🤖 machine learning

Green Physics-Informed Machine Learning Models For Structural Health Monitoring

This paper evaluates the environmental impact of structural health monitoring models by demonstrating how physics-informed "grey-box" approaches can achieve high extrapolative performance with reduced computational costs and carbon emissions compared to purely data-driven "black-box" models.

Original authors: Daisy R Bradley, Elizabeth J Cross

Published 2026-05-01

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Daisy R Bradley, Elizabeth J Cross

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to teach a robot how to predict how a bridge or an airplane wing will vibrate when the wind blows. Usually, we teach robots using a "Black Box" approach: we feed them thousands of photos of the bridge in different winds and say, "Learn the pattern." This works, but it's like trying to learn to swim by reading a million books about water without ever getting in the pool. It takes a lot of time, a lot of energy, and if the robot sees a wind pattern it hasn't seen before, it might get confused.

This paper introduces a smarter way called "Green Physics-Informed Machine Learning." Think of this as giving the robot a Swim Instructor (the physics) alongside the books. Instead of just memorizing data, the robot is told the basic rules of how water (or metal) moves. This is called a "Grey Box" model because it's a mix of data and human knowledge.

Here is the simple breakdown of what the authors, Daisy Bradley and Elizabeth Cross, did and found:

1. The Goal: Saving Energy (The "Green" Part)

Computers use electricity, and making that electricity often creates carbon emissions (like smoke from a factory). The authors wanted to know: Does teaching a robot with physics rules save energy compared to just feeding it data?

They compared three types of "robots" (mathematical models):

Black-1 (The Pure Learner): Has no prior knowledge. It must learn everything from scratch using only data.
Grey-1 (The Student with a Hint): Knows the general rules of physics but still has to figure out the specific details (like the exact period of a vibration) by looking at data.
Grey-2 (The Expert): Knows the physics rules perfectly and doesn't need to guess the details at all.

2. The Experiment: The "Toy" and the "Real Thing"

They tested these models in two ways:

The Toy Box: A simple math problem (like a wavy line) to see how the models behave.
The Real Thing: A real lab test of a small metal airplane wing (the GARTEUR structure) to see if it holds up in the real world.

The Big Discovery:
The models with physics knowledge (the Grey Boxes) needed much less data to get the job done right.

The "Pure Learner" needed 80% of the data to get a good answer.
The "Expert" (Grey-2) only needed 20% of the data to get the same good answer.

3. The Catch: Complexity vs. Data

Here is the twist. Teaching a robot physics makes the robot slightly more complicated to build (it has more "knobs" or settings to tune).

In the simple "Toy" test: The "Expert" model was the winner. Because it needed so little data, it finished its work super fast, using the least amount of electricity. The savings from needing less data were bigger than the cost of the extra complexity.
In the "Real" airplane test: It was a bit more mixed. Sometimes, the extra complexity of the physics model made it take more time and energy, especially when the amount of data was very small. However, when they used a denser set of data (more points to learn from), the "Expert" model shined again, using 89% less energy than the pure learner.

4. The Bottom Line

The authors conclude that adding physics to machine learning is like giving a student a map.

If you are walking a short distance (small data sets), carrying the map might feel like extra weight.
But if you are walking a long distance (large, complex engineering problems), having the map means you don't have to wander around guessing. You get to the destination much faster, saving a huge amount of energy.

In short: By mixing human engineering knowledge with computer learning, we can teach machines to do their job with less data, which means they run faster and burn less carbon—making structural health monitoring (checking if bridges and planes are safe) more sustainable.

1. Problem Statement

The paper addresses the growing environmental impact of Machine Learning (ML) in engineering, specifically within Structural Health Monitoring (SHM). While ML is increasingly used for regression and classification tasks in engineering, purely data-driven ("black-box") approaches have significant limitations:

Data Scarcity: They often fail to generalize to unseen environmental or operational conditions (EOCs) where training data is unavailable.
High Computational Cost: To achieve high accuracy, black-box models typically require massive datasets, leading to long training times and high carbon emissions.

The authors hypothesize that Physics-Informed Machine Learning (PIML) or "grey-box" models, which embed physical knowledge into the model architecture, can reduce the amount of training data required. The core research question is whether the reduction in data requirements (and thus runtime) outweighs the potential increase in model complexity caused by embedding physics, ultimately resulting in lower carbon emissions.

