🔬 physics

A general framework for disturbance compensation in airborne and seaborne gravimetry supported by machine learning

This paper proposes a general framework for compensating environmental disturbances in airborne and seaborne gravimetry by integrating multi-sensor systems, supervised machine learning, and a dedicated laboratory training platform, while addressing the unique challenges of validating such methods for gravity measurements through specific experimental solutions.

Original authors: Lorenzo Iafolla, Massimo Chiappini, Roberto Carluccio, Stefano Chiappini, Emiliano Fiorenza, Carlo Lefevre, Pasqualino Loffredo, Marco Lucente, Francesco Santoli

Published 2026-09-01

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Lorenzo Iafolla, Massimo Chiappini, Roberto Carluccio, Stefano Chiappini, Emiliano Fiorenza, Carlo Lefevre, Pasqualino Loffredo, Marco Lucente, Francesco Santoli

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

Measuring gravity is one of the most precise tasks in science, yet it is also one of the most easily spoiled. Gravity is the invisible pull that keeps our feet on the ground, but to scientists, it is a subtle signal that reveals the hidden density of rocks beneath the Earth's surface, helping to map oil reserves, locate underground water, and understand the planet's shifting mass. To capture this signal, researchers use instruments called gravimeters, which are essentially ultra-sensitive scales. However, these scales are notoriously fragile. They do not just feel the pull of the Earth; they also feel the heat of the sun, the vibration of a passing truck, and the tilt of the ground beneath them. When these instruments are placed on moving platforms like airplanes or ships, the problem becomes overwhelming. The motion of the vehicle creates forces that are thousands of times stronger than the tiny variations in gravity the scientists are trying to find. It is like trying to hear a whisper in the middle of a roaring stadium. For decades, engineers have tried to build better shields or write complex mathematical formulas to cancel out this noise, but the environment is often too chaotic for simple fixes.

A team of researchers from Italy has proposed a new way to solve this problem by teaching the instruments to learn from their own mistakes. Instead of relying solely on rigid physical laws or perfect hardware, they developed a framework that combines a cluster of different sensors with a specialized computer program known as machine learning. The core idea is simple but powerful: if you can show a computer exactly how a sensor behaves when it is being disturbed, the computer can learn to predict and remove that disturbance in real time. To do this, the researchers built a laboratory training ground that mimics the chaotic conditions of the real world. They placed their sensitive instruments on a platform that could be heated, cooled, tilted, and shaken in precise, controlled ways. By recording how the sensors reacted to these specific, known changes, they created a massive library of data. They then fed this data into a machine learning algorithm, which studied the patterns until it could distinguish between the true signal of gravity and the false noise created by temperature changes or the tilting of the platform.

The team tested this approach in two distinct scenarios to see if it could handle different types of interference. First, they focused on temperature. Even a tiny change in heat can cause the metal parts inside a sensor to expand or contract, leading to false readings. In their experiment, they surrounded a high-precision sensor with eleven different thermometers to map the temperature inside the device from every angle. They then used heating mats and lamps to create random, shifting patterns of heat and cold. A traditional method might try to correct for this using a single thermometer and a simple formula, but the researchers found that this was not enough. By using all eleven temperature readings and letting the machine learning model analyze the complex relationships between them, they were able to strip away the thermal noise with far greater accuracy. The result was a dramatic reduction in error, proving that the computer could learn to ignore the heat in a way that standard math could not.

The second test tackled the difficulty of measuring gravity while the platform is moving and tilting. When an airplane banks or a ship rolls, the sensor inside is no longer pointing straight down at the Earth's center. This tilt mixes the gravity signal with the motion of the vehicle, creating a confusing jumble of data. To train their system, the researchers built a platform that could tilt back and forth while simultaneously moving up and down to simulate the feeling of gravity changing. They used a reference sensor to measure the exact motion of the platform, giving the computer a "correct answer" to compare against. When they let the machine learning model process the data from the tilting sensor and the gyroscopes that measured the rotation, it outperformed the best existing mathematical models. The computer successfully reconstructed the true vertical motion, reducing the error by more than half compared to the standard method. This showed that the system could learn to untangle the complex web of forces acting on the sensor, even when the platform was moving in ways that made the gravity signal almost invisible.

The researchers are careful to note that this is not a finished product ready to be installed on every plane or ship tomorrow. The work is a proof of concept, demonstrating that the method works in a controlled laboratory setting. They acknowledge that the real world is far messier than their training platform, and that future work will need to refine the hardware and the software to handle the unpredictability of actual flights and voyages. However, the study establishes a clear path forward. By combining a multi-sensor system with a dedicated training environment and machine learning, they have shown that it is possible to teach instruments to compensate for disturbances that were previously thought to be too complex to fix. This approach does not replace the need for high-quality sensors, but it gives them a new level of intelligence, allowing them to see the true signal of the Earth even when the world around them is shaking and shifting.

