Biochemical Process Control Using Simulated Data for Quality Prediction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for producing biochemical products, such as vaccines or adjuvants, face challenges in developing new processes quickly and efficiently while ensuring the desired quality and quantity, particularly when historical data is sparse, leading to inaccurate AI supervision and increased waste and time delays.
Innovation Solution
A method that trains a predictive model using a combination of historical and simulated data, where a physical model generates simulated quality attributes, allowing for accurate predictions and reducing the need for reactants and minimizing waste, by deploying a neural network or other algorithms to correct actuation parameters in real-time.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If historical data is used to train the predictive model, then the model can be trained with real experimental data, but the training database is insufficient when historical data is sparse
Solution Approach 1:
The patent creates a virtual copy of the physical process through a physical model that simulates process behavior and generates synthetic training data. This virtual process replicates the essential characteristics of the real biochemical process, allowing the predictive model to be trained on simulated data that mirrors real operational conditions without requiring extensive historical experimental data.
Solution Approach 2:
The patent performs preliminary actions by developing and validating the physical model before actual process optimization begins. The model is pre-trained with available sparse historical data and then used to generate additional synthetic training data in advance, preparing a comprehensive training database before the predictive model needs to make accurate predictions.
2Measurement precision
If more experimental data is collected to train the predictive model, then the model accuracy improves, but the time required for data collection and processing increases
Solution Approach 1:
Instead of collecting extensive experimental data through time-consuming real-world experiments, the patent uses a virtual copy (physical model) to generate synthetic training data instantaneously. This eliminates the time delay associated with conducting, recording, and processing real experimental trials while providing sufficient data for model training.
Solution Approach 2:
The physical model is developed and validated in advance to serve as a ready-to-use data generation tool. Once established, the model can rapidly produce training datasets without requiring additional experimental time, thus preliminary preparation of the modeling framework eliminates ongoing data collection delays.
3Measurement precision
If real experimental data is used for training, then the predictive model reflects actual process behavior, but the waste of reactants and materials increases
Solution Approach 1:
The patent replaces physical material consumption with virtual simulation. The physical model replicates process behavior and generates training data without requiring actual reactants, reagents, or biochemical materials. This virtual copying approach maintains predictive accuracy while eliminating material waste associated with extensive experimental trials.
Solution Approach 2:
The physical model serves itself by generating its own training data through simulation rather than requiring external experimental data collection. The model uses mathematical representations of process physics and chemistry to produce synthetic datasets, making the system self-sufficient and eliminating the need for material-intensive external experiments.
Data Source
Figure 1~5
Figure 6~9
Figure 10~11
AI summary
The invention relates to a method for optimising a process (PROC) to produce a biochemical product (P) defined by a quality attribute, the process being controlled by an actuation parameter (C) and being monitored to get a measured value (T), the method comprising the following steps : - training a predictive model (PRED) on a training database ; - deploying the trained predictive model (PRED) to provide a correction actuation parameter (dC) when a predicted quality attribute (pQA) is out of a targeted quality attribute interval (QAmin, QAmax); the method comprising a step of designing a physical model of the process (PROC) able to provide a simulated quality attribute, the training database comprising simulated quality attributes computed from the physical model and experimental quality attributes computed from biochemical products (P) previously produced.