2. Methodology

The study employs a comparative framework using Gaussian Process (GP) regression models, chosen for their interpretability and flexibility in embedding physics without altering the core architecture.

A. Model Architecture

Three distinct GP models were designed to test varying levels of "greyness" (integration of physical knowledge):

Black-1 (Black-box): A standard GP with a constant mean and a Squared Exponential (SE) kernel. It has 4 hyperparameters to learn (length-scale, output variance, mean, noise).
Grey-1 (Semi-physics): An SE kernel multiplied by a Periodic kernel (acting as a proxy for physical insight regarding the system's periodicity). The period parameter is bounded within 10% of the true value. It has 6 hyperparameters.
Grey-2 (High-physics): Similar to Grey-1, but the period parameter is fixed at the true value (not optimized during training). It has 5 hyperparameters.

B. Experimental Setup

The study utilized two test cases:

Toy-box Case: A synthetic surface defined by $y = \sqrt{|x_2|} \sin(4x_1)$ , designed to test extrapolation capabilities.
Engineering Case (SHM): Real-world data from the GARTEUR benchmark (a metallic aircraft structure). The task was to predict acceleration responses spatially and temporally. Two data layouts were tested:
- Original: Sparse data points.
- Upsampled: Denser data points (200 points per 10% coverage).

C. Evaluation Metrics & Tools

Performance Threshold: Models were trained with varying data coverage (10% intervals) until they achieved a Normalized Mean Squared Error (NMSE) < 10.
Carbon Estimation: The CodeCarbon package was used to estimate $gCO_2e$ emissions based on runtime, CPU Thermal Design Power (TDP), and electricity carbon intensity.
Hardware: Experiments ran on a Dell Inspiron 5680 (Intel Core i7-8700, 32GB RAM).

3. Key Contributions

Green Lens on PIML: This is the first work to explicitly evaluate PIML models through the lens of environmental sustainability (carbon emissions) rather than just predictive accuracy.
Trade-off Analysis: The study quantifies the trade-off between model complexity (number of hyperparameters) and data efficiency. It demonstrates that embedding physics reduces the data needed to reach a performance target but increases the computational cost per iteration due to more hyperparameters.
Scalability Insight: The authors establish that while simple models show mixed results, the carbon-saving potential of PIML scales significantly with larger datasets, where the reduction in training data volume dominates the increased complexity cost.

4. Results

The results indicate that the environmental benefit of PIML is highly dependent on the dataset size and the specific configuration of the physics integration.

Metric	Toy-box Case	GARTEUR (Original)	GARTEUR (Upsampled)
Data Reduction	Grey-2 required 75% less data than Black-1.	Grey-2 required 57% less data.	Grey-2 required 75% less data.
Emissions ( $gCO_2e$ )	Grey-2 reduced emissions by 73.7% (0.124 vs 0.471).	Grey-2 increased emissions by 31.9% (0.0984 vs 0.0746).	Grey-2 reduced emissions by 89.7% (0.209 vs 2.04).
Key Driver	Data reduction outweighed hyperparameter cost.	Small data size meant hyperparameter optimization cost dominated.	Large data size meant data reduction savings dominated.

Toy-box & Upsampled GARTEUR: The physics-informed models (especially Grey-2) achieved the target accuracy with significantly less data. The resulting reduction in training time led to massive carbon savings (up to ~90% reduction).
Original GARTEUR: Due to the very small dataset size, the computational overhead of optimizing extra hyperparameters in the grey-box models outweighed the time saved by having fewer data points, resulting in slightly higher emissions compared to the black-box model.

5. Significance and Conclusion

The paper concludes that Physics-Informed Machine Learning is a viable strategy for "Green Computing" in engineering, but its success is conditional on dataset scale.

Small Datasets: The overhead of learning additional physical parameters may negate the benefits of reduced data requirements.
Large Datasets: The cubic complexity of GP training ( $O(N^3)$ ) means that even a modest reduction in training data size ( $N$ ) leads to exponential reductions in runtime and carbon emissions. In these scenarios, the "cost" of extra hyperparameters is negligible compared to the savings from data reduction.

Future Work: The authors plan to extend this research to larger-scale models (e.g., Neural Networks and PINNs) and more complex engineering problems to validate that this trend holds as models scale to industrial levels. The study provides a foundational framework for engineers to design sustainable ML models by explicitly balancing physical insight against computational burden.