inefficiencies in processing data with varying dimensional sizes, especially due to limitations in experimental samples and instrument time.
Innovation Solution
A deep learning-based method is developed that constructs a simulated dataset using a mathematical model, generates labels for training, and employs a deep learning network model with a multi-head attention module and adaptive optimization to process DOSY data, eliminating the need for exact component knowledge and allowing for efficient processing across varying dimensions.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If traditional deep learning methods are used for DOSY processing, then processing speed can be improved, but the methods require large numbers of training samples which are impossible to obtain due to experimental limitations
Solution Approach 1:
The patent creates synthetic training data by copying and transforming simulated DOSY spectra through various operations (adding noise, varying parameters, mixing samples) to generate a large training dataset without requiring additional physical experiments. This allows the deep learning model to be trained on abundant synthetic data that mimics real experimental conditions.
Solution Approach 2:
The patent performs preliminary generation of simulated DOSY data and training dataset creation before the actual DOSY processing is needed. By pre-generating diverse training samples with known ground truth values and pre-training the model, the system avoids the need to collect large numbers of real experimental samples during actual use.
2Ease of operation
If conventional neural networks are used, then processing can be performed, but the dimensional size of testing data must match training data which varies with experimental parameters
Solution Approach 1:
The patent implements a dynamic architecture where the neural network can accept variable input dimensions. The model dynamically adjusts its processing based on the actual size of the input gradient dimension, allowing it to handle different experimental conditions without retraining. This is achieved through flexible layer designs that can adapt to varying data sizes.
Solution Approach 2:
The patent creates a universal model that can process DOSY data with any gradient dimension size. The network architecture is designed to be multi-functional, handling both training and inference with varying dimensions, and can process different types of DOSY experiments (varying numbers of gradient steps, different gradient strengths) with the same trained model.
3Adaptability or versatility
If ILT methods are used for DOSY processing, then prior knowledge of component quantities is not required, but complicated adjustments of regularization parameters are needed to optimize results
Solution Approach 1:
The patent implements a self-service approach where the deep learning model automatically learns optimal processing parameters during training and applies them during inference without user intervention. The model self-adjusts to handle different experimental conditions and automatically determines the number of components and their properties, eliminating the need for users to manually tune regularization parameters.
Solution Approach 2:
The patent transforms the complex parameter adjustment problem into a learning problem. Instead of requiring users to manually adjust regularization parameters, the model learns optimal parameter settings from synthetic training data and generalizes them to real data. The network internally manages parameter optimization through its learned weights and biases, converting a manual tuning process into an automated learning process.
Data Source
AI summary
A method for processing two-dimensional (2D) nuclear magnetic resonance (NMR) Diffusion Ordered Spectroscopy (DOSY) based on deep learning comprises constructing a simulated dataset by generating simulated data using a mathematical model based on signal characteristics of the 2D NMR DOSY, generating labels for training a deep learning network model, wherein the labels comprise a first two-dimensional matrix, and two dimensions of the first two-dimensional matrix comprise chemical shift and diffusion coefficients, constructing the deep learning network model and setting training parameters of the deep learning network model, training the deep learning network model using the simulated dataset, and testing the deep learning network model.