Coating Composition Prediction Using CNN and Active Learning
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The development of new compositions for paints, varnishes, printing inks, grinding resins, and pigment concentrates is time-consuming and costly due to the complex interactions of raw materials, making it difficult to predict desired properties without extensive empirical testing.
Innovation Solution
A method using a convolutional neural network (CNN) and an active learning module to iteratively train a neural network with selected test compositions, expanding the training data record to enhance predictive power, allowing for the generation of compositions with desired properties.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of manufacture
If conventional empirical methods are used to develop new compositions, then the development process is straightforward and easy to implement, but it is time-consuming and cost-intensive due to the need for extensive chemical synthesis and testing
Solution Approach 1:
The patent replaces the mechanical/chemical system of trial-and-error synthesis and testing with an information-processing system (neural network) that computationally predicts composition properties. The CNN processes composition data and predicts properties without physical synthesis, substituting computational algorithms for laboratory experimentation.
Solution Approach 2:
The patent creates a virtual model (neural network) that copies and simulates the complex relationships between composition components and properties. Instead of physically testing each composition, the system uses a trained neural network model to predict outcomes based on learned patterns from training data.
2Measurement precision
If a cognitive computer system with learning logic is trained using existing chemical formulation data, then prediction capability is improved, but creating sufficiently large and balanced training data records is complex and expensive
Solution Approach 1:
The patent performs preliminary actions by pre-processing and organizing training data before neural network training. The system prepares balanced training datasets in advance, performing data cleaning, normalization, and balancing operations to ensure the neural network receives high-quality input data, thereby improving prediction accuracy while managing complexity.
Solution Approach 2:
The neural network system performs self-service by automatically learning from training data and improving its own prediction capabilities. The learning logic autonomously processes training examples, adjusts its internal parameters, and enhances prediction accuracy without requiring manual intervention for each prediction task.
3Reliability
If the training data record is expanded by synthesizing and analyzing more test compositions, then the neural network's predictive power is enhanced, but the time and material consumption increases
Solution Approach 1:
The patent applies partial action by using a carefully selected subset of training data that is sufficient to achieve reliable prediction without requiring exhaustive coverage of all possible compositions. The system identifies and uses the most informative training examples, avoiding the need to synthesize and analyze every possible composition variant.
Solution Approach 2:
The patent changes parameters of the training process, such as the size and composition of training datasets, the architecture of the neural network, and the training algorithms used. By optimizing these parameters, the system achieves high predictive power with reduced training time and resource consumption.
4Measurement precision
If extensive empirical testing is conducted to assess composition properties, then accurate property assessment is achieved, but the process becomes cost-intensive and time-consuming
Solution Approach 1:
The patent replaces physical property measurement systems with a computational prediction system. The neural network predicts composition properties based on input data, substituting laboratory measurement instruments and procedures with algorithm-based predictions, thereby improving productivity while maintaining acceptable accuracy.
Data Source
AI summary
The method includes using known compositions to train the convolutional neural network, a loss function being minimized for the training; examining whether the value of a loss function meets a predefined criterion, the following steps being carried out selectively in the case where the criterion is not met—selecting a test composition from a set of predefined test compositions by an active learning module; activating a chemical apparatus for producing and examining compositions for paints, varnishes, printing inks, grinding resins, pigment concentrates or other coating substances for the purpose of producing and examining the selected test composition; training the convolutional neural network, using the selected test composition and the properties thereof detected by the apparatus; generating a prediction composition for paints, varnishes, printing inks, grinding resins, pigment concentrates or other coating substances by inputting an input vector into the convolutional neural network; and outputting the prediction composition.