Bridging physics, simulations, and AI to deliver faster, explainable insights in manufacturing operations
Physics-Based Digital Twins
Digital Twins have become a widely discussed concept alongside the rise of the Industrial Internet of Things (IIoT) and the broader Industry 4.0 movement. While the underlying themes have existed for decades, advances in internet infrastructure and cloud computing have transformed what’s possible. Today, it’s not just about monitoring machines remotely but about predicting and optimizing industrial processes using data, models, and intelligence.
The early promise of digital twins centered around real-time visibility: tracking the current state of assets through sensors, dashboards, and alerts. Use cases like predictive maintenance and prescriptive control dominated industry conversations, offering solutions for machine downtime and process deviations. These ideas were compelling—and partially effective—but often fell short of delivering deeper insights. Why?
The Limits of Data-Only Approaches
While the phrase “more data leads to better models” is often true, in industrial settings it comes with caveats:
- Data availability is limited, especially across the full range of operating conditions (the process envelope).
- Failure data is scarce—most systems are designed to avoid failure, so models often lack examples of what "bad" looks like.
- Inputs and outputs are complex—unlike simple systems, industrial processes often involve multiple interacting variables (e.g., powder flow, temperature, material properties), and KPIs (e.g., throughput, quality) depend on a blend of design, materials, and control actions.
Why Physics Matters in Manufacturing
Industrial systems follow the laws of physics. For decades, engineers have relied on process models—ranging from statistical regressions to first-principles-based 1D models—to analyze and design manufacturing processes.
More recently, high-fidelity simulation tools such as:
- Computational Fluid Dynamics (CFD)
- Finite Element Method (FEM)
- Discrete Element Method (DEM)
…have become essential for understanding process behavior, predicting performance, and optimizing operations.
However, traditional simulations are slow, resource-intensive, and typically used only by specialized R&D teams—not in real-time operations.
Physics-Informed Machine Learning
This is where Physics-Based Digital Twins shine.
By using validated simulations (e.g., CFD or DEM) to generate synthetic data across a range of scenarios, we can train machine learning models that:
- Learn from both physics and data
- Require fewer real-world samples
- Offer faster predictions with explainable results
These models can be deployed as advisors or soft sensors, providing insights to engineers, operators, and plant managers in real time.
A More Complete Digital Twin
What makes these twins complete is not just their predictive accuracy—but their explainability and generalizability. They bridge the gap between R&D simulations and shop-floor decisions.
By integrating physics and data, these twins can:
- Cover a broader design space
- Offer quick feedback loops
- Support better process understanding and decision-making
Real-World Example: Powder Blending in Food and Pharma
One of the most impactful applications of physics-based digital twins we've worked on is in the area of powder blending—a critical unit operation in both pharmaceutical and food industries.
Ensuring blend uniformity is notoriously challenging. Why? Because powder properties—like particle size distribution and flowability—often vary due to:
- Supplier-to-supplier differences
- Environmental exposure during storage
- Changes in batch size or equipment across manufacturing partners
To analyze such complexity, we leverage the Discrete Element Method (DEM) to simulate particle dynamics and mixing behavior. However, two key challenges arise:
- Material Calibration: Accurate DEM simulations depend on reliable input parameters like friction, cohesion, and particle density. These are not always known.
- Computational Speed: Full DEM simulations are time-consuming and not suited for real-time use.
To overcome both, we developed a physics-guided machine learning approach:
- Step 1: A rotating drum experiment captures the dynamic angle of repose using a machine vision system. An ML model then predicts the most suitable DEM material properties based on the captured profile.
- Step 2: A second ML model, trained using DEM simulations and validated against experimental results (e.g., assay tests), predicts blend uniformity in real time. It considers both global metrics (like the Mixing Index) and local variations (e.g., mass fraction at different sampling zones).
The result is a comprehensive digital twin that:
- Integrates physics (DEM)
- Is grounded in real measurements
- Runs in real time
- And is deployed as a web application for easy access by non-expert users
A deeper dive into this case study will be covered in our upcoming follow-up blog.
Why Intelimek
At Intelimek, we help companies in pharma, food, automotive, and steel industries adopt physics-based digital twins to drive performance, consistency, and insight in their operations. Our approach combines high- fidelity modeling, data science, and domain expertise—bridging the gap between science and operations.