Polymer Production Prediction Using Reinforcement Learning Feedback

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing methods for predicting polymeric product properties are limited in identifying complex and subtle correlations between production parameters and formulation data, often requiring intuitive hypothesis formation which may miss essential relationships.

Innovation Solution

The use of reinforcement learning to develop a prediction model that describes functional relationships between production parameters, formulation data, and product properties, allowing for incremental building and refinement of the model based on user input and feedback.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If traditional analytical approaches and supervised learning are used to build prediction models, then the model development process is straightforward and controllable, but the model fails to identify complex and subtle correlations between production parameters and product properties

Engineering Contradiction:
Improveprediction accuracyVSAvoidmodel complexity
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The patent replaces traditional analytical and supervised learning methods with reinforcement learning algorithms. This substitution enables the system to automatically explore and identify complex, non-linear correlations between production parameters and product properties without requiring pre-defined hypotheses or manual feature engineering, thereby achieving higher prediction accuracy for complex polymeric product relationships.

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

Solution Approach 2:

The patent transforms the prediction model development approach by changing from static, hypothesis-driven parameter relationships to dynamic, exploration-driven parameter optimization. The reinforcement learning agent continuously adjusts model parameters based on feedback from production data, enabling the system to adapt to subtle and complex correlations that traditional methods miss.

Inventive Principle:
Principle #35Parameter changes

2Measurement precision

If reinforcement learning is used to develop prediction models, then the model can identify complex and subtle correlations, but the computational resources and time required increase significantly

Engineering Contradiction:
Improveprediction accuracyVSAvoidmodel development time
Core Design Contradiction:
Measurement precisionVSLoss of time

Solution Approach 1:

The patent implements preliminary action by using reinforcement learning to pre-train prediction models on historical production data before actual production scenarios. This allows the model to learn complex correlations in advance, so that during actual production, the model can provide accurate predictions quickly without requiring extensive real-time computation or iteration.

Inventive Principle:
Principle #10Preliminary action

3Measurement precision

If reinforcement learning is used to develop prediction models, then explorative capabilities improve and complex correlations are identified, but the computational resources required increase

Engineering Contradiction:
Improveprediction accuracyVSAvoidcomputational resource consumption
Core Design Contradiction:
Measurement precisionVSUse of energy by moving object

Solution Approach 1:

The patent applies partial action by implementing reinforcement learning selectively for identifying complex correlations in specific production scenarios rather than applying it universally. The system uses reinforcement learning exploratively only when traditional methods fail to achieve sufficient accuracy, thereby reducing overall computational resource consumption while still capturing the benefits of explorative model development where needed.

Inventive Principle:
Principle #16Partial or excessive action

Data Source

PatentEP3850593B1Method for improving prediction relating to the production of a polymeric product
Publication Date: 2022.10.05 COVESTRO INTELLECTUAL PROPERTY GMBH & CO KG
  • EP3850593B1 patent drawingFigure 1
  • EP3850593B1 patent drawingFigure 2

AI summary

The invention relates to a method for improving prediction relating to the production of a polymeric product, wherein a prediction model (1) is provided for describing a functional relationship between production parameters (2), which production parameters (2) comprise formulation portions data specifying raw material portions used for the production of a respective polymeric product and comprise processing parameters data specifying machine process properties during the production of that polymeric product, and product properties data (3) associated with that polymeric product on a computer system (5), which production parameters (2) and product properties data (3) form data entry properties for a respective polymeric product, wherein user input is provided comprising user product targets specifying only a part of the data entry properties, wherein a new data entry (7) with data entry properties is generated by the computer system (5) for realizing the user product targets, wherein for the new data entry (7) the specified data entry properties are determined, wherein a reward metric (8) is determined by the computer system (5) based on the determined data entry properties and wherein based upon the reward metric (8) the prediction model (1) is updated by the computer system (5).