The difficulty is that a resonance spectrum is exquisitely sensitive to everything at once. The positions, heights, and shapes of the peaks respond not only to the property an engineer wants to measure, such as internal pressure, but also to a host of confounding factors: manufacturing tolerances in the vessel’s dimensions, the quality of the coupling between transducer and wall, temperature drift, and countless other structural quirks. Disentangling the wanted signal from the unwanted variation is a formidable inverse problem. Classical chemometric methods, borrowed from electromagnetic spectroscopy where spectral trends are comparatively simple, struggle with the way acoustic peaks shift, split, merge, or vanish as a structure deforms.

Deep learning offers a way out. Convolutional neural networks, which excel at extracting meaningful patterns from complex one-dimensional signals, can in principle learn to map a messy spectrum directly onto the quantity of interest. But neural networks are voracious: they typically need large, labeled datasets to train without overfitting, and for most real-world ARS applications such datasets are impractical to collect. Each new sensing target—a different vessel, a different material, a different property—would require its own expensive experimental campaign. Simulating the data instead seems attractive, since finite element models can generate thousands of spectra quickly and cheaply. The catch is the domain gap: any mismatch between the simulated physics and the real world, whether in material constants, geometry, or boundary conditions, produces spectra that differ subtly but consequentially from real measurements. A network trained purely on synthetic data learns the features of the simulation, not of reality, and its predictions on laboratory data can collapse dramatically.

In a study published in Frontiers in Acoustics, researchers J. Greenhall and colleagues at Los Alamos National Laboratory report a training framework that closes this gap without using a single labeled real-world measurement. Their approach borrows from computer vision a technique known as Correlation Alignment, or CORAL, an unsupervised domain adaptation method that nudges a neural network to learn features that are common to both the synthetic and the real-world data domains. The team demonstrates the method on a deceptively simple task: reading the internal pressure of a sealed stainless-steel vessel purely from its acoustic resonances, with the network trained on simulated spectra and unlabeled experimental recordings.

The architecture at the heart of the work is a residual network, or ResNet, adapted for one-dimensional spectral input. The raw spectrum passes through three ResNet blocks, each containing four convolutional layers with rectified linear unit activations and batch normalization, with kernel counts and lengths tuned to capture features across the 0.5 to 1.5 kilohertz measurement band. Residual connections that bypass each block’s convolutional stack mitigate the vanishing gradient problem, improving convergence. A global average pooling layer then condenses the final convolutional output into a fixed-length feature vector of 480 values. This vector feeds two heads during training: a label regressor that predicts pressure from synthetic samples, and the CORAL alignment mechanism that compares how features are distributed across the two domains.

CORAL works by explicitly matching the statistics of the latent features. For each training batch, the method computes the covariance matrices of the synthetic and real-world feature vectors and penalizes the difference between them, using the Frobenius norm as the alignment loss. Simultaneously, the network minimizes the mean squared error between predicted and true pressures on the labeled synthetic data. A tunable hyperparameter, the Lagrange multiplier lambda, balances the two objectives: small values let the network fit the simulations, while larger values force it to prioritize domain-invariant representations. Because the alignment loss never touches real-world labels, the method is genuinely unsupervised on the target side—only unlabeled experimental spectra are needed during training.

The experimental demonstration centered on a stainless-steel vessel roughly 185 millimeters in diameter and 222 millimeters tall, sealed with a polymethyl methacrylate lid and instrumented with two piezoelectric transducers glued to its outer wall, one emitting and one receiving, separated by about 100 millimeters. The researchers stepped the internal pressure between 25 and 575 Torr in 25 Torr increments while a capacitance manometer recorded the true pressure, repeating the experiment four times and periodically removing and reattaching the transducers to mimic real-world coupling variability. In total, 92 labeled experimental spectra were collected—far too few to train a deep network conventionally. On the simulation side, the team used COMSOL finite element software with shell physics to model the vessel’s vibrational modes, applying a static pressure difference across the wall and then sweeping frequency to synthesize the resonance spectrum, perturbing the vessel’s height, diameter, and wall thickness with random variations to broaden the training distribution.

The results are striking. A conventionally trained network, fit only to the synthetic spectra, failed almost completely on real data, producing a root mean square error of 179.8 Torr and a coefficient of determination of negative 0.12—worse than predicting the mean. The same architecture trained with CORAL alignment achieved an RMSE of 48.5 Torr and an R-squared of 0.92 on the real-world measurements, a roughly 73 percent reduction in error and an improvement of 1.04 in the coefficient of determination. Visualizations of the learned feature space using t-distributed stochastic neighbor embedding showed that CORAL training pulled the synthetic and real-world feature distributions into overlapping clusters, confirming that the network had learned representations that transcend the domain gap.

The team also mapped out how the method’s key knobs affect performance, offering a practical recipe for future applications. Increasing the Lagrange multiplier from zero rapidly stabilized training and reduced error, with performance plateauing near a value of 0.25, which the researchers adopted for subsequent experiments. Dataset size mattered too: more than 100 synthetic samples were needed for stable results, and beyond roughly 600 samples the error reached its lowest levels. Perhaps counterintuitively, adding more variation to the simulated geometry helped the CORAL-trained model but hurt the conventionally trained one, which drifted further from the real data as the synthetic distribution broadened. Even for CORAL, however, excessive variation beyond a normalized amplitude of about 0.2 slightly degraded accuracy, indicating that an optimal spread of simulation parameters exists and should be tuned for each new sensing problem.

The implications extend well beyond pressure gauges. Because the framework requires only a plausible baseline simulation, a set of physical parameters to perturb, and a stream of unlabeled real-world measurements, it could in principle be applied to any ARS target—structural health monitoring of bridges and pipelines, material characterization, biomedical diagnostics, or manufacturing quality assurance—wherever the resonance spectrum is sensitive to the property of interest. The authors suggest that the same strategy may generalize to other sensing modalities where synthetic data are abundant but experimental labels are scarce, from electromagnetic spectroscopy to ultrasound imaging. If the approach proves as portable as it appears, the bottleneck that has kept resonance-based sensing a laboratory curiosity—hunger for real-world training data—may finally be lifted, letting cheap simulated physics do the heavy lifting that once required months of painstaking measurement.

Subject of Research: Unsupervised domain adaptation for synthetic-data-trained acoustic resonance sensing models

Article Title: Generalizing synthetic data-trained acoustic predictive models to real-world measurements

Article References: Greenhall, J., Pereira, D., Prisbrey, M., Vakhlamov, P., Chavez, C., Pantea, C., & Davis, E. S. (2026). Generalizing synthetic data-trained acoustic predictive models to real-world measurements. Frontiers in Acoustics, 4, Article 1867954. https://doi.org/10.3389/facou.2026.1867954

Image Credits: AI Generated

DOI: 10.3389/facou.2026.1867954

Keywords: acoustic resonance spectroscopy, domain adaptation, CORAL, machine learning, nondestructive evaluation, finite element simulation, pressure sensing, neural networks, synthetic data, noninvasive sensing, Los Alamos National Laboratory, structural health monitoring