Introducing Artificial Neural Networks in the Physics Laboratory: A Compound Pendulum Case Study

This study demonstrates that integrating Artificial Neural Networks into undergraduate physics laboratories as a complementary tool for a compound pendulum experiment significantly enhances the precision and consistency of determining gravitational acceleration compared to traditional analytical methods, without replacing the fundamental experimental approach.

The Gist

Imagine you are trying to measure the strength of Earth's gravity (how fast things fall) using a classic physics experiment: a swinging metal bar, known as a compound pendulum.

For decades, physics students have done this by swinging the bar, timing it with a stopwatch, measuring the length, and doing some math on paper to calculate gravity. It's a great lesson, but it's messy. Your hand might shake when you start the stopwatch, the air might push the bar slightly, or the ruler might be a tiny bit off. These little "noise" factors make your final answer a little fuzzy, like a photo taken with a shaky camera.

This paper introduces a new, modern twist: teaching students to use Artificial Neural Networks (ANNs)—a type of computer brain—to help clean up that mess.

Here is the story of what they did, explained simply:

1. The Old Way vs. The New Way

• The Old Way (The Manual Labor): Students swing the bar, time it, and use a formula. Because of human error and air resistance, their result for gravity was 1009.03, but they weren't sure of the last digits. Their "uncertainty" (the range of error) was quite wide, like saying, "It's probably between 1002 and 1016."
• The New Way (The Computer Brain): The students took all their messy measurements and fed them into a computer program designed to think like a brain. This program didn't just do math; it learned the patterns in the data. It looked at thousands of swings and figured out, "Ah, when the bar is this long and swings this fast, gravity is exactly this."

2. The "Smart Brain" Analogy

Think of the Artificial Neural Network (ANN) like a super-taster at a restaurant.

• The Traditional Method is like a chef who tastes a soup once, guesses the salt level based on a recipe, and writes it down. If the chef is tired or the spoon is dirty, the guess might be off.
• The ANN is like a chef who has tasted that same soup 1,000 times. Even if the salt shaker is sticky or the spoon is slightly bent, the chef's brain has learned the true flavor profile. The chef can ignore the little mistakes and tell you the exact salt level with incredible precision.

In this study, the "chef" (the ANN) was trained on 70% of the students' data. It learned the relationship between the length of the bar, the time it took to swing, and the angle it started at. Then, it was tested on the remaining 30% of data it had never seen before.

3. The Results: Sharper Focus

The results were fascinating:

• The Traditional Method said gravity was 1009.03 ± 6.82. The big number after the plus/minus means the result is a bit fuzzy.
• The ANN Method said gravity was 1009.029858 ± 0.000592.

Notice the difference? The computer didn't just guess the same number; it guessed it with laser precision. The "fuzziness" (uncertainty) was almost completely gone. It's like going from a blurry, pixelated photo to a high-definition 4K image.

4. Why This Matters for Students

The authors aren't trying to replace the swinging bar experiment. They want to keep the hands-on fun of physics. Instead, they are adding a "second layer" of learning:

• Bridging the Gap: It teaches students that physics isn't just about formulas on a chalkboard; it's also about how we handle messy, real-world data.
• Learning to "Listen" to Data: Students learn that sometimes the data tells a story that simple math misses. The computer found the "hidden rhythm" in the noise.
• Future-Proofing: By doing this, students learn a skill used in modern science and engineering: Machine Learning. They learn that sometimes, the best way to solve a problem isn't to write a perfect equation, but to let a computer learn the pattern from the data itself.

The Bottom Line

This paper shows that you can take a classic, old-school physics lab and upgrade it with modern technology. By letting a "computer brain" help analyze the data, students get a clearer, more precise answer than they could get with a stopwatch and a ruler alone. It's a perfect example of how old wisdom (physics) and new tools (AI) can work together to see the world more clearly.

Technical Summary

1. Problem Statement

Traditional undergraduate physics laboratories rely on analytical methods and manual calculations to determine physical constants, such as the acceleration due to gravity ($g$). In the context of the compound pendulum experiment, these conventional methods face several limitations:

• Idealized Assumptions: Theoretical models assume ideal conditions (frictionless pivots, negligible air resistance, small oscillations) that rarely exist in real-world experiments.
• Experimental Noise: Real data is subject to random errors (e.g., human reaction time, air currents) and systematic errors (e.g., instrument calibration), leading to significant uncertainties.
• Pedagogical Gap: Students often struggle with complex error quantification and lack exposure to modern data-driven modeling techniques. While high-precision instruments (like cold atom gravimeters) exist, they are too expensive and technically complex for standard undergraduate labs.

