Virtual Reference Sample Calibration for Cell Culture
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing calibration methods for predicting nutrient concentrations in cell cultures are inefficient, requiring extensive cell culture processes, leading to high costs, risk of contamination, and reduced model robustness due to varying cell culture conditions.
Innovation Solution
A calibration system that generates a data set of spectral data and objective variables using machine learning to create a calibration model, eliminating the need for cell culture sampling and allowing for training on diverse conditions, thereby reducing costs and improving model accuracy and robustness.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If conventional calibration methods using actual cell culture samples are used, then model accuracy may be improved through real data, but time consumption and cost increase significantly due to extensive cell culture processes
Solution Approach 1:
The patent creates virtual reference samples that replicate the spectral characteristics and compositional information of actual cell culture samples without requiring physical cell culture processes. These synthetic datasets contain simulated spectral data and corresponding concentration values, enabling model training and validation while eliminating the time-consuming and resource-intensive actual cultivation processes.
2Measurement precision
If conventional calibration methods using actual cell culture samples are used, then model accuracy may be improved through real data, but contamination risks increase due to handling and sampling operations
Solution Approach 1:
The patent generates virtual reference samples through computational methods that simulate spectral data and compositional information without requiring physical sampling of cell cultures. This eliminates all handling, storage, and analysis operations on actual biological samples, thereby completely removing the risk of contamination during the calibration process.
3Reliability
If calibration models are trained on data from varying cell culture conditions, then model robustness should be improved, but difficulty in creating a universal model increases due to condition variability
Solution Approach 1:
The patent systematically varies parameters in the virtual reference samples to represent different cell culture conditions, including changes in component concentrations, spectral characteristics, and measurement conditions. By controlling and documenting these parameter variations in the synthetic datasets, the method enables training of robust models that generalize across conditions while maintaining manageable complexity through structured data generation.
Solution Approach 2:
The calibration model trained on virtual reference samples designed to encompass diverse cell culture conditions achieves universal applicability across different cultivation scenarios. The synthetic training data is constructed to represent a broad range of possible conditions, enabling a single model to function effectively for multiple applications and culture types without requiring separate calibration procedures.
4Quantity of substance
If extensive cell culture processes are used for calibration, then sufficient training data can be obtained, but cost increases due to media, reagents, and operational expenses
Solution Approach 1:
The patent generates large quantities of virtual reference samples through computational algorithms that simulate spectral data and compositional information without consuming physical resources. This approach provides abundant training data for comprehensive model development while eliminating all costs associated with cell culture media, reagents, consumables, and operational expenses required for actual biological sample preparation.
Data Source
AI summary
A calibration device generates a data set of a reference sample including spectral data of the reference sample containing a plurality of components and each objective variable determined by a content of each of the components of the reference sample, and trains, by machine learning using the data set of the reference sample, a machine learning model that outputs at least one objective variable among the objective variables of each of the components in response to input of the spectral data.