Domain-Aware Neural Modeling for Sparse Industrial Process Prediction

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Conventional models, such as physics-based and data-driven models, face challenges in accurately predicting the behavior of complex industrial entities like rotary kilns due to sparse and dynamic measurements, leading to decreased prediction accuracy and convergence issues in multiphysics processes.

Innovation Solution

A domain aware data driven (DADD) model is developed using a grouped neural network architecture (GNNA) that integrates physics-informed neural networks (PINNs) with sequential training, where sub-process governing equation components are identified and trained sequentially, utilizing sparse measurements and domain knowledge to enhance convergence and prediction accuracy.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If physics-based models are used to model industrial systems, then model interpretability and physical consistency are improved, but the models face challenges when model parameters change or model properties are unavailable

Engineering Contradiction:
Improvemodel reliabilityVSAvoidmodel adaptability
Core Design Contradiction:
ReliabilityVSAdaptability or versatility

Solution Approach 1:

The patent segments the complex industrial system into multiple sub-processes, each governed by its own equation components. This segmentation allows the model to handle parameter changes in individual sub-processes independently while maintaining overall system reliability through the structured framework.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent employs dynamic neural network architectures that can adapt to changing model parameters and properties. The neural networks are designed to learn and adjust to new conditions, providing both the reliability of physics-based modeling and the adaptability needed when system parameters change.

Inventive Principle:
Principle #15Dynamics

2Adaptability or versatility

If data-driven models are used to model industrial systems, then model flexibility is improved, but the models lack generalizability and soft sensing capabilities

Engineering Contradiction:
Improvemodel flexibilityVSAvoidmodel generalizability
Core Design Contradiction:
Adaptability or versatilityVSReliability

Solution Approach 1:

The patent merges data-driven neural network approaches with physics-based governing equations into a unified framework. This combination allows the model to inherit the flexibility of data-driven methods while gaining the generalizability and physical consistency of physics-based models through the integrated loss functions and constraints.

Inventive Principle:
Principle #5Merging (Combining)

Solution Approach 2:

The patent utilizes parameter changes in the neural network architecture to balance flexibility and generalizability. By dynamically adjusting network parameters and hyperparameters based on available data and physical constraints, the model achieves both adaptability to specific conditions and generalizability across different operating scenarios.

Inventive Principle:
Principle #35Parameter changes

3Reliability

If PINNs are used for complex industrial systems with multiphysics processes, then integration of physical principles and data-driven insights is improved, but convergence challenges occur

Engineering Contradiction:
Improvemodel integration capabilityVSAvoidtraining convergence time
Core Design Contradiction:
ReliabilityVSLoss of time

Solution Approach 1:

The patent segments the complex multiphysics system into distinct sub-processes with separate governing equation components. This segmentation allows for targeted training strategies where each sub-process can be trained with appropriate loss functions and hyperparameters, significantly improving convergence compared to treating the entire system as a single unified model.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent implements preliminary actions in the training process, such as pre-processing data, pre-defining appropriate loss functions for each sub-process, and pre-configuring the neural network architecture before full training begins. These preliminary steps reduce the complexity of the optimization landscape and accelerate convergence for complex multiphysics systems.

Inventive Principle:
Principle #10Preliminary action

4Ease of operation

If conventional training strategies are applied to PINNs for industrial entities, then training process is simplified, but rapid convergence cannot be achieved

Engineering Contradiction:
Improvetraining simplicityVSAvoidtraining speed
Core Design Contradiction:
Ease of operationVSProductivity

Solution Approach 1:

The patent employs dynamic training strategies that adapt during the training process. The loss function weights, learning rates, and other hyperparameters are dynamically adjusted based on the training progress and the specific characteristics of each sub-process, enabling rapid convergence while maintaining ease of operation through automated adaptation.

Inventive Principle:
Principle #15Dynamics

Solution Approach 2:

The patent systematically changes key training parameters such as loss function weights, learning rates, and network architecture parameters during training. These parameter changes are orchestrated to accelerate convergence for each sub-process while maintaining overall training simplicity through a structured parameter adjustment framework.

Inventive Principle:
Principle #35Parameter changes

Data Source

PatentEP4664356A1Method and system for domain aware data driven (DADD) modeling of an industrial entity
Publication Date: 2025.12.17 TATA CONSULTANCY SERVICES LTD
  • EP4664356A1 patent drawingFigure 1
  • EP4664356A1 patent drawingFigure 2
  • EP4664356A1 patent drawingFigure 3

AI summary

This disclosure relates generally to a method and system for a domain aware data driven model (DADD) for optimizing complex operations of an industrial entity expressed by process governing equations (PGEs). State-of-the-art methods face convergence challenges when applied to complex systems of industrial entity. Moreover, process descriptors available for these complex systems often tend to be sparse. The disclosed method involves domain aware neural network modeling of the complex industrial entities. The method involves obtaining process governing equations (PGEs), spatiotemporal domain co-ordinates and sparse measurements of the industrial entity. The domain aware model is generated by extracting sub-process governing equation component from the plurality of PGEs, followed by designing a grouped neural network architecture (GNNA) having individual neural network subset parameter of each sub-process governing equation component. The neural network architecture is sequentially trained and fine-tuned to finally predict process descriptors for the industrial entity.