Technical Summary: A General Framework for Disturbance Compensation in Airborne and Seaborne Gravimetry Supported by Machine Learning

Problem Statement
High-accuracy gravimetry on moving platforms (airborne and seaborne) faces a critical metrological challenge: the separation of the gravity signal from overwhelming environmental and kinematic disturbances. While static gravimeters can achieve sub-microgal precision, mobile applications are plagued by motion-induced noise, platform rotations, and environmental variability (temperature, pressure, humidity) that can exceed the gravity signal by several orders of magnitude. Conventional compensation methods—such as inherent sensor insensitivity, physical insulation, opposing inputs (e.g., gimbals), and analytical corrections (e.g., Kalman filters)—often fail to address complex, non-linear interactions between multiple disturbances or are impractical due to weight, volume, and power constraints. Furthermore, the nature of gravity as a measurand makes it difficult to generate controlled variations in a laboratory setting for calibration, hindering the development of robust correction models.

Methodology
The authors propose a general framework for disturbance compensation built on three pillars:

Multi-Sensor System: A primary sensor (e.g., accelerometer) measures the measurand (gravity) but is also sensitive to disturbances. Secondary sensors (e.g., thermometers, gyroscopes, barometers) measure the specific disturbances. The system requires precise synchronization of all data streams to enable effective signal fusion.
Machine Learning (ML): Instead of relying solely on analytical models, the framework employs supervised machine learning (specifically regression algorithms) to learn the complex, non-linear mapping between the multi-sensor inputs and the desired output. The ML model is trained to predict either the corrected measurand directly or the disturbance-induced error to be subtracted.
Dedicated Training Platform: To overcome the difficulty of varying gravity in a lab, the authors introduce a laboratory-based training platform capable of physically reproducing operational disturbances and, crucially, generating a known reference signal for the measurand. Two training strategies are defined:
- General Strategy: Actuators vary both the measurand and disturbances, with a high-accuracy reference sensor providing the "ground truth" label.
- Simplified Strategy: Only disturbances are varied. The ML model learns to predict the error induced by disturbances, assuming linear superposition, and the corrected signal is derived by subtracting this predicted error from the raw sensor output.

Experimental Case Studies
The framework was validated through two distinct experimental setups:

Case 1: Temperature and Thermal Gradient Rejection: A high-sensitivity tri-axial accelerometer was instrumented with 11 thermometers. The training platform used heating mats and lamps to induce random temperature and thermal gradient fluctuations. The ML model was trained to predict the error in the accelerometer output based on the multi-point temperature data.
Case 2: Pitch and Roll Rejection for Vertical Acceleration: A multi-sensor system comprising an accelerometer and a Fiber Optic Gyroscope (FOG) was mounted on a custom-built platform. Using the equivalence principle, the platform generated controlled vertical oscillations (simulating gravity variations) and pitch/roll tilts. The reference signal was derived from high-precision linear encoders measuring the platform's displacement. The ML model was trained to reconstruct the vertical acceleration from the noisy, rotation-contaminated accelerometer signal and gyroscope data.

Key Results

Temperature Rejection: The ML-based approach significantly outperformed conventional single-thermometer linear regression. For the horizontal axis, the Root Mean Square Error (RMSE) was reduced from $6.6 \times 10^{-6} \, \text{m/s}^2$ to approximately $2.0 \times 10^{-6} \, \text{m/s}^2$ , with the coefficient of determination ( $R^2$ ) improving from negative values to $\sim 0.91$ . The vertical axis showed similar improvements, reaching an $R^2$ of 0.999.
Pitch and Roll Rejection: In reconstructing vertical acceleration, the ML-based model achieved an RMSE of $0.9 \times 10^{-3} \, \text{m/s}^2$ (90 mGal), a 51% reduction compared to the analytical system model baseline ( $2.0 \times 10^{-3} \, \text{m/s}^2$ ). The ML approach demonstrated superior ability to capture residual system-specific effects and non-linearities not accounted for in the analytical model.

Significance and Claims
The paper claims that this framework offers a complementary tool to conventional methods, extending disturbance rejection capabilities to regimes where analytical modeling and physical compensation are impractical. The primary contribution is methodological rather than operational; the work demonstrates the feasibility of the approach and defines the necessary procedures for implementation.

Key assertions include:

Data-Driven Flexibility: The method automatically accounts for system-specific characteristics (non-linearities, misalignments, assembly imperfections) without requiring explicit physical modeling.
Controlled Training: The training platform allows for the generation of datasets spanning the full range of expected operating conditions, mitigating the risk of extrapolation errors common in field-based training.
Adaptability: The framework is not a "plug-and-play" solution but requires dedicated hardware and training strategies tailored to specific applications (e.g., the specific actuator design for simulating gravity variations).
Future Direction: The authors suggest that future iterations could benefit from "model-based machine learning," where physical laws constrain the ML predictions, and from more advanced training platforms (e.g., hexapods) and refined sensor integration (e.g., co-located IMUs).

The authors explicitly state that the reported results are proof-of-concept demonstrations under controlled laboratory conditions and should not be interpreted as final performance specifications for operational airborne or seaborne gravimeters.

Technical Summary: A General Framework for Disturbance Compensation in Airborne and Seaborne Gravimetry Supported by Machine Learning

More like this