# Machine Learning in Digital Twin

## 0. Post Details
- Reference: [Nele et al., Towards the application of machine learning in digital twin technology: a multi-scale review, 2024.](https://link.springer.com/article/10.1007/s42452-024-06206-4)
- Post by: Jedd Yang
- Date: 2025-05-12
- Keywords: Computational Mechanics, Digital Twin

## 1. The Three Layers

The article explores the digital twin (DT) framework, breaking it into 3 layers.

### 1.1 Perception Layer
This layer serves as the gateway between the physical asset and its DT by capturing real-time data via sensors and IoT devices. It performs essential preprocessing:
1. Sensor fusion — integrating data from multiple sensors via regression to estimate a system’s state. Uncertainties probabilistic methods (Kalman Filters (KFs)) are often used for noise reduction.
2. Outlier rejection — ensuring data used for training models is representative and not out-of-distribution data or outliers.
3. Data pre‑processing — Applying scaling methods, such as normalisation and standardisation to clean up the data.

### 1.2 Digital Object Layer
This layer builds on a data‑driven models of the physical system, simulating and forecasting the system using the preprocessed data. These models are essentially statistical in nature.
1. Regression for static modeling — using Linear or Polynomial regression. Methods such as Ordinary Least Squares (OLS), Lp-Norm estimators, or gradient descent are applied for optimizing model parameters.
2. Auto-Regressive Polynomial modeling for dynamics — modeling the current state by referencing previous states or observations, capturing system dependencies over time.
3. Neural Networks (NN) for regression and classification — learning patterns by iteratively adjusting the weights between neurons to minimize the difference between predicted and actual outputs. Physics-informed NN are especially blooming.

### 1.3 Application Analysis Layer
This layer integrates outputs from both the previous two to extract actionable insights and drive decisions. Advances analytics methodologies may be used in this layer:
1. Clustering analysis — grouping samples contained in a dataset based on their similarity, which helps anomaly detection.
2. Dimensionality reduction — simplifying the problem by reducing data dimensions while preserving key information, enabling better visualization and faster analysis.
3. Fuzzy logic — handling uncertainty and imprecise data in complex systems by applying human-like reasoning, often used in control systems where binary logic falls short.
4. Reinforcement learning — optimizing decision-making through interaction with the environment, making it useful for real-time control and adaptive optimization.
5. Large language models — controlling DT through natural language interfaces, simplifying human-machine interaction.