The study aims to bridge this gap by integrating Artificial Neural Networks (ANNs) as a complementary tool to traditional experimental methods, enhancing precision and introducing students to machine learning concepts without replacing the fundamental physics.

2. Methodology

The research employs a hybrid approach combining a standard physical experiment with a data-driven ANN model.

A. Experimental Setup (Traditional)

• Apparatus: A 100 cm metal bar with pre-drilled holes at 5 cm intervals acting as a compound pendulum.
• Procedure: The time period ($T$) for 10 oscillations was measured using a digital timer for five repetitions at various pivot points.
• Variables:
  • Input Parameters: Initial angular displacement ($\theta$), Effective length ($L_{eff}$), and Time period ($T$).
  • Target Output: Acceleration due to gravity ($g$).
• Data Collection: Measurements were taken for angular displacements between $1^\circ$ and $5^\circ$. The dataset was split into 70% training, 15% validation, and 15% testing subsets.

B. ANN Architecture and Training

• Network Structure: A feedforward neural network with:
  • Input Layer: 3 neurons ($\theta, L_{eff}, T$).
  • Hidden Layer: 9 neurons (determined as optimal via trial-and-error) using the Hyperbolic Tangent (tanh) activation function.
  • Output Layer: 1 neuron (predicting $g$) using a linear activation function.
• Training Algorithm: The Levenberg–Marquardt algorithm was used for optimization, utilizing the Jacobian matrix to iteratively update weights and biases.
• Loss Functions: Mean Squared Error (MSE) and Mean Absolute Error (MAE) were used to quantify the difference between predicted and actual values.
• Preprocessing: Min-max normalization was applied to scale data within the $[0, 1]$ range to ensure stability and faster convergence.
• Stopping Criteria: Training was halted based on the "early stopping" criterion when the validation error began to increase (indicating overfitting), specifically at epoch 127.

3. Key Contributions

• Pedagogical Innovation: The study proposes a novel framework for undergraduate labs where students perform the physical experiment and then use the same data to train an ANN. This bridges the gap between classical experimental physics and modern data science.
• Precision Enhancement: It demonstrates that ANNs can significantly reduce the uncertainty associated with experimental noise compared to traditional analytical calculations.
• Optimization of Model Complexity: The paper systematically analyzes the impact of the number of hidden nodes (1–30) on model performance, providing a practical guide for selecting optimal architectures to avoid underfitting and overfitting in physics contexts.
• Validation of Data-Driven Physics: It validates that ANNs can learn the non-linear physical relationships governing the compound pendulum directly from experimental data without requiring explicit derivation of complex equations during the prediction phase.

4. Results

The study compared the results derived from standard analytical methods against the ANN predictions:

Metric	Traditional Experimental Method	ANN Model Prediction
Mean Value of $g$	$1009.03 \text{ cm/s}^2$	$1009.029858 \text{ cm/s}^2$
Uncertainty/Error	$\pm 6.82 \text{ cm/s}^2$ (Standard Deviation)	$\pm 0.0005925 \text{ cm/s}^2$ (MAE)
Precision	Lower (High variance due to noise)	Extremely High (Variance reduced by $\sim 10^4$)
Correlation ($R$)	N/A	$\approx 1.0$ (across training, validation, and test sets)

• Optimal Architecture: The model with 9 hidden nodes yielded the lowest overall MAE ($0.0005925$) and MSE ($0.00150054$).
• Overfitting Analysis: Configurations with very high node counts (e.g., 12, 16, 20) showed near-zero training error but high validation/test errors, confirming overfitting. Conversely, very low node counts led to underfitting.
• Error Distribution: Histograms of prediction errors showed a sharp central peak near zero for all datasets, confirming the model's ability to generalize and minimize random fluctuations.

5. Significance

• Educational Impact: The approach transforms a standard physics lab into an interdisciplinary learning experience. Students gain hands-on experience with machine learning, learning to interpret model reliability, generalization, and the trade-off between model complexity and accuracy.
• Scientific Utility: The ANN model effectively filters out random experimental noise (e.g., timing errors, air currents), providing a more consistent estimate of $g$ than the raw experimental average.
• Scalability: This framework can be extended to other undergraduate physics experiments (e.g., optics, thermodynamics) to modernize laboratory education.
• Future Applications: Once trained, the ANN can rapidly predict $g$ for new input conditions without repeating physical experiments, serving as a tool for virtual laboratory analysis and parameter optimization.

In conclusion, the paper successfully demonstrates that integrating ANNs into undergraduate physics education does not replace traditional methods but enhances them, offering students a powerful tool to understand the interplay between experimental data, theoretical models, and computational intelligence.