Neural Network NMR Training Data Generation

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Current methods for determining the concentration of molecules in NMR samples are prone to errors due to subjective human interpretation and are limited by the complexity of NMR spectra, which often include overlapping signals and noise, making it difficult to accurately identify and quantify molecules.

Innovation Solution

A computer-implemented method generates synthetic NMR spectra as a training dataset for neural networks, allowing them to autonomously identify and quantify molecules by simulating realistic training data, including background signals and noise, to improve prediction accuracy and objectivity.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If human experts manually interpret NMR spectra to identify and quantify molecules, then subjective expertise can be applied, but errors due to subjectivity and inconsistency increase

Engineering Contradiction:
Improveconcentration determination accuracyVSAvoidinterpretation consistency
Core Design Contradiction:
Measurement precisionVSReliability

Solution Approach 1:

The patent replaces the mechanical system of human expert interpretation with an automated computational system. A neural network model processes NMR spectra to determine molecule concentrations, eliminating subjective human interpretation while maintaining or improving measurement precision. The system substitutes human cognitive processes with algorithmic processing, achieving consistent and reproducible results across different samples and analysts.

Inventive Principle:
Principle #28Mechanics substitution (Replace mechanical system)

Solution Approach 2:

The patent creates a computational model that copies and learns from expert knowledge. By training the neural network on labeled NMR spectra data, the system captures expert interpretation patterns and reproduces them automatically. This copying of expert expertise into a machine-learning model enables consistent application of sophisticated analytical reasoning without human subjectivity.

Inventive Principle:
Principle #26Copying

2Ease of operation

If simple numerical integration is used to determine molecule concentration, then the process is straightforward, but it fails when signals overlap with background or other peaks

Engineering Contradiction:
Improveconcentration determination simplicityVSAvoidconcentration determination accuracy
Core Design Contradiction:
Ease of operationVSMeasurement precision

Solution Approach 1:

The patent transforms the NMR spectrum from raw signal data into a feature representation suitable for neural network processing. The system extracts relevant spectral features and converts them into a format that captures complex signal relationships, enabling the model to distinguish target molecule signals from overlapping background and interference signals while maintaining operational simplicity.

Inventive Principle:
Principle #35Parameter changes

Solution Approach 2:

The patent replaces the simple but insufficient numerical integration method with an advanced neural network-based analysis system. This substitution enables the handling of complex overlapping signals that simple integration cannot resolve, significantly improving measurement precision for concentration determination in challenging spectral conditions.

Inventive Principle:
Principle #28Mechanics substitution (Replace mechanical system)

3Measurement precision

If spectral fitting with parameter optimization is used, then overlapping signals can be resolved, but prior knowledge of spectra is required and the process becomes complex

Engineering Contradiction:
Improvesignal peak identification accuracyVSAvoidspectral analysis complexity
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The patent performs preliminary training of the neural network model using labeled NMR spectra data before actual analysis. During this training phase, the system learns spectral patterns, peak characteristics, and concentration relationships. This preliminary action enables the model to automatically handle complex spectral fitting and peak identification during actual use without requiring real-time parameter optimization or prior spectral knowledge from the user.

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The patent implements a self-learning system where the neural network automatically adapts to different spectral patterns and molecular compositions. The model performs self-service by autonomously identifying peaks, resolving overlaps, and determining concentrations without requiring manual intervention, prior spectral knowledge, or complex parameter optimization procedures, thereby reducing operational complexity while maintaining high precision.

Inventive Principle:
Principle #25Self-service

4Measurement precision

If more FIDs are acquired and averaged to improve signal-to-noise ratio, then measurement sensitivity increases, but measurement time increases

Engineering Contradiction:
Improvesignal-to-noise ratioVSAvoidmeasurement time
Core Design Contradiction:
Measurement precisionVSLoss of time

Solution Approach 1:

The patent applies a neural network model that can extract meaningful signal information even from spectra with lower signal-to-noise ratios. By using advanced pattern recognition and feature extraction capabilities, the system achieves accurate concentration determination without requiring extensive signal averaging, thereby reducing the number of FIDs needed and shortening measurement time while maintaining measurement precision.

Inventive Principle:
Principle #16Partial or excessive action

Data Source

PatentEP4356153B1Systems and methods for provisioning training data to enable neural networks to analyze signals in NMR measurements
Publication Date: 2024.12.18 BRUKER BIOSPIN MRI GMBH
  • EP4356153B1 patent drawingFigure 1
  • EP4356153B1 patent drawingFigure 2
  • EP4356153B1 patent drawingFigure 3A~4

AI summary

A system (100), method and computer program product for generating a data record of a training dataset (140) set configured to train a neural network (230) for determination of the concentration (215c) of a particular target molecule in an NMR sample (201). An NMR spectrum (213) associated with a known concentration (c1) of the target molecule is obtained. The obtained NMR spectrum is adjusted by applying a random shift to generate an adjusted NMR spectrum (213a). A background generator 132 adds a background spectrum which reflects contributions of impurities in the NMR sample. The resulting NMR spectrum together with the information about the concentration of the target molecule is then stored as a new data record of the training dataset (140).