Page 67 - IJAMD-2-2

International Journal of AI for
Materials and Design
A unified industrial AI foundation framework

influence model predictions, guiding feature selection. PINNs incorporate physical laws and equations into the
In addition, deep learning-based feature representation learning process, improving model reliability in scenarios
learning can discover complex patterns in datasets. where data are sparse but prior knowledge is available.
Tools such as tsfresh, PyCaret, and Shapley Additive In addition, multi-modal learning approaches, which
Explanations (SHAP) provide automated pipelines combine information from multiple modalities – such as
for feature extraction, feature importance, and feature images, text, audio, tabular data, and sensor measurements
selection. Therefore, effective feature engineering ensures – enable the development of more comprehensive models
that AI models focus on the most informative inputs while suitable for complex industrial scenarios. 49,50 Furthermore,
minimizing redundancy and noise. foundation models, which are pre-trained on large-scale
diverse datasets, offer a promising approach for transferable
4.2.4. Data visualization and adaptable AI across different industrial domains.
Data visualization plays an important role in making These models can be fine-tuned for specific tasks, allowing
complex datasets interpretable for both humans and AI for efficient deployment with reduced training time and
51
systems. Visualization supports tasks such as exploratory improved generalization. In the process of algorithm
44
data analysis, feature space visualization, anomaly design, developers may utilize prior knowledge and
detection, and real-time monitoring. For instance, feature experience to create new model architectures or iteratively
space visualization using t-SNE or PCA projections can refine existing models through systematic experimentation
help detect outliers or clusters in high-dimensional sensor and testing. When open-source implementations are
data. Real-time dashboards that visualize data streams available, they can be reproduced and further extended
from machines and production lines help operational based on established methods.
teams identify issues early and improve system reliability.
Moreover, industrial practitioners are encouraged to utilize 4.3.2. Model interpretability
interactive visualization tools, such as Plotly, Tableau, and Model interpretability is crucial for ensuring that AI
D3, to enable dynamic data exploration. models are transparent and explainable. One important
direction is interpretable ML, which focuses on developing
4.3. Model module models that are transparent by design, such as regression
With structured knowledge and AI-ready data in place, models, decision trees, and generalized additive models.
the model module is responsible for developing intelligent, These models, along with techniques such as rule-based
adaptable, interpretable, and efficient AI models to drive learning and sparse linear models, can provide clear
industrial AI applications in Industry 4.0. This module and consistent explanations, making them suitable for
provides four key aspects to help developers systematically industrial applications. 52,53 However, with the increasing
design, adapt, and refine models that meet complex use of deep learning models, which are often treated as
industrial requirements. black boxes, it has become challenging to understand and
explain how such models make predictions. To address
4.3.1. Algorithm design this, developers are encouraged to use post hoc explanation
54
Algorithm design serves as the foundation for developing methods, such as the SHAP values, Local Interpretable
55
AI models. Developers should select or design models Model-agnostic Explanations, and integrated gradients.
that align with data characteristics, domain constraints, These methods enable feature importance analysis and
and application requirements. Typical options range from provide explanations of model predictions in both global
traditional ML and statistical methods – such as decision and local contexts. In addition, causal inference techniques
tree, random forest, support vector machine, gradient can enhance model interpretability by identifying cause-
boosting, k-nearest neighbors, Gaussian mixture model, and-effect relationships between input variables and model
and various clustering algorithms – to deep learning outcomes. Approaches such as structural causal models,
architectures, including convolutional neural networks counterfactual analysis, and invariant causal prediction can
(CNNs), recurrent neural networks, graph neural networks, be used to provide more robust and stable explanations,
transformers, and their variants. 45,46 These algorithms are particularly in dynamic industrial settings. 56
applicable to problems involving both continuous variables
(e.g., vibration signals, temperature, and pressure) and 4.3.3. Model adaptation and generalization
discrete variables (e.g., operational states, fault modes, and Model adaptation refers to the ability of an AI model to
control events). Beyond traditional ML and deep learning adjust its parameters or structure to accommodate new
architectures, physics-informed neural networks (PINNs) data distributions, domains, or operational scenarios
have emerged as valuable tools in industrial AI. 47,48 without requiring complete retraining. Meanwhile, model

Volume 2 Issue 2 (2025) 61 doi: 10.36922/IJAMD